SIGIR 2026 · Conference Program

Personalized generative recommender systems have emerged as a promising solution for fashion recommendation. However, existing methods primarily rely on implicit visual embeddings from historical interactions, which often contain preference-irrelevant information and result in insufficient user behavior modeling. Moreover, these models typically generate only item images, providing limited interpretability. To address these limitations, we propose DualFashion, a Dual-Diffusional Generative Fashion Recommendation Architecture that jointly models image and text modalities for personalized and explainable recommendation. DualFashion adopts a dual-diffusion Transformer with image and text branches, where structured attribute-level captions and visual outfit information are jointly used as conditioning signals to model user behavior. The proposed architecture produces both fashion item images and textual descriptions, ensuring visual compatibility while providing explicit semantic interpretability. Furthermore, we introduce a text-augmented fine-tuning strategy that enhances generation diversity and enables effective cross-modal knowledge transfer without incurring heavy computational costs. Extensive experiments on iFashion and Polyvore-U across Personalized Fill-in-the-Blank and Generative Outfit Recommendation tasks demonstrate that DualFashion achieves strong performance in behavior modeling, interpretability, and efficiency compared to state-of-the-art methods. Our code and model checkpoints are available at https://github.com/LinkMingzhe/DualFashion.

Sequential recommendation predicts the next items a user will interact with by modeling evolving preferences over time. Recent diffusion-based generative recommenders show promise in capturing complex dependencies, but they typically treat temporal context as an external conditioning signal rather than integrating temporal transitions into the diffusion dynamics. In this paper, we introduce TDRec (Temporally-aware Diffusion for sequential Recommendation), a novel framework that integrates temporal progression into both forward and reverse processes: at each diffusion step, a position's latent is updated by noise injection and by mixing with its preceding latent. We derive a closed-form solution for this temporal mixing process, proving that it allows for efficient parallel training with O(1) complexity relative to sequence length. Furthermore, we establish the existence of a corresponding DDPM-like reverse process and a reparameterized objective, ensuring efficient optimization and sampling without incurring extra computational overhead. Empirical results on three public datasets demonstrate that TDRec consistently outperforms state-of-the-art baselines, including recent diffusion models. Ablation studies confirm the effectiveness of the temporal scheduler and sequence-reduction module in generating coherent, context-aware predictions. Code is available at https://github.com/wuyaoericyy/TDRec.

Multi-interest recommendation is essential for the matching stage. By generating multiple user representations, it can better cover the diverse interests derived from user interaction history. Ideally, multi-interest models should effectively identify the underlying latent factors — the specific themes, intents, or preferences — within historical behaviors. However, conventional methods typically rely on weighted aggregation (e.g., Attention), which we argue leads to over-smoothed representations. This aggregation dilutes the intensity of significant patterns that appear only locally, blending them into a blurry average. To address this, we propose DMI, a model-agnostic diffusion framework that distills precise interests by amplifying co-occurring latent factors across behaviors. Distinct from prior diffusion works that reconstruct the single next item—which risks collapsing diverse interests—DMI reconstructs the interest vectors themselves to preserve their distributional independence. To support this, we introduce a cross-transformer module that adaptively extracts interest-specific information from designated historical interactions, transforming the diffusion process from an unconditional one into a guided, interest-disentangled pathway. In addition, we design a gradient back-propagation strategy to decouple the joint optimization of the reconstruction and recommendation losses, thereby improving training stability. Extensive offline experiments demonstrate DMI's superiority over existing methods, achieving an average relative improvement of 11.2% across all metrics on Amazon Books datasets while increasing recommendation diversity by 11.8%. Successfully deployed in a real-world recommender system, DMI effectively enhances user satisfaction and system performance at scale, serving the major traffic of hundreds of millions of daily active users.

Generative recommendation (GR) has shown great promise in industrial applications, particularly for candidate generation and end-to-end recommendations. However, existing GR training paradigms suffer from two fundamental mismatches with real-world deployment requirements. First, they optimize for point-wise prediction of a single ground-truth item, whereas practical systems must produce a diverse, high-value set of candidates. Second, they treat all user interactions as equally informative, ignoring their inherent differences in utility. Although reward-based fine-tuning offers a partial remedy, it often lacks token-level supervision. To address these challenges, we reformulate GR as a sequential set-generation problem and propose GFlowGR, a GFlowNet-based fine-tuning framework that explicitly aligns generation probabilities with item-level utilities. GFlowGR comprises three tightly integrated components, each addressing a key limitation of conventional fine-tuning: a trajectory sampler that constructs training trajectories from candidate sets to enable set-wise learning, a behavior-aware reward model that quantifies item utility to support value-aware optimization, and a GFlowNet objective that provides token-level supervision. Extensive experiments on three real-world datasets with two representative LLM-based GR backbones show consistent and significant improvements over strong baselines, validating the effectiveness of our approach. For real-world deployment, GFlowGR has been integrated into Taobao 's search advertising businesses, delivering a 0.4% relative improvement in annual revenue since its launch in mid-2025, corresponding to billion-level monetary gains. Code is available at https://github.com/Applied-Machine-Learning-Lab/SIGIR26_GFlowGR.

Generative recommendation has emerged as a transformative paradigm for capturing the dynamic evolution of user intents in sequential recommendation. While flow-based methods improve the efficiency of diffusion models, they remain hindered by the ''Noise-to-Data'' paradigm, which introduces two critical inefficiencies: prior mismatch, where generation starts from uninformative noise, forcing a lengthy recovery trajectory; and linear redundancy, where iterative solvers waste computation on modeling deterministic preference transitions. To address these limitations, we propose a Flow-based Average Velocity Establishment (Fave) framework for one-step generation recommendation that learns a direct trajectory from an informative prior to the target distribution. Fave is structured via a progressive two-stage training strategy. In Stage 1, we establish a stable preference space through dual-end semantic alignment, applying constraints at both the source (user history) and target (next item) to prevent representation collapse. In Stage 2, we directly resolve the efficiency bottlenecks by introducing a semantic anchor prior, which initializes the flow with a masked embedding from the user's interaction history, providing an informative starting point. Then we learn a global average velocity, consolidating the multi-step trajectory into a single displacement vector, and enforce trajectory straightness via a JVP-based consistency constraint to ensure one-step generation. Extensive experiments on three benchmarks demonstrate that Fave not only achieves state-of-the-art recommendation performance but also delivers an order-of-magnitude improvement in inference efficiency, making it practical for latency-sensitive scenarios. Code is available at https://github.com/Blue130/Fave

Offline reinforcement learning (RL) is a useful approach for recommender systems because it can optimize long-term user feedback from logged interaction data without online exploration. A key challenge is the multi-modal nature of user preferences: a user may like several unrelated item types, so a unimodal policy (for example, a Gaussian) tends to average across modes and generate actions that do not match any interest. Recent diffusion-based policies can model complex preference distributions, but they often require many denoising steps. We propose PerfRec (Preference-aware Flow for Recommendation), a flow-matching offline RL framework that learns an expressive behavioral policy and distills it into an efficient one-step policy. PerfRec (i) trains a conditional flow model to clone the logged action distribution, (ii) trains twin Q-networks using next actions sampled from the learned flow policy, and (iii) trains an advantage-conditioned one-step policy with Q-guidance for improvement and a distillation loss that keeps the policy close to the flow policy. We use binary advantage conditioning to separate high-advantage and low-advantage regions of the flow-induced action distribution, so that at inference we can sample from the high-advantage mode with a single forward pass. Experiments on five benchmark datasets and one online simulation platform show that PerfRec improves recommendation performance over strong offline RL baselines.

Conversational Recommender Systems (CRSs) are interactive systems that use multi-turn natural language dialogue to understand evolving user preferences and provide personalized recommendations. To achieve this goal, CRSs rely on preference elicitation strategies to actively gather informative preference cues from users; however, the timing and selection of these strategies during a conversation remain largely unexplored. While many existing studies emphasize eliciting explicit item attributes and tend to adopt relatively static elicitation strategies, the use of item-based preference elicitation and how it varies across different dialogue stages remains less explored. In this work, we conduct a systematic investigation of preference elicitation strategies from a stage-aware perspective. We provide empirical evidence that optimal preference elicitation strategies are stage-dependent and context-sensitive: attribute-based inquiries are effective in early stages, while item-based strategies become superior as preferences refine. To support this paradigm, we introduce InPE, a dataset enriched with fine-grained annotations for elicitation necessity and strategy selection. With this dataset, we propose COPE (COnversational Preference Elicitation via Mixture of Experts), a novel architecture for strategy modeling. Extensive offline evaluation on our dataset indicates that context-aware preference elicitation strategies are beneficial for conversational recommendation. In addition, the analysis of the predicted strategies uncovers consistent stage-wise tendencies in dialogue progression, providing empirical evidence of common interaction patterns in conversational recommendation systems. Our dataset is available at https://github.com/juanfacabian/InPE.

Recent years have seen a surge of research into conversational recommender systems (CRS). Among existing datasets, ReDial is the most widely used benchmark, cited in hundreds of studies. However, variations in how the dataset is preprocessed and used in experiments, particularly in the definition of ground-truth items, make it difficult to compare results across studies. These comparisons are further complicated by confounding factors such as the choice of the underlying large language model (LLM) and the use of external data sources. In this work, we revisit seven prominent CRS methods across three architectural families and evaluate them under standardized conditions. Our reproducibility study reveals a ''granularity gap,'' where fine-grained ranking (Recall@1) is highly sensitive to implementation details, while our replicability analysis shows that nearly 50% of reported accuracy stems from ''repetition shortcuts'' that are absent in novelty-focused evaluation. Furthermore, we find that performance gains are often driven more by the capacity of the LLM backbone than by specific architectural innovations. Finally, by applying user-centric utility metrics, we demonstrate that traditional recall frequently overstates a system's actual conversational effectiveness. This work establishes a transparent, controlled baseline and promotes evaluation practices that prioritize novelty and interaction efficiency.

Personalization is crucial for enhancing both the effectiveness and user satisfaction of language technologies, particularly in information-seeking tasks like question answering. Current approaches for personalizing large language models (LLMs) often rely on retrieval-augmented generation (RAG), followed by reinforcement learning with scalar reward signals to teach models how to use retrieved personal context. We believe that these scalar rewards sometimes provide weak, non-instructive feedback, limiting learning efficiency and personalization quality. We introduce VAC, a novel framework for personalized response generation that replaces scalar rewards with natural language feedback (NLF) that are generated conditioned on the user profiles and the question narratives. NLF serves as a rich and actionable supervision signal, allowing the policy model to iteratively refine its outputs and internalize effective personalization strategies. Training alternates between optimizing the feedback model and fine-tuning the policy model on the improved responses, resulting in a policy model that no longer requires feedback at inference. Evaluation on the LaMP-QA benchmark that consists of three diverse domains demonstrates consistent and significant improvements over the state-of-the-art results. Human evaluations further confirm the superior quality of the generated responses. These results demonstrate that NLF provides more effective signals for optimizing personalized question answering.

Alignment of Large Language Models (LLMs) aims to align outputs with human preferences, and personalized alignment further adapts models to individual users. This relies on personalized reward models that capture user-specific preferences and automatically provide individualized feedback. However, developing these models faces two critical challenges: the scarcity of feedback from individual users and the need for efficient adaptation to unseen users. We argue that addressing these constraints requires a paradigm shift from fitting static user models to ''learning to learn'' adaptation. To realize this, we propose Meta Reward Modeling (MRM), which reformulates personalized reward modeling as a meta-learning problem. Specifically, we represent each user's reward model as a weighted combination of base reward functions, and optimize the initialization of these weights using a Model-Agnostic Meta-Learning (MAML)-style framework to support fast adaptation under limited feedback. To ensure robustness, we introduce the Robust Personalization Objective (RPO), which places greater emphasis on hard-to-learn users during meta optimization. Extensive experiments on personalized preference datasets validate that MRM enhances few-shot personalization, improves user robustness, and consistently outperforms baselines. We release code at https://github.com/ModalityDance/MRM.

The remarkable text understanding and generation capabilities of large language models (LLMs) have revitalized the field of general recommendation based on implicit user feedback. Rather than deploying LLMs directly as recommendation models, a more flexible paradigm leverages their ability to interpret users' historical interactions and semantic contexts to extract structured profiles that characterize user preferences. These profiles can be further transformed into actionable high-dimensional representations, serving as powerful signals to augment and strengthen recommendation models. However, the mechanism by which such profiles enhance recommendation performance within the feature space remains insufficiently understood. Moreover, existing studies predominantly rely on nonlinear alignment and fusion strategies to incorporate these profiles, which often lead to semantic loss and fail to fully exploit their potential. To address these limitations, we revisit profiles from a retrieval perspective and propose a simple yet effective recommendation framework built upon distribution shaping (ProMax) in this paper. We begin by employing dense retrieval to uncover the collaborative relationships between user and item profiles within the feature space. Based on this insight, we introduce a dual distribution-reshaping process, in which the profile distribution acts as a guiding signal to steer the recommendation model toward learning user preferences for unseen items beyond the scope of observed interactions. We apply ProMax to four classic recommendation methods on three public datasets. The results indicate that ProMax substantially improves base model performance and outperforms existing LLM-based recommendation approaches.

\beginabstract Bundle recommender system is divided into two main stages: bundle editing and bundle recommendation. While substantial research progress has been made in each stage, in practical application scenarios, bundle compositions and the final recommended bundles mutually influence each other: the continuously editing bundle compositions affect the recommendation results, while user feedback on recommended bundles in turn guides the refinement of bundle compositions. This paper presents the first comprehensive reproducibility study of the complete bundle recommendation pipeline. We implement eight bundle-level editing methods, nine item-level editing methods, and seven state-of-the-art bundle recommendation models, and evaluate their performance across six real-world datasets. Our empirical analysis reveals several key findings. First, bundle-level editing faces the challenge of generating high-quality bundles. Second, in the item-level editing, the replacement operation emerges as a universal bottleneck across all methods. Third, in the recommendation stage, recommendation models exhibit varying performance across different interaction density scenarios (e.g., cold-start). Finally, bundle recommendation suffers degraded performance when integrating item-level editing and bundle recommendation within a unified pipeline. Overall, there is the systemic limitation of bundle recommendation: prior work has focused on optimizing individual stages independently, disregarding the interdependencies throughout the entire recommendation system. These findings highlight the urgent need to develop end-to-end solutions that can holistically address the bundle editing and recommendation workflow. Our repository is now available for public access via https://github.com/anyr123/Bundle_Edit_Rec_SIGIR26. \endabstract

Temporal relation classification is the task of determining the temporal relation between pairs of temporal entities in a text. Despite recent advancements in natural language processing, temporal relation classification remains a considerable challenge. Early attempts framed this task using a comprehensive set of temporal relations between events and temporal expressions. However, due to the task complexity, datasets have been progressively simplified, leading recent approaches to focus on the relations between event pairs and to use only a subset of relations. In this work, we revisit the broader goal of classifying interval relations between temporal entities by considering the full set of relations that can hold between two time intervals. The proposed approach, Interval from Point, involves first classifying the point relations between the endpoints of the temporal entities and then decoding these point relations into an interval relation. Evaluation on the TempEval-3 dataset shows that this approach can yield effective results, achieving a temporal awareness score of $70.1$ percent, a new state-of-the-art on this benchmark.

Legal information retrieval demands high precision, yet traditional ''Retrieve-then-Rerank'' pipelines with two separate models suffer from cascading error propagation and knowledge disconnects between stages. To address these issues, we propose RULER, a Robust Unified LLM-based Efficient Retrieval that integrates efficient Bi-Encoder retrieval and high-precision Cross-Encoder reranking within a parameter-sharing architecture. To mitigate the Phantom Hits problem that irrelevant documents are assigned unreasonably high confidence, we introduce a Distribution-Robust Data Construction strategy that explicitly simulates pure-negative candidate groups. This is coupled with a Dynamic Margin Ranking Objective and Maximum Entropy Regularization, which collectively enforce uncertainty on irrelevant samples and enhance robustness. Extensive experiments on the JuDGE and LeCaRDv2 benchmarks demonstrate that RULER achieves state-of-the-art performance, outperforming all independent retrievers in retrieval tasks and surpassing competing unified architectures—where retriever and reranker parameters are shared—in high-precision reranking.

Table Retrieval (TR) has traditionally been formulated as an ad-hoc retrieval problem, where relevance is primarily determined by topical semantic similarity. With the growing adoption of LLMs-based agentic systems, access to structured data is increasingly instruction-driven, where relevance is conditional on explicit content and schema constraints rather than topical similarity alone. We therefore formalize Instruction-Following Table Retrieval (IFTR), a new task that requires models to jointly satisfy topical relevance and fine-grained instruction constraints. We identify two core challenges in IFTR: (i) sensitivity to content scope, such as inclusion and exclusion constraints, and (ii) awareness of schema-grounded requirements, including column semantics and representation granularity—capabilities largely absent in existing retrievers. To support systematic evaluation, we introduce FollowTable, the first large-scale benchmark for IFTR, constructed via a taxonomy-driven annotation pipeline. We further propose a new metric, termed the Instruction Responsiveness Score , to evaluate whether retrieval rankings consistently adapt to user instructions relative to a topic-only baseline. % Our results indicate that existing retrieval models struggle to follow fine-grained instructions over tabular data. In particular, they exhibit systematic biases toward surface-level semantic cues and remain limited in handling schema-grounded constraints, highlighting substantial room for future improvements. Our benchmark is released publicly. https://github.com/AriKing11/FollowTable_Benchmark

Tabular data represents a major source of structured knowledge for open-domain question answering (QA) and enterprise data lakes, yet effective table retrieval remains challenging due to the structure–semantics gap imposed by tabular layouts. Heuristic table linearization often leads to semantic loss, particularly for implicit queries, while recent supervised retrieval models (e.g., Birdie, Contr) rely heavily on large-scale query logs and labeled QA pairs, limiting their applicability in Day-0 cold-start scenarios. We propose GeCo-TR (Generative Schema and Contrastive Table Retrieval), a zero-shot table retrieval framework that eliminates the need for supervised QA data by shifting from direct query-to-table learning to modeling the intrinsic structural semantics of the table corpus. GeCo-TR introduces UHMI, a unified hybrid representation that integrates table structure with linked knowledge graph entities, and employs a hybrid neural–symbolic retrieval mechanism that dynamically combines dense semantic retrieval, symbolic graph traversal, and sparse lexical matching. This design enables robust semantic generalization while enforcing explicit structural constraints, resulting in high-precision and high-recall retrieval for implicit queries in a zero-shot setting. Extensive experiments on public benchmarks demonstrate the effectiveness of GeCo-TR; notably, on Open-WikiTable, it achieves 97.5% Recall@5 in the zero-shot setting, ranking second among all evaluated methods despite requiring no query-level supervision.

Content-based research paper recommendation (CbRPR) has seen advances in computer science and biomedicine, but remains unexplored for mathematics, where paper relatedness is more conceptual than explicit textual or citation-based similarity. Mathematics papers may be connected through shared proof techniques, logical implications, or natural generalizations, yet exhibit minimal textual or citation overlap, rendering existing CbRPR ineffective. To address this gap, we first conduct an expert-driven study characterizing mathematical recommendations, revealing that relevance is inherently aspect -driven. Grounded in this insight, we introduce GoldRiM (small, expert-annotated) and SilverRiM (large, automatically derived), the first datasets for aspect -aware CbRPR in mathematics. Recognizing that LLM embeddings of mathematical content alone yield suboptimal representation, we propose AchGNN, an aspect -conditioned heterogeneous GNN that jointly models textual semantics, citation structure, and author lineage. Across GoldRiM and SilverRiM, AchGNN consistently outperforms prior aspect -based CbRPR methods, achieving substantial gains across all evaluated aspects. We conduct ablation studies to analyze the contributions of individual aspect supervision, authorship lineage, and graph-structural signals to AchGNN's performance. To assess domain generality, we further evaluate AchGNN on the Papers with Code dataset of machine learning publications, demonstrating that our aspect -aware approach effectively transfers beyond mathematics. We deploy our system on the MaRDI platform to help mathematicians with recommendations and release datasets and code publicly for reproducibility: github.com/gipplab/MathAspectRecSys.

Keyphrase extraction aims to automatically identify concise phrases that effectively represent the content of a document. While recent methods leveraging pre-trained language models (PLMs) have significantly improved the extraction of keyphrases with strong global semantic relevance, they often fall short in capturing the local contextual importance of keyphrases tied to specific subtopics dispersed in a document. In this paper, we propose a novel context-centric framework, MUDY, that effectively captures multi-granular contextual salience of candidate keyphrases. MUDY employs two complementary components: (1) a prompt-based scoring that estimates the generation likelihood of each candidate keyphrase, augmented with candidate-aware weighting to better reflect its local contextual importance, and (2) a self-attention-based scoring that utilizes multi-granular attention patterns from PLMs to assess candidate significance at both the document-wide and segment-specific levels. Evaluations on four real-world datasets demonstrate that MUDY outperforms state-of-the-art baselines in top-k accuracy at various cutoff thresholds. In-depth quantitative and qualitative analyses further highlight the efficacy of context-centric keyphrase extraction with multi-granular saliency. For reproducibility, the source code of MUDY is available at https://github.com/HgKang1/MUDY.

Automated fact-checking aims to verify the veracity of claims based on related evidence, and has become increasingly important as large language models (LLMs) make it easier to generate and disseminate misinformation at scale. In open settings, effective fact-checking requires models to iteratively retrieve relevant evidence and reason over noisy and incomplete information. While recent LLM-based approaches have shown promising reasoning capabilities, prompt-based methods remain limited by the inherent behaviors of base LLMs, and supervised fine-tuning methods typically require costly annotated reasoning trajectories. In this paper, we propose R3Check, a rule-guided reinforcement learning framework that enables LLMs to perform iterative retrieval–reasoning for multi-hop fact-checking. R3Check formulates the retriever as an external environment and optimizes the model using Group Relative Policy Optimization, relying only on final veracity labels and format-based rewards rather than explicit reasoning annotations. To mitigate the mutual interference between retrieval and reasoning that arises under joint training, we introduce a two-stage curriculum that first trains structured reasoning under closed fact-checking with gold evidence, and then jointly optimizes retrieval and reasoning with real-time retrieval. An importance-based sampling strategy further strengthens effective supervision signals during training. Despite using only a 7B backbone, R3Check outperforms existing baselines and even powerful reasoning LLMs, under both given-evidence and real-time retrieval settings, while producing interpretable reasoning chains. This work demonstrates the potential of pure reinforcement learning to induce effective retrieval–reasoning behaviors for fact-checking under weak supervision.

Misinformation spreading over the Internet poses a significant threat to both societies and individuals, necessitating robust and scalable fact-checking that relies on retrieving accurate and trustworthy evidence. Previous methods rely on semantic and social-contextual patterns learned from training data, which limits their generalization to new data distributions. Recently, Retrieval Augmented Generation (RAG) based methods have been proposed to utilize the reasoning capability of LLMs with retrieved grounding evidence documents. However, these methods largely rely on textual similarity for evidence retrieval and struggle to retrieve evidence that captures multi-hop semantic relations within rich document contents. These limitations lead to overlooking subtle factual correlations between the evidence and the claims to be fact-checked during evidence retrieval, thus causing inaccurate veracity predictions. To address these issues, we propose a Web-enhanced Knowledge Graph retrieval Fact-Checking agentic framework (WKGFC), which exploits authorized open knowledge graph as a core resource of evidence. LLM-enabled retrieval is designed to assess the claims and retrieve the most relevant knowledge subgraphs, forming structured evidence for fact verification. To augment the knowledge graph evidence, we retrieve web contents for completion. The above process is implemented as an automatic Markov Decision Process (MDP): A reasoning LLM agent decides what actions to take according to the current evidence and the claims. To adapt the MDP for fact-checking, we use prompt optimization to fine-tune the agentic LLM. Our extensive experiments over datasets in three categories (Wikipedia, websites, and article summaries) show that WKGFC outperform several advanced state-of-the-art fact-checking methods in balanced accuracy score by over 5% absolute. These results highlight the effectiveness of knowledge-centric evidence retrieval for fact-checking under open-world settings.

Entity Matching (EM) is a fundamental task in data integration, traditionally studied over structured data such as tables and knowledge graphs. In modern repositories, real-world objects are often represented across multiple modalities, including \emph{structured entities} with symbolic attributes and \emph{visual entities} in image-centric collections. This motivates cross-modal entity matching, which aims to identify visual-structured entity pairs that refer to the same object. Existing methods typically rely on pretrained vision-language models to compute entity pair similarity and derive correspondence via local ranking, which, however, can be unreliable given noisy and ambiguous cross-modal data and may produce globally inconsistent correspondences across related entities. Multimodal large language models (MLLMs) offer richer cross-modal cues for matching, but exhaustive MLLM reasoning over large candidate spaces is prohibitively expensive. To address these limitations, in this work, we propose \ourq, a blackboard-based multi-agent framework for zero-shot cross-modal entity matching that \emph{performs iterative, self-correcting refinement by fusing multiple matching signals, explicitly regulating global consistency, and selectively invoking an MLLM only for hard cases}. We further construct two new benchmarks from real-world visual and structured data. Extensive experiments show that \ourq consistently outperforms competitive baselines, providing an effective solution for zero-shot cross-modal entity matching. We release our data and code at \url{https://github.com/Q-17/CoMCo}.

Entity matching is a fundamental task in a wide range of retrieval and knowledge applications, aiming to identify whether two objects correspond to the same real-world entity across heterogeneous sources. Typical variants include entity resolution (ER), entity linking (EL), and entity alignment (EA). While recent unified matchers have made progress through multi-task training with comprehensive annotations, real-world pipelines often operate under scarce supervision, where labeled data is incomplete and fails to cover the full spectrum of matching scenarios. In this regime, supervised unified models degrade substantially, and deployable compact LLMs remain unreliable: lightweight fine-tuning and in-context learning yield inconsistent behavior and can even exhibit negative effects under scenario shifts. To fill in this gap, we propose øurs, a meta-rule induction and retrieval framework for unified entity matching under scarce supervision. Instead of relying on parametric adaptation, øurs converts limited supervision into explicit natural-language rules, abstracts them into reusable meta-rules via hierarchical clustering, and retrieves the most relevant meta-rules to guide the LLM's inference for each input instance. This design improves robustness by grounding decisions on explicit and reusable evidence, instead of relying solely on implicit adaptation or prompt demonstrations. Extensive experiments show that øurs achieves state-of-the-art performance on unified entity matching under scarce supervision.

False claims about medical information can have damaging consequences. Although LLMs have been used to verify such claims, their success depends on access to accurate content and robust reasoning capabilities. Prior research assumes access to gold evidence during inference or relies on expensive, compute-intensive scaling strategies. However, these approaches fall short when applied to real-world verification tasks, where relevant evidence is hard to obtain and efficiency is crucial. To address this, we introduce SciCheck, a novel solution that integrates web evidence retrieval with a process of reasoning distillation. Our approach fine-tunes a small language model using reasoning traces generated by an LLM. The distillation process involves a data preparation pipeline that avoids data leakage and filters reasoning traces to retain only those leading to correct answers. It also combines web and gold evidence during training to improve robustness, while evaluation is performed with web retrieval only. We performed an extensive experimental evaluation on different claim verification datasets. The results demonstrate that SciCheck outperforms competing approaches and proprietary LLMs such as Gemini 2.5 Flash in most scenarios with lower computational cost.

Current neural biomedical entity linking (BioEL) models treat ontologies as flat dictionaries, ignoring the rich terminological knowledge (TBox) that defines concept boundaries. Consequently, they struggle with contextual ambiguity, often retrieving logically inconsistent candidates based solely on surface similarity. We present OntoEL, a neuro-symbolic framework that shifts BioEL from surface-level matching to logic-grounded reasoning. OntoEL integrates differentiable fuzzy EL⊥ reasoning into the retrieval pipeline as a consistency-aware re-ranker, employing a hybrid strategy: structural TBox reasoning is delegated to classical polynomial-time reasoners, while the sigmoidal Reichenbach implication performs soft type-consistency evaluation, effectively resolving the "implication bias" gradient pathology in previous neuro-symbolic methods. By enforcing ontological axioms as differentiable soft constraints, OntoEL aligns neural representations with logical truth. Comprehensive experiments on three benchmarks (MedMentions, BC5CDR, and NCBI Disease) demonstrate state-of-the-art performance, surpassing strong baselines by up to 4.2% in Accuracy@1. On highly ambiguous mentions requiring ontological reasoning, our method corrects 71.2% of retrieval errors, proving the efficacy of incorporating logical semantics into neural retrieval.

Recent advancements in information retrieval have highlighted the potential of integrating visual and textual information, yet effective reranking for image-text documents remains challenging due to the modality gap and scarcity of aligned datasets. Meanwhile, existing approaches often rely on large models (7B--32B parameters) with reasoning-based distillation, incurring unnecessary computational overhead while primarily focusing on textual modalities. In this paper, we propose Rank-Nexus, a multimodal image-text document reranker that performs listwise qualitative reranking on retrieved lists incorporating both images and texts. To bridge the modality gap, we introduce a progressive cross-modal training strategy. We first train each modality separately: leveraging abundant text reranking data, we distill ranking knowledge into the text branch using GPT-4o as the teacher model. For images, where labeled data is scarce, we construct distilled pairs from multimodal large language model (MLLM) captions on image retrieval benchmarks. Subsequently, we train on a joint image-text reranking dataset. Rank-Nexus achieves outstanding performance on text reranking benchmarks (TREC, BEIR) and the challenging image reranking benchmarks (INQUIRE, MMDocIR), using only a lightweight 2B pretrained visual-language model. This efficient design ensures strong generalization across diverse multimodal scenarios without excessive parameters or reasoning overhead.

Dense retrievers bridge vocabulary gaps but suffer from semantic drift, ranking topically similar passages that are not actually relevant to the query. Sparse retrievers like BM25 are generally less prone to such false positives but miss paraphrases. We observe that these failure modes are complementary, with relevant passages concentrated in regions where both retrievers score highly, while single-retriever confidence proves unreliable. Building on this insight, we propose AgreRank, a re-ranking method that leverages sparse-dense consensus for query expansion. AgreRank identifies anchor documents where both retrievers agree, expands the query through these verified anchors, and applies geometric gating to suppress drift from the original intent. On TREC Deep Learning 2021 and 2022, AgreRank improves over Dense+CE by 3.4--5.7% and over RRF+CE by 3.7--5.6% under identical retrieval and reranking components, achieving nDCG@10 of 0.706 and 0.627 with a lightweight 33M-parameter cross-encoder. These gains require no additional parameters or training, demonstrating that more effective use of retrieval signals offers a practical alternative to model scaling.

Zero-shot document re-ranking with Large Language Models (LLMs) has evolved from Pointwise methods to Listwise and Setwise approaches that optimize computational efficiency. Despite their success, these methods predominantly rely on generative scoring or output logits, which face bottlenecks in inference latency and result consistency. In-Context Re-ranking (ICR) has recently been proposed as an O(1) alternative method. ICR extracts internal attention signals directly, avoiding the overhead of text generation. However, existing ICR methods simply aggregate signals across all layers; layer-wise contributions and their consistency across architectures have been left unexplored. Furthermore, no unified study has compared internal attention with traditional generative and likelihood-based mechanisms across diverse ranking frameworks under consistent conditions. In this paper, we conduct an orthogonal evaluation of generation, likelihood, and internal attention mechanisms across multiple ranking frameworks. We further identify a universal ''bell-curve'' distribution of relevance signals across transformer layers, which motivates the proposed Selective-ICR strategy that reduces inference latency by 30%--50% without compromising effectiveness. Finally, evaluation on the reasoning-intensive BRIGHT benchmark shows that precisely capturing high-quality in-context attention signals fundamentally reduces the need for model scaling and reinforcement learning: a zero-shot 8B model matches the performance of 14B reinforcement-learned re-rankers, while even a 0.6B model outperforms state-of-the-art generation-based approaches. These findings redefine the efficiency-effectiveness frontier for LLM-based re-ranking and highlight the latent potential of internal signals for complex reasoning ranking tasks. Our code and results are publicly available at https://github.com/ielab/Selective-ICR.

Large language models (LLMs) are increasingly applied to ranking tasks in retrieval and recommendation. Although reasoning prompting can enhance ranking utility, our preliminary exploration reveals that its benefits are inconsistent and come at a substantial computational cost, suggesting that when to reason is as crucial as how to reason. To address this issue, we propose a reasoning routing framework that employs a lightweight, plug-and-play router head to decide whether to use direct inference (Non-Think) or reasoning (Think) for each instance before generation. The router head relies solely on pre-generation signals: i) compact ranking-aware features (e.g., candidate dispersion) and ii) model-aware difficulty signals derived from a diagnostic checklist reflecting the model's estimated need for reasoning. By leveraging these features before generation, the router outputs a controllable token that determines whether to apply the Think mode. Furthermore, the router can adaptively select its operating policy along the validation Pareto frontier at deployment time, enabling dynamic allocation of computational resources toward instances most likely to benefit from Think under varying system constraints. Experiments on three public ranking datasets with different scales of open-source LLMs show consistent improvements in ranking utility with reduced token consumption (e.g., +6.3% NDCG@10 with –49.5% tokens on MovieLens with Qwen3-4B), demonstrating reasoning routing as a practical solution to the accuracy-efficiency trade-off.

Deep research has emerged as an important task that aims to address hard queries that need extensive open-web exploration. To tackle it, most prior work equips large language model (LLM)-based agents with opaque web search APIs, enabling agents to iteratively issue search queries, retrieve external evidence, and reason over it. Despite search's essential role in deep research, black-box web search APIs leave the behaviour of established text ranking methods in deep research largely unclear. To fill this gap, we reproduce key findings and best practices for text ranking methods in deep research. We examine their effectiveness from three perspectives: (i) retrieval units (documents vs. passages), (ii) pipeline configurations (different retrievers, re-rankers, and re-ranking depths), and (iii) query characteristics (the mismatch between agent-issued queries and the training queries of text rankers). We perform experiments on BrowseComp-Plus, a deep research dataset with a fixed corpus, evaluating 2 open-source agents, 5 retrievers, and 3 re-rankers. We find that agent-issued queries typically follow web-search-style syntax (e.g., quoted exact matches), favouring lexical, learned sparse, and multi-vector retrievers; passage-level units are more efficient under limited context windows, and avoid the difficulties of document length normalisation in lexical retrieval; re-ranking is highly effective. We further propose a query-to-question (Q2Q) method that translates agent-issued queries into natural-language questions, significantly reducing the query mismatch.

The classical cascading pipeline of retrieve--rerank suffers from a bounded recall problem, stemming from limitations of the first-stage retriever. Most current approaches address the bounded recall problem by improving the first-stage retriever, but this incurs substantial training and inference costs, especially to handle queries that require substantial reasoning. To circumvent the computational costs of reasoning-based retrievers, we replicate the findings of GAR, Graph-based Adaptive Reranking, on the BRIGHT reasoning-intensive retrieval benchmark. GAR addresses the bounded recall problem by modifying the reranking process itself through iterative exploration of a corpus graph, but it was previously only tested on models designed for topical and question-answering-style queries. Hence, reproduce GAR in reasoning-intensive settings with reasoning and non-reasoning reranking models. We observe that the quality of the reranker's signal plays an important role in identifying additional relevant documents within the corpus graph. Overall, we find that GAR boosts the effectiveness of reasoning-intensive retrieval across a variety of models while contributing minimally to computational overheads. Ultimately, this work enables more practical deployment of retrieval systems that can address reasoning-intensive queries.

Retrieval-augmented generation (RAG) has shown strong potential for deploying large language models at the edge, yet existing designs largely rely on generic and monolithic knowledge bases that are poorly matched to the heterogeneous queries and resource-constrained edge computing environments. Through extensive empirical analysis, we find that domain-specialized knowledge bases, when deployed on individual edge servers, deliver substantially higher retrieval accuracy and generation quality than generic knowledge bases under identical resource budgets. Based on this, we propose SRAG, a distributed RAG system that enforces knowledge specialization at the edge. Each edge server maintains a domain-aware specialized knowledge base by retaining domain-aligned knowledge and decoupling out-of-domain content. SRAG uses a buffer-based knowledge migration mechanism to redistribute out-of-domain content to better-matched edge servers, enabling efficient global knowledge utilization without central coordination. To handle domain-mismatched queries, SRAG employs lightweight cross-node routing guided by compact metadata summaries, avoiding full knowledge replication. Together, these mechanisms form an end-to-end workflow for decentralized edge RAG. Experiments show that SRAG improves retrieval relevance, generation quality, and storage efficiency, while reducing end-to-end latency.

Retrieval-Augmented Generation (RAG) techniques have emerged as a promising direction to merge the non-parametric knowledge into Large Language Models (LLMs), thereby alleviating factual errors, hallucinations and outdated knowledge. Existing RAG methods, which append multiple retrieved documents or passages to the input of LLMs, will inevitably increase the context length, resulting in not only significant computational overhead and inference latency, but also performance degradation. Although reranking or compression modules have been introduced to address these challenges, they overlook the contextual preferences of the generative LLMs itself and may inadvertently discard information that is crucial for generation accuracy. To this end, we introduce InnerRAG, which incentivizes RAG via Inner Adaptive Context Selection. InnerRAG is a novel paradigm that empowers LLMs to autonomously select relevant context during generation. Our proposed InnerRAG endows the model to accurately identify the documents that are most helpful for generation from long contexts. By endowing the model with this capability, InnerRAG facilitates more effective exploitation of external knowledge without being misled by disturbed information, leading to substantial improvements in generation quality while maintaining computational efficiency. Extensive experiments across multiple benchmarks and human evaluations demonstrate that our method consistently outperforms state-of-the-art RAG baselines. Moreover, our framework is orthogonal and complementary to in-context RAG approaches, offering further performance improvements when combined.

Recent work has suggested that adding irrelevant documents to the input of retrieval-augmented generation (RAG) systems can improve question-answering performance, a phenomenon referred to as the '' Power of Noise. '' This motivated investigations into the role of noise in information retrieval. In this paper, we reproduce the main findings of Cuconasu et al. \citecuconasu2024power and evaluate the robustness of the effect under extended experimental settings. We first confirm that the phenomenon holds under the original setup, which uses earlier-generation LLMs, restrictive prompting and constrained decoding settings. We subsequently introduce a series of extensions to investigate the underlying causes of the noise effect, examining the authors' original design choices including the use of different models, instruction prompting, and relaxed output length constraints. Across these ablations, the Power-of-Noise pattern proves highly sensitive to inference configuration: it can appear, weaken, or disappear under small changes to prompt formulation and decoding limits. Combined with our error analysis, which shows substantial contributions from truncation and malformed generations, this variance indicates that the original effect cannot be robustly confirmed as a general benefit of noisy retrieval under these experimental conditions. More broadly, our work highlights the importance of carefully scrutinizing inference design in retrieval-augmented generation systems. Our code is available at https://github.com/ina0105/The-Power-of-Noise-Reproduction.

Retrieval-augmented generation (RAG) systems are frequently evaluated via fact-based metrics, yet standard implementations retrieve passages or static propositions. This unit mismatch between evaluation and retrieval objects hinders maintenance when corpora evolve and fails to capture superseded facts or source disagreements. We propose NuggetIndex, a retrieval system that stores atomic information units as managed records, so called nuggets. Each record maintains links to evidence, a temporal validity interval, and a lifecycle state. By filtering invalid or deprecated nuggets prior to ranking, the system prevents the inclusion of outdated information. We evaluate the approach using a nuggetized MS MARCO subset, a temporal Wikipedia QA dataset, and a multi-hop QA task. Against passage and unmanaged proposition retrieval baselines, NuggetIndex improves nugget recall by 42%, increases temporal correctness by 9 percentage points without the recall collapse observed in time-filtered baselines, and reduces conflict rates by 55%. The compact nugget format reduces generator input length by 64% while enabling lightweight index structures suitable for browser-based and resource-constrained deployment. We release our implementation, datasets, and evaluation scripts.

Retrieval-Augmented Generation (RAG) enhances Large Language Models (LLMs) by grounding generation in external knowledge to improve factuality and reduce hallucinations. Yet most deployments assume a centralized corpus, which is infeasible in privacy-aware domains where knowledge remains siloed. This motivates federated RAG (FedRAG), where a central LLM server collaborates with distributed silos without sharing raw documents. In-context RAG violates this requirement by transmitting verbatim documents, whereas parametric RAG encodes documents into light weight adapters that merge with a frozen LLM at inference, avoiding raw-text exchange. We adopt the parametric approach but face two unique challenges induced by FedRAG: high storage and communication from per-document adapters, and destructive aggregation caused by indiscriminately merging multiple adapters. We present FedMosaic, the first federated RAG framework built on parametric adapters. FedMosaic clusters semantically related documents into multi-document adapters with document-specific masks to reduce overhead while preserving specificity, and performs selective adapter aggregation to combine only relevance-aligned, non-conflicting adapters. Experiments show that FedMosaic achieves an average 10.9% higher accuracy than state-of-the-art methods in four categories, while lowering storage costs by 78.8% to 86.3% and communication costs by 91.4%, and never sharing raw documents.

\begin{abstract}For information access systems, it is not enough for outputs or answers to be relevant or correct; they must also be \emph{permitted}. This paper highlights a research gap: non-disclosure obligations often concern \emph{propositions} that must not be stated, while deployed Retrieval-Augmented Generation (RAG) systems enforce restrictions through record-level handling such as document removal or access-control lists. Such systems can appear compliant while still disclosing revoked facts in generated answers. Most RAG evaluations fail to assess this important aspect of compliance. We explore the nature of this gap and introduce \emph{Forgetting-by-Design (FBD)}, a mechanism-agnostic audit that runs probes across paired system states before and after revocation. \emph{FBD} separates compliance into two observable channels---retrieval/citation \emph{exposure} and answer-level \emph{disclosure or abstention}---and reports the cost of compliance using matched lawful controls and substitution-aware utility signals. We instantiate \emph{FBD} in a reproducible RAG setting and show how the resulting report card reveals failures that single metrics miss: retrieval exposure can be suppressed while answer-level leakage persists, and interventions that reduce disclosure can still degrade lawful utility or collapse citation coverage.\end{abstract}

Multimodal recommendation systems have achieved success in capturing user interests, yet user interests are inherently diverse. Existing GCN-based methods adopt a uniform approach to model user-item interaction graphs, failing to differentiate between ''popular interests'' and ''niche interests''. This deficiency leads to problems such as the information of popular items overshadowing that of niche items during message passing and multimodal fusion, thereby compromising the accuracy and diversity of recommendations. To address these problems, we propose a novel Popular-Niche Graph Wavelet Learning Framework for Multimodal Recommendation (PNGRec). PNGRec first extracts user behavior signals and item popularity signals from the user-item interaction graph, then decomposes the frequency-domain signals of interests into a Popular User-Item Graph and a Niche Graph via frequency-domain signal decomposition, aiming to capture users' popular and niche interests respectively. Furthermore, we introduce a User Interests Four Quadrants Graph Learning mechanism, which enhances the diversity of users' interests under multimodal perception by finely dividing user behavior representations and item popularity representations. Finally, the model is optimized through a self-supervised multi-task joint optimization loss. Extensive experiments on four real-world industrial datasets demonstrate the effectiveness of our proposed interest-divided modeling approach. The source code is publicly available at https://github.com/orangeheyue/PNGRec.

Multimodal recommendation benefits from leveraging rich content signals such as images and texts to alleviate interaction sparsity, yet existing graph-based approaches are still hindered by (i) noisy user—item edges that are treated as static during training and (ii) inconsistent representation spaces across interaction-driven and modality-induced graph views. To address these issues, we propose DIGEST, a multi-graph framework that propagates trainable ID embeddings on a denoised user—item graph and a fused modality-induced item—item graph, and interleaves message passing with dynamic graph refinement that iteratively reweights existing edges to suppress noisy connections. To enable reliable semantic transfer across views, DIGEST further introduces a dual contrastive alignment that (i) aligns the collaborative and semantic item views and (ii) constrains the semantic graph representations to projected multimodal features, together with a lightweight dimension decorrelation regularizer and adaptive gated fusion to reduce redundancy and stabilize multi-view learning. Extensive experiments on three Amazon benchmark datasets demonstrate that DIGEST consistently outperforms state-of-the-art multimodal recommenders, achieving up to 8.43% relative improvement on NDCG@20 and 7.66% on Recall@20 over the strongest baselines.

Multimodal recommendation combines the user historical behaviors with the modal features of items to capture the tangible user preferences, presenting superior performance compared to the conventional ID-based recommender systems. However, existing methods still encounter two key problems in the representation learning of users and items, respectively: (1) the initialization of multimodal user representations is either agnostic to historical behaviors or contaminated by irrelevant modal noise, and (2) the widely used KNN-based item-item graph contains noisy edges with low similarities and lacks audience co-occurrence relationships. To address such issues, we propose MLLMRec, a novel preference reasoning paradigm with graph refinement for multimodal recommendation. Specifically, on the one hand, the item images are first converted into high-quality semantic descriptions using a multimodal large language model (MLLM), thereby bridging the semantic gap between visual and textual modalities. Then, we construct a behavioral description list for each user and feed it into the MLLM to reason about the purified user preference profiles that contain the latent interaction intents. On the other hand, we develop the threshold-controlled denoising and topology-aware enhancement strategies to refine the suboptimal item-item graph, thereby improving the accuracy of item representation learning. Extensive experiments on three publicly available datasets demonstrate that MLLMRec achieves the state-of-the-art performance with an average improvement of 21.48% over the optimal baselines. The source code is provided at https://github.com/Yuzhuo-Dang/MLLMRec.

We study the problem of inferring substitutable and complementary items, which underpins applications such as alternative and follow-up purchase suggestions. Existing approaches typically learn from behavior-derived item-item associations using GNNs or leverage item content alone. However, these methods often overlook two key challenges: (i) user behaviors (e.g., co-view/co-purchase) only provide noisy weak supervision, and (ii) behavior signals are long-tailed, leaving many items with sparse associations. We propose MMSC, a self-supervised multi-modal relational representation learning framework that combines a multi-modal foundation model adapted to encode item metadata and a self-supervised denoising module that learns relationship-aware representations from noisy user behaviors, unified by a hierarchical aggregation mechanism. We further use LLM-assisted supervision to mitigate noise in behavior-derived supervision during training. Experiments on five real-world datasets show that MMSC consistently outperforms existing baselines by 26.1% for substitutable and 39.2% for complementary item inference, while remaining effective for cold-start items.

Recent advancements in multimodal recommendations, which leverage diverse modality information to mitigate data sparsity and improve recommendation accuracy, have gained significant attention. However, existing multimodal recommendations overlook the critical role of user representation initialization. Unlike items, which are naturally associated with rich modality information, users lack such inherent information. Consequently, item representations initialized based on meaningful modality information and user representations initialized randomly exhibit a significant semantic gap. To this end, we propose a Semantically Guaranteed User Representation Initialization (SG-URInit). SG-URInit constructs the initial representation for each user by integrating both the modality features of the items they have interacted with and the global features of their corresponding clusters. SG-URInit enables the initialization of semantically enriched user representations that effectively capture both local (item-level) and global (cluster-level) semantics. Our SG-URInit is training-free and model-agnostic, meaning it can be seamlessly integrated into existing multimodal recommendation models without incurring any additional computational overhead during training. Extensive experiments on multiple real-world datasets demonstrate that incorporating SG-URInit into advanced multimodal recommendation models significantly enhances recommendation performance. Furthermore, the results show that SG-URInit can further alleviate the item cold-start problem and also accelerate model convergence, making it an efficient and practical solution for multimodal recommendations.

Multimodal recommendation improves user modeling by integrating collaborative signals with heterogeneous item content. In real applications, user interests evolve over time and exhibit nonstationary dynamics, where different preference factors change at different rates. This challenge is amplified in multimodal settings because visual and textual cues can dominate decisions under different temporal regimes. Despite strong progress, most multimodal recommenders still rely on static interaction graphs or coarse temporal heuristics, which limits their ability to model continuous preference evolution with fine-grained temporal adaptation. To address these limitations, we propose TimeMM, a time-conditioned spectral filtering framework for dynamic multimodal recommendation. TimeMM instantiates Time-as-Operator by mapping interaction recency to a family of parametric temporal kernels that reweight edges on the user--item graph, producing component-specific representations without explicit eigendecomposition. To capture non-stationary interests, we introduce Adaptive Spectral Filtering that mixes the operator bank according to temporal context, yielding prediction-specific effective spectral responses. To account for modality-specific temporal sensitivity, we further propose Spectral-Aware Modality Routing that calibrates visual and textual contributions conditioned on the same temporal context. Finally, a ranking-space Spectral Diversity Regularization encourages complementary expert behaviors and prevents filter-bank collapse. Extensive experiments on real-world benchmarks demonstrate that TimeMM consistently outperforms state-of-the-art multimodal recommenders while maintaining linear-time scalability.

Reasoning in Large Language Models (LLMs) has recently shown strong potential in enhancing generative recommendation through deep understanding of complex user preference. Existing approaches follow a reason-then-recommend paradigm, where LLMs perform step-by-step reasoning before item generation. However, this paradigm inevitably suffers from reasoning degradation (i.e., homogeneous or error-accumulated reasoning) due to the lack of intermediate verification, thus undermining the recommendation. To bridge this gap, we propose a novel reason-verify-recommend paradigm, which interleaves reasoning with verification to provide reliable feedback, guiding the reasoning process toward more faithful user preference understanding. To enable effective verification, we establish two key principles for verifier design: 1) reliability ensures accurate evaluation of reasoning correctness and informative guidance generation; and2) multi-dimensionality emphasizes comprehensive verification across multi-dimensional user preferences. Accordingly, we propose an effective implementation called VRec. It employs a mixture of verifiers to ensure multi-dimensionality, while leveraging a proxy prediction objective to pursue reliability. Experiments on four real-world datasets demonstrate that VRec substantially enhances recommendation effectiveness and scalability without compromising efficiency.

Large language models (LLMs) have recently been adopted for recommendation by framing user preference modeling as a language generation problem. However, existing latent reasoning approaches typically represent user intent with a single latent vector, which struggles to capture the inherently multi-faceted nature of user preferences. We propose Factorized Latent Reasoning (FLR), a novel framework for LLM-based sequential recommendation that decomposes latent reasoning into multiple disentangled preference factors. FLR introduces a lightweight multi-factor attention module that iteratively refines a latent thought representation, where each factor attends to distinct aspects of the user's interaction history. To encourage diversity and specialization, we design orthogonality, attention diversity, and sparsity regularization objectives, and dynamically aggregate factor contributions for the final prediction. We further integrate FLR with an efficient reinforcement learning strategy based on group-relative policy optimization, enabling stable alignment directly in the latent reasoning space. Experiments on multiple benchmarks show that FLR consistently outperforms strong baselines while improving robustness and interpretability. Our data and code are available at https://github.com/ToAdventure/FLR.

Large Language Models (LLMs) have demonstrated strong potential in recommendations due to their powerful reasoning capabilities. However, existing methods typically rely on a single reasoning path to drive the entire Top-K recommendations. This paradigm is prone to reasoning path collapse, where limiting exploration of potentially superior and diverse reasoning paths within the LLMs space. As a result, both the accuracy and diversity of the recommendation outcomes are constrained. To address this issue, we propose a novel model, Divergent Reasoning for LLM-based Recommendation, named DivReason. Inspired by the structure of intellect theory, which emphasizes a two-stage cognitive process of divergent thinking followed by convergent thinking, DivReason is designed with two core components: the Divergent Reasoning Path Generation Module and the Reasoning Path Aggregation Module. In the first module, DivReason introduces a training-free form of controlled uncertainty to promote diverse reasoning, leveraging Monte Carlo Dropout and Directional Perturbation to expand exploration in the latent reasoning space. In the Reasoning Path Aggregation Module, we adaptively select a subset of high-quality reasoning paths from the entire path pool and aggregate them into a unified reasoning representation. Meanwhile, we further adopt an alternating reinforcement learning strategy to optimize the model, explicitly balancing accuracy and diversity during training. Extensive experimental results show that DivReason effectively mitigates the issue of reasoning path collapse, while improving both the accuracy and diversity of LLM-based recommendations.

Recent progress in large language model (LLM) based generative recommendation (GR) shows that leveraging LLM world knowledge can substantially improve performance. However, existing methods rely on fixed, manually designed instructions to generate semantic knowledge and directly incorporate it into GR, which has two limitations: (1) fixed instructions cannot capture the multidimensional heterogeneity of user interests; (2) uncontrollable knowledge fusion may conflict with behavioral signals and harm recommendations.To address these limitations, we propose \textbf{LWGR}, a framework that leverages \textbf{L}agrangian constraints to transfer users' personalized \textbf{W}orld knowledge from LLMs into \textbf{G}enerative \textbf{R}ecommendation. LWGR enhances GR along two axes: knowledge extraction and fusion. It builds user personalized soft instructions to extract behavior-relevant LLM world knowledge. Then, it formulates knowledge fusion as an optimization problem with explicitly bounded performance degradation, solved via a Lagrangian primal–dual method that selectively incorporates beneficial knowledge.We further design two training strategies for different LLM scales and a deployment scheme that combines nearline precomputation with lightweight online serving. Experiments on multiple public datasets and one industrial dataset show that LWGR outperforms eight state-of-the-art baselines by up to 11.23\% and brings a 1.35\% revenue lift on a large-scale advertising platform, demonstrating its effectiveness and practicality.

Large Language Models have revolutionized recommender systems (LLM4Rec) by leveraging their generative capabilities to model complex user preferences. However, existing LLM4Rec methods primarily rely on token-level objectives, making it difficult to optimize list-level and non-differentiable metrics (e.g., NDCG, fairness) that define actual recommendation quality. While Best-of-N (BoN) directly optimizes these metrics during inference, its high computational cost hinders real-world deployment. To address this, BoN Alignment aims to distill the search capability into the model itself, yet current approaches suffer from two critical limitations: (1) Indiscriminate Supervision, where the static reference fails to distinguish the relative quality of candidates exceeding its empirical range, leading to a loss of ranking guidance; and (2) Gradient Decay, where the effective supervision signal rapidly diminishes as the evolving policy improves, resulting in inefficient optimization. To overcome these challenges, we propose BLADE (Bayesian List-wise Alignment via Dynamic Estimation). Unlike static approaches, BLADE introduces a Bayesian framework that continuously updates the target distribution by fusing historical priors with dynamic evidence from the model's current rollouts. This mechanism constructs a self-evolving target that adapts to the model's growing capabilities, ensuring the training signal remains informative throughout the learning process. Extensive experiments on three real-world datasets demonstrate that BLADE significantly outperforms state-of-the-art baselines. Crucially, it breaks the static performance upper bound, achieving sustained gains in both ranking accuracy (Recall, NDCG) and complex list-wise metrics (Fairness, Diversity). The code is available via https://github.com/RegionCh/BLADE.

Large language models (LLMs) have transformed recommender systems through strong semantic understanding and generalization. However, the design of item identifiers remains a critical bottleneck that directly affects recommendation quality. Traditional metadata-based identifiers introduce length variability and semantic ambiguity, whereas existing collaborative indexing (CID) approaches often neglect item attributes, show limited cross-dataset generalizability, and incur high computational cost at scale. To address these limitations, we propose a Graph Neural Network (GNN)–based item indexing framework with three coordinated innovations. First, we construct attribute-enriched co-occurrence graphs and use a GNN encoder to fuse item features with collaborative signals, yielding semantically informed representations that work well for attribute-rich catalogs. Second, we replace recursive spectral clustering with hierarchical agglomerative clustering on GNN embeddings, enabling direct control of index length via tree depth and reducing hyperparameter tuning across datasets. Third, we exploit localized message passing rather than global eigendecomposition, which provides considerably better runtime efficiency and is amenable to mini-batch training, supporting online index updates as interactions evolve. Across five benchmarks, GID achieves strong average ranking performance, showing larger improvements on sparse and attribute-rich datasets while remaining competitive in dense settings. The framework is robust under both seen and unseen prompt templates, which supports practical LLM-based recommendation. On sequential recommendation, GID improves HR@10 by 7.9% on average over the strongest baseline in each dataset.

Deep hashing is widely used in large-scale image retrieval systems due to its efficient retrieval performance. However, its susceptibility to adversarial attacks limits its security in practical applications. Adversarial training is the most effective method for improving robustness, but it often leads to a significant trade-off between robustness and retrieval accuracy. In this paper, we conduct spectral analysis and find that generating high-quality hash codes requires wide-frequency response models, whereas adversarial training forces the model into spectral collapse, degrading it to a low-frequency response model and weakening its discriminability. To address this issue, we propose a Spectral-Adaptive Adversarial Hashing (SAAH) framework, which selectively preserves discriminative and task-relevant frequency components while suppressing adversarially unstable ones, enabling robust hashing without sacrificing retrieval performance. Extensive experiments on benchmark datasets demonstrate that SAAH consistently achieves a superior balance between retrieval accuracy and adversarial robustness, achieving the best performance in both retrieval accuracy and robustness compared with existing robust hashing methods.

Large language model-powered sequential recommender systems (LLM-SRSs) have recently demonstrated remarkable performance, enabling recommendations through prompt-driven inference over user interaction sequences. However, this paradigm also introduces new security vulnerabilities, particularly text-level manipulations, rendering them appealing targets for promotion attacks that purposely boost the ranking of specific target items. Although such security risks have been receiving increasing attention, existing studies typically rely on an unrealistic assumption of access to either the victim model or prompt to unveil attack mechanisms. In this work, we investigate the item promotion attack in LLM-SRSs under a more realistic setting where both the system prompt and victim model are unknown to the attacker, and propose a Prompt-Unknown Dual-poisoning Attack (PUDA) framework. To simulate attacks under this full black-box setting, we introduce an LLMbased evolutionary refinement strategy that infers discrete system prompts, enabling the training of an effective surrogate model that mimics the behaviors of the victim model. Leveraging the distilled prompt and surrogate model, we devise a promotion attack that adversarially revises target item texts under semantic constraints, which is further complemented by the highly plausible, surrogategenerated poisoning sequences to enable cost-effective target item promotion. Extensive experiments on real-world datasets demonstrate that PUDA consistently outperforms state-of-the-art competitors in boosting the exposure of unpopular target items. Our findings reveal critical security risks in modern LLM-SRSs even when both prompts and models are protected, and highlight the need for more robust defensive means.

Large Language Models (LLMs) have emerged as powerful re-rankers. Recent research has however showed that simple prompt injections embedded within a candidate document (i.e., jailbreak prompt attacks) can significantly alter an LLM's ranking decisions. While this poses serious security risks to LLM-based ranking pipelines, the extent to which this vulnerability persists across diverse LLM families, architectures, and settings remains largely under-explored. In this paper, we present a comprehensive empirical study of jailbreak prompt attacks against LLM rankers. We focus our evaluation on two complementary tasks: (1) Preference Vulnerability Assessment, measuring intrinsic susceptibility via attack success rate (ASR); and (2) Ranking Vulnerability Assessment, quantifying the operational impact on the ranking's quality (nDCG@10). We systematically examine three prevalent ranking paradigms (pairwise, listwise, setwise) under two injection variants: decision objective hijacking and decision criteria hijacking. Beyond reproducing prior findings, we expand the analysis to cover vulnerability scaling across model families, position sensitivity, backbone architectures, and cross-domain robustness. Our results characterize the boundary conditions of these vulnerabilities, revealing critical insights such as that encoder-decoder architectures exhibit strong inherent resilience to jailbreak attacks. We publicly release our code and additional experimental results at https://github.com/ielab/LLM-Ranker-Attack.

Advancements in computer science are raising concerns and preoccupations about the privacy of users' data submitted to and used by Information Retrieval (IR) systems. IR systems, such as search engines, integrate new generative information access pipelines that implement effective and efficient document retrieval and answer generation. However, critical challenges arise in Privacy-Preserving IR (PPIR): Are such data leaking personal information or being used to train generative systems? Are current privacy solutions sufficient to guarantee user privacy and limit such information leakage? Are users' queries and retrieved documents protected throughout the entire retrieval process? How has the privacy threats landscape changed, and in which directions should the IR and Privacy research community investigate to address such new risks? In this perspective paper, we provide initial answers to these questions, analysing state-of-the-art solutions for protecting user privacy when accessing information and highlighting areas of concern. We propose a new PPIR research agenda to address the unsolved problem of private data use and access. The agenda includes novel privacy research pillars aimed at addressing objectives grounded in gaps in the literature, user survey findings, and structured interviews with experts from research, industry, and regulatory bodies. By defining these privacy research pillars, we advise the IR community to pursue research toward a more resilient privacy direction that can address future challenges stemming from rapidly advancing technology eager for user data.

Synthesizing high-quality tabular data is essential for privacy-preserving data analysis. However, this task remains challenging due to two key factors: (1) distribution complexity : imbalanced and skewed data make it challenging to learn the data distribution accurately; and (2) semantic coherence : implicit relationships and logical dependencies among fields must be preserved to ensure valid and meaningful synthetic samples. To address these issues, we propose SynDiSC, a high-quality tabular data synthesis approach that enforces both distributional and semantic consistency. It comprises three core designs: (1) a distribution-aware encoding that effectively handles heterogeneous data types and complex distributions; (2) a multi-dimensional semantic conditioning that leverages multi-dimensional conditional dependencies to enforce semantic validity during generation; and (3) a conditional consistency controller that guides the generator to produce diverse samples satisfying multiple conditional constraints while mitigating mode collapse. Extensive experiments on datasets from various application domains demonstrate that SynDiSC significantly improves data quality, conditional controllability, and downstream task performance compared to state-of-the-art methods. Our code and data samples are open-sourced in the GitHub repository. https://github.com/Knightz9/SynDiSC.

Data protection laws and regulations require service providers to disclose data practices in privacy policies, specifying what personal information is processed and for what purposes. For compliance, these data practices must adhere to the data minimization principle, limiting the processing of personal information to what is directly relevant and necessary for the service purposes. However, data minimization is context-dependent, making violations difficult to define and quantify in privacy policies. Meanwhile, privacy policies are semantically complex and unstructured, hindering accurate extraction of fine-grained data practices and large-scale automated evaluation. To address these issues, we propose DataMini, a human--LLM collaborative evaluation framework for identifying violations of the data minimization principle in privacy policies. First, DataMini categorizes data minimization violations into two dimensions: inherent violations and contextual violations, establishing fine-grained evaluation criteria. Second, we construct a compliance baseline by mining high-frequency patterns from large-scale privacy policies and integrating expert knowledge to derive compliance mappings for human--LLM collaborative evaluation. Finally, the compliance baseline can automatically verify data practices that satisfy the data minimization principle, enabling the framework to focus exclusively on identifying suspected violations to improve efficiency and accuracy. Extensive evaluations demonstrate that DataMini exhibits superior data practice extraction accuracy of 83.46% and achieves an F1-score of 0.8180 for identifying data minimization violations in privacy policies, reducing manual evaluation effort by approximately 80%.

Video Moment Retrieval (VMR) aims to retrieve temporal moments in videos that align with natural language queries, a task requiring cross-modal reasoning between video and text. Among various supervision paradigms, point-supervised VMR has emerged as a practical solution, leveraging single-frame annotations to significantly reduce annotation costs while maintaining competitive retrieval performance. However, this sparse supervision approach induces cross-modal representation shift. This shift complicates the model's ability to accurately capture action sequences and associate text with visual content. To tackle this, we propose a novel framework called pseuDo fRame-based tempOral and semaNtic rEfinement (DRONE) with two key modules: (1) Pseudo-Frame Temporal Alignment (PTA), which embeds textual queries as pseudo-frames to enhance temporal coherence, and (2) Curriculum-Guided Semantic Refinement (CSR), which uses a progressive contrastive learning strategy to refine semantic representations from easy to hard cases. Extensive experiments show that DRONE achieves effective retrieval performance while keeping annotation costs low.

Video Moment Retrieval (VMR) aims to retrieve target video moments that correspond to natural language queries. Most existing methods rely on a positive-only assumption that the queried moment always exists within the video, which limits their reliability in practical scenarios. Departing from this restrictive setting, we study Video Moment Retrieval with Relevance Feedback (VMR-RF), which requires models to both retrieve relevant moments and reject irrelevant queries. This task remains challenging due to the following issues: 1) Intrinsic Semantic Interference caused by visually similar but irrelevant moments, and 2) Propagative Decision Irreversibility induced by unidirectional relevance prediction. In light of these, we introduce AdversaRial Reliability cOrrection netWork (ARROW) for VMR-RF. ARROW adopts an active discriminative strategy through two synergetic components: a Gradient-induced Semantic Adversary (GSA) that probes model vulnerabilities by actively amplifying semantic interference, and an Adversarial Reliability Predictor (ARP) that quantifies prediction stability under such interference to effectively suppress unreliable decisions. Extensive experiments validate the effectiveness of ARROW.

Video Corpus Moment Retrieval (VCMR) requires models to efficiently retrieve and precisely locate specific moments relevant to natural language queries within a massive, untrimmed video corpus. However, existing discriminative approaches typically rely on shallow visual-textual feature matching mechanisms, which often struggle to capture fine-grained semantic differences. To address this limitation, we propose Video-GAR, a novel framework that reframes the conventional retrieval task from superficial matching to generative understanding, positing that the capability for query reconstruction evidences deep semantic comprehension. Specifically, Video-GAR orchestrates three synergistic components: To overcome the computational efficiency bottleneck, we construct a Bi-Mamba backbone that leverages the linear complexity of state-space models for efficient global context modeling. Building on these representations, we introduce a generation-augmented fusion module, in which a training-only decoder acts as a semantic regularizer to implicitly calibrate cross-modal attention without increasing inference overhead. Finally, to ensure fine-grained precision, we propose a boundary-aware localization strategy that integrates boundary modeling with categorical supervision. Experiments on two benchmark datasets demonstrate that Video-GAR significantly improves retrieval and localization accuracy while maintaining outstanding inference speed.

Long video understanding (LVU) is challenging due to rich and complicated multimodal clues in long temporal range. Current methods adopt reasoning to improve the model's ability to analyze complex video clues in long videos via text-form reasoning. However, the existing literature suffers from the fact that the text-only reasoning under fixed video context may exacerbate hallucinations since detailed crucial clues are often ignored under limited video context length due to the temporal redundancy of long videos. To address this gap, we propose Video-TwG, a curriculum reinforced framework that employs a novel Think-with-Grounding paradigm, enabling video LLMs to actively decide when to perform on-demand grounding during interleaved text–video reasoning, selectively zooming into question-relevant clips only when necessary. Video-TwG can be trained end-to-end in a straightforward manner, without relying on complex auxiliary modules or heavily annotated reasoning traces. In detail, we design a Two-stage Reinforced Curriculum Strategy, where the model first learns think-with-grounding behavior on a small short-video GQA dataset with grounding labels, and then scales to diverse general QA data with videos of diverse domains to encourage generalization. Further, to handle complex think-with-grounding reasoning for various kinds of data, we propose the TwG-GRPO algorithm, which features the fine-grained grounding reward, self-confirmed pseudo reward, and accuracy-gated mechanism. Finally, we propose to construct a new TwG-51K dataset that facilitates training. Experiments on Video-MME, LongVideoBench, and MLVU show that Video-TwG consistently outperforms strong LVU baselines. Further ablation validates the necessity of our Two-stage Reinforced Curriculum Strategy and shows our TwG-GRPO better leverages diverse unlabeled data to improve grounding quality and reduce redundant groundings without sacrificing QA performance. https://github.com/hlchen23/Video-TwG

Spatio-Temporal Video Grounding (STVG) aims to retrieve the spatio-temporal tube of a target object or person in a video given a text query. Most existing approaches perform frame-wise spatial localization within a predicted temporal span, resulting in redundant computation, heavy supervision requirements, and limited generalization. Weakly-supervised variants mitigate annotation costs but remain constrained by the dataset-level train-and-fit paradigm with an inferior performance. To address these challenges, we propose the Agentic Spatio-Temporal Grounder (ASTG) framework for the task of STVG in an open-world and zero-shot setting. Specifically, two specialized agents SRA (Spatial Reasoning Agent) and TRA (Temporal Reasoning Agent) constructed leveraging modern Multi-modal Large Language Models (MLLMs) work collaboratively to retrieve the target tube in an autonomous and self-guided manner. Following a propose-and-evaluation paradigm, ASTG duly decouples spatio-temporal reasoning and automates the tube extraction, verification and temporal localization processes. With a dedicated visual memory and dialogue context, ASTG achieves architectural efficiency by minimizing the number of reasoning calls compared to exhaustive per-frame reasoning, eliminating the logical redundancy inherent in joint-reasoning systems. Experiments on popular benchmarks demonstrate the superiority of the proposed approach where it outperforms existing weakly-supervised and zero-shot approaches by a margin and is comparable to some of the fully-supervised methods.

Open-Vocabulary Temporal Action Detection (OV-TAD) aims to localize and classify action segments of unseen categories in untrimmed videos, where effective alignment between action semantics and video representations is critical for accurate detection. However, existing methods struggle to mitigate the semantic imbalance between concise, abstract action labels and rich, complex video contents, inevitably introducing semantic noise and misleading cross-modal alignment. To address this challenge, we propose DFAlign, the first framework that leverages diffusion-based denoising to generate foreground knowledge for the guidance of action–video alignment. Following the 'conditioning, denoising and aligning' manner, we first introduce the Semantic-Unify Conditioning (SUC) module, which unifies action-shared and action-specific semantics as conditions for diffusion denoising. Then, the Background-Suppress Denoising (BSD) module generates foreground knowledge by progressively removing background redundancy from videos through denoising process. This foreground knowledge serves as effective intermediate semantic anchor between video and text representations, mitigating the semantic gap and enhancing the discriminability of action-relevant segments. Furthermore, we introduce the Foreground-Prompt Alignment (FPA) module to inject extracted foreground knowledge as prompt tokens into text representations, guiding model's attention towards action-relevant segments and enabling precise cross-modal alignment. Extensive experiments demonstrate that our method achieves state-of-the-art performance on two OV-TAD benchmarks. The code repository is provided as follows: https://github.com/Sasa77777779/DFAlign_SIGIR26.git.

Web search engines have long been the primary gateway to online information, but generative AI chatbots are emerging as alternative tools. Yet we lack empirical understanding of how users choose between these tools and how their information-seeking behaviors differ across these contexts. Understanding how users select, use, and assess information retrieval (IR) systems is essential for evaluating them in ways that reflect user expectations across contexts. To shed light on user perceptions and usage of IR systems, we surveyed 84 participants about their recent search engine and GenAI chatbot use, analyzing task types, contextual triggers, outcomes, and tool preferences. Our findings reveal that users approach search engines and GenAI chatbots with distinct philosophies. With search engines, they follow a 'Browse and Verify' approach, prioritizing reliability and source verifiability. With GenAI chatbots, they adopt a 'Delegate and Consume' approach, valuing quick summarization and personalized content. Despite these stated preferences, actual usage shows that users frequently turn to GenAI chatbots for actionable guidance in high-stakes domains such as health and relationship matters, hinting at a divergence between stated philosophies and real-world behavior. These findings underscore the need for IR systems that support reliable information assessment, transparent sources, and user-guided evaluation.

We investigate between-user drivers of query variability through a controlled between-subject, full-factorial user study that manipulates age, gender, and language proficiency across six backstory-driven search tasks. From initial queries, session logs, and post-task interviews, we quantify how demographic and task factors shape query-, task-, and session-level behaviors. We further derive a small set of interpretable latent search dimensions from user evidence to analyze and simulate heterogeneous query behavior. Our results show that age is the most consistent predictor of query formulation and search interaction patterns. Gender and language differences are more selective, and task context is further associated with these patterns. The latent dimensions help explain the differences as variation in search strategy rather than uniform differences in engagement or ability. The paper provides a trait-informed view of heterogeneous search behavior that supports more user-aware analysis and robustness-oriented evaluation in IR.

We conduct a lab eye-tracking study to examine how users interact with search engines that place Generative Artificial Intelligence (GenAI) results above traditionally ranked search results, also known as "ten blue links", and we use an engagement scale to evaluate their experience. Our aim is to study how users interact with search engine interfaces that incorporate GenAI content, assess users' willingness to scroll past GenAI content to view the traditional search results, and explore how these interactions differ from existing literature on scanning search engine result pages. % We show that GenAI content is changing how people search, but the "golden triangle" remains valid, where the top-left section of the search page attracts the most attention. Searchers are still engaging with the blue links in patterns consistent with the literature; however, they engage significantly more with GenAI content. Finally, we outline future directions to deepen our understanding of search behavior in the era of GenAI.

Cognitive style, a user's habitual approach to information processing, offers a promising approach to personalising information searching (IS) systems, yet the underlying style-related search behaviours remain poorly understood. This study investigates how the Wholist–Analytic and Verbal–Imagery dimensions of cognitive style influence search behaviour and prefrontal cortex (PFC) activation during the Post-focus stage of exploratory search. Forty participants completed comparison search tasks while we recorded behavioural metrics, subjective workload ratings, and functional near-infrared spectroscopy (fNIRS) data. Our results demonstrate that cognitive style significantly predicted search engine results page (SERP) interaction patterns: analytics and imagers adopted structured navigation with detailed reading, whereas wholists and verbalisers preferred sporadic navigation with rapid scanning. Critically, fNIRS revealed distinct PFC activation patterns, specifically in the Ventrolateral PFC (VLPFC) and Dorsolateral PFC (DLPFC), corresponding to these behavioural differences. By mapping neuro-cognitive profiles to IR behaviours, this work provides empirical grounding for designing ''style-aware'' adaptive IR interfaces that tailor information density and navigational support to individual cognitive profiles.

Carousel interfaces have been the de-facto standard for streaming media services for over a decade. Yet, there has been very little research into user behavior with such interfaces, which thus remains poorly understood. Due to this lack of empirical research, previous work has assumed that behaviors established in single-list web-search interfaces, such as the F-pattern and the examination hypothesis, also apply to carousel interfaces, for instance when designing click models or evaluation metrics. We analyze a recently-released interaction and examination dataset resulting from an eye-tracking study performed on carousel interfaces to verify whether these assumptions actually hold. We find that (i)~the F-pattern holds only for vertical examination and not for horizontal swiping; additionally, we discover that, when conditioned on a click, user examination follows an L-pattern unique to carousel interfaces; (ii)~click-through-rates conditioned on examination indicate that the well-known examination hypothesis does not hold in carousel interfaces; and (iii)~contrary to the assumptions of previous work, users generally ignore carousel headings and focus directly on the content items. Our findings show that many user behavior assumptions, especially concerning examination patterns, do not transfer from web search interfaces to carousel recommendation settings. Our work shows that the field lacks a reliable foundation on which to build models of user behavior with these interfaces. Consequently, a re-evaluation of existing metrics and click models for carousel interfaces may be warranted.

LLM-powered search agents are increasingly being used for multi-step information seeking tasks, yet the IR community lacks empirical understanding of how agentic search sessions unfold and how retrieved evidence is reflected in later queries. This paper presents a large-scale log analysis of agentic search based on 14.44M search requests (3.97M sessions) collected from DeepResearchGym, i.e., an open-source search API accessed by external agentic clients. We sessionize the logs, assign session-level intents and step-wise query-reformulation labels using LLM-based annotation, and propose Context-driven Term Adoption Rate (CTAR) to quantify whether newly introduced query terms are lexically traceable to previously retrieved evidence. Our analyses reveal distinctive behavioral patterns. First, over 90% of multi-turn sessions contain at most ten steps, and 89% of inter-step intervals fall under one minute. Second, behavior varies by intent. Fact-seeking sessions exhibit high repetition that increases over time, while sessions requiring reasoning sustain broader exploration. Third, query reformulations are often traceable to retrieved evidence across steps. On average, 54% of newly introduced query terms appear in the accumulated evidence context, with additional traceability to earlier steps beyond the most recent retrieval. These findings provide candidate signals for repetition-aware stopping, intent-adaptive retrieval budgeting, and explicit cross-step context tracking. We released the anonymized logs, making them available at a public HuggingFace repository.

We study efficient multi-vector retrieval for late interaction in any modality. Late interaction has emerged as a dominant paradigm for information retrieval across modalities, but its computation and storage costs grow linearly with document length, making it costly for multimodal corpora. To address this limitation, we explore methods for compressing multi-vector document representations under a constant vector budget. We introduce four approaches for index compression, including a novel attention-guided clustering (øurs). øurs uses an attention-guided mechanism to identify the most semantically salient parts of a document as cluster centroids and to weight token aggregation. Evaluating these methods on retrieval tasks spanning 4 modality settings (\beir, \vidore, \msrvtt, μltivent), we show that øurs consistently outperforms other parameterized compression methods (sequence resizing and memory tokens), provides greater flexibility in index size than non-parametric hierarchical clustering, and achieves competitive or improved performance compared to a full, uncompressed index.

Reproducibility must validate architectural robustness, not just numerical accuracy. We evaluate ColBERT-v2 and ConstBERT across five dimensions, finding that while ConstBERT reproduces within 0.05% MRR@10 on MS-MARCO, both models show a drop of 86–97% on long, narrative queries (TREC ToT 2025). Ablations prove this failure is architectural: performance plateaus at 20 words because the MaxSim operator's uniform token weighting cannot distinguish signal from filler noise. Furthermore, undocumented backend parameters create an 8-point gap due to ConstBERT's sparse centroid coverage, and fine-tuning with 3× more data actually degrades performance by up to 29%. We conclude that architectural constraints in multi-vector retrieval cannot be overcome by adaptation alone. Code: https://github.com/utshabkg/multi-vector-reproducibility.

The computational costs of the transformer-based multi-vector retrieval model ColBERT depend on the number of vectors used to represent queries and documents. Common strategies to lower the costs thus prune the vectors to a fixed number or relative to the sequence length. We compare standard pruning approaches like weighted token pruning or IDF-based pruning and analyze the impact on the downstream effectiveness of respective ColBERT models. Our experiments indicate that weighted pruning can yield a better effectiveness--efficiency trade-off than other pruning techniques, but we also find that very simplistic pruning techniques can yield very effective ColBERT models when trained properly.

Late-interaction models such as ColBERT offer competitive performance across various retrieval tasks but require storing a dense embedding for each document token, leading to a substantial index storage overhead. Past works address this by attempting to prune low-importance token embeddings based on statistical and empirical measures, but they often either lack formal grounding or are ineffective. To address these shortcomings, we introduce a framework grounded in hyperspace geometry and cast token pruning as a Voronoi cell estimation problem in the embedding space. By interpreting each token's influence as a measure of its Voronoi region, our approach enables principled pruning that retains retrieval quality while reducing index size. Through our experiments, we demonstrate that this approach serves not only as a competitive pruning strategy but also as a valuable tool for improving and interpreting token-level behavior within dense retrieval systems.

This study addresses the challenge of improving dense retrieval performance for queries containing numerical conditions, such as ''companies with more than one billion dollars in R&D expenditure.'' Although recent research has underscored the limitations of standard models in handling numeric information across domains such as finance, e-commerce, and medicine, existing solutions typically decompose queries into textual and numerical components and score them separately using dedicated methods. These approaches intrude upon late-interaction retrieval models such as ColBERT and incur considerable challenges in deployment, latency, and maintainability. To overcome these limitations, we propose NumColBERT, an inference-time non-intrusive method that enhances numerically conditioned retrieval while preserving the original late-interaction mechanism and providing unified scoring across textual and numerical content. Because NumColBERT retains the standard ColBERT indexing and MaxSim scoring pipeline, existing optimizations and ecosystem components developed for ColBERT can be directly reused, facilitating practical deployment. NumColBERT introduces a Numerical Gating Mechanism and a Numerical Contrastive Learning objective to enable numerical conditions to contribute more effectively to retrieval within the standard ColBERT scoring mechanism. The gating mechanism dynamically amplifies the influence of tokens carrying critical numerical constraints while suppressing context-neutral mentions such as model numbers or dates. The contrastive objective explicitly shapes the embedding space to reflect numerical magnitudes and conditions, enabling numerical values to be distinguished within the shared representation space. Experimental results show that NumColBERT substantially outperforms standard fine-tuning baselines and achieves accuracy that matches or exceeds that of prior approaches that rely on separate textual and numerical scoring. These findings demonstrate the feasibility of numerically conditioned retrieval with a non-intrusive inference pipeline and present a maintainable solution for real-world deployment.

Multi-vector dense retrieval models, such as ColBERT, achieve strong retrieval effectiveness by modelling fine-grained token-level interactions between queries and documents. Methods such as PLAID use centroid-based quantisation of each token's vector to reduce the index size and speed up retrieval while maintaining strong effectiveness. In this work, we introduce PLAID-PRF, a method that performs Pseudo-Relevance Feedback (PRF) over PLAID to reformulate ColBERT's query vectors based on the top-retrieved results. In contrast with prior methods that perform PRF on multi-vector retrieval models, PLAID-PRF keeps computational costs low by leveraging the internal PLAID centroid vectors, treating them similarly to tokens in traditional PRF methods. The method selects a small and diverse set of high-utility expansion vectors and appends them to the original query, rerunning PLAID to refine both candidate generation and final scoring. Extensive experiments on the standard in-domain MSMARCO and four out-of-domain BEIR benchmarks show that PLAID-PRF consistently improves retrieval effectiveness over various baselines. In particular, PLAID-PRF improves over PLAID by up to 4.3% nDCG@10 and 7.3% MRR@10, while introducing substantially less computation overhead than prior PRF methods. The results demonstrate that our proposed centroid-aware PRF method offers an effective and lightweight mechanism to improve the quality of top-ranked retrieved results. Overall, this work enables effective and efficient feedback-aware late-interaction retrieval without expensive query-time document-token clustering.

Disentangled recommendation within the Variational Autoencoder (VAE) framework aims to capture multiple user interests. While effective, these VAEs are fundamentally constrained by their reliance on interaction data alone, lacking the rich external semantic knowledge needed to properly structure and separate latent interests. Meanwhile, Large Language Models (LLMs) excel at deriving profound user preference signals from textual data. Prevailing methods for integrating LLMs into recommendation, however, either focus on single-interest modeling or perform a shallow fusion by aligning LLM and VAE representation spaces. Thus, they fail to fundamentally shape the VAE's latent space for multi-interest learning, hindering recommendation performance. To bridge this gap, we propose to fundamentally shift the integration point: instead of aligning representation spaces, we bridge the LLM-generated embedding space directly to the VAE's parameter space. Our framework designs hypernetworks to generate parameters governing interest discovery of a disentangled VAE conditioned on LLM-derived user embeddings. This direct bridge injects rich semantic knowledge into the model's learning foundation, preserving the VAE's power for disentangled representation while dramatically enhancing its modeling capability with informative, LLM-structured priors. Extensive experiments on benchmark datasets show that our method yields significant performance gains, unveiling the untapped potential of using LLMs for disentangled recommendation.

Recently, large language models (LLMs) have advanced recommendation systems (RSs), and recent works have begun to explore how to integrate LLMs into industrial RSs. While most approaches deploy LLMs offline to generate and pre-cache augmented representations for RSs, high-dimensional representations from LLMs introduce substantial storage and computational costs. Thus, it is crucial to compress LLM representations effectively. However, we identify a counterintuitive phenomenon during representation compression: Mid-layer Representation Advantage (MRA), where representations from middle layers of LLMs outperform those from final layers in recommendation tasks. This degraded final layer renders existing compression methods, which typically compress on the final layer, suboptimal. We interpret this based on modularity theory that LLMs develop spontaneous internal functional modularity and force the final layer to specialize in the proxy training task. Thus, we propose Modular Representation Compression (MARC) to explicitly control the modularity of LLMs. First, Modular Adjustment explicitly introduces compression and task adaptation modules, enabling the LLM to operate strictly as a representation-learning module. Next, to ground each module to its specific task, Modular Task Decoupling uses information constraints and different network structures to decouple tasks. Extensive experiments validate that MARC addresses MRA and produces efficient representations. Notably, MARC achieved a 2.82% eCPM lift in an online A/B test within a large-scale commercial search advertising scenario.

Large language models (LLMs) have become an important semantic infrastructure for modern recommender systems. A prevailing paradigm integrates LLM-derived semantic embeddings with collaborative representations via representation alignment, implicitly assuming that the two views encode a shared latent entity and that stronger alignment yields better results. We formalize this assumption as the global low-complexity alignment hypothesis and argue that it is stronger than necessary and often structurally mismatched with real-world recommendation settings. We propose a complementary perspective in which semantic and collaborative representations are treated as partially shared yet fundamentally heterogeneous views, each containing both shared and view-specific factors. Under this shared-plus-private latent structure, enforcing global geometric alignment may distort local structure, suppress view-specific signals, and reduce informational diversity. To support this perspective, we develop complementarity-aware diagnostics that quantify overlap, unique-hit contribution, and theoretical fusion upper bounds. Empirical analyses on sparse recommendation benchmarks reveal low item-level agreement between semantic and collaborative views and substantial oracle fusion gains, indicating strong complementarity. Furthermore, controlled alignment probes show that low-capacity mappings capture only shared components and fail to recover full collaborative geometry, especially under distribution shift. These findings suggest that alignment should not be treated as the default integration principle. We advocate a shift from alignment-centric modeling to complementarity fusion-centric, complementarity-aware design, where shared factors are selectively integrated while private signals are preserved. This reframing provides a principled foundation for the next generation of LLM-enhanced recommender systems.

Large language models (LLMs) have recently emerged as powerful training-free recommenders. However, their knowledge of individual items is inevitably uneven due to imbalanced information exposure during pretraining, a phenomenon we refer to as knowledge gap problem. To address this, most prior methods have employed a naive uniform augmentation that appends external information for every item in the input prompt. However, this approach not only wastes limited context budget on redundant augmentation for well-known items but can also hinder the model's effective reasoning. To this end, we propose KnowSACKP(Knowledge-aware Selective Augmentation with Comparative Knowledge Probing) to mitigate the knowledge gap problem. KnowSACKP estimates the LLM's internal knowledge by evaluating its capability to capture collaborative relationships and selectively injects additional information only where it is most needed. By avoiding unnecessary augmentation for well-known items, KnowSACKP focuses on items that benefit most from knowledge supplementation, thereby making more effective use of the context budget. KnowSACKP requires no fine-tuning step, and consistently improves both recommendation accuracy and context efficiency across four real-world datasets. Our code is available at https://github.com/nowhyun/KnowSA_CKP.

We introduce RecPFN, a prior-fitted network that brings in-context learning to sequential recommendation. RecPFN is pretrained entirely on synthetic clickstream environments sampled from a broad structural causal prior, enabling it to amortize Bayesian-style inference from a small support set. At inference, a lightweight decoder-only transformer conditions on a handful of domain sequences and produces next-item predictions for queries in a single forward pass, without any weight updates. Across eight public benchmarks, RecPFN achives state-of-the-art zero-shot performance while remaining strongly competitive with supervised methods in low-compute and low-data regimes. It is deployment-efficient and robust to domain shift, outperforming strong zero-shot baselines that rely on large real-interaction corpora. RecPFN provides a practical path toward generalizable, data-efficient recommenders and opens avenues for richer priors, longer-context ICL, and multimodal extensions. Code for training and evaluation will be made publicly available by the conference date.

Graphs are prevalent in numerous real-world applications. Previous methods directly model graph structures and achieve significant success. However, these methods encounter bottlenecks due to the inherent irregularity of graphs. An innovative solution is converting graphs into textual representations, thereby harnessing the powerful capabilities of Large Language Models (LLMs) to process and comprehend graphs. In this paper, we present a comprehensive review of methodologies for applying LLMs to graphs, termed LLM4graph. The core of LLM4graph lies in transforming graphs into texts for LLMs to understand and analyze. Thus, we propose a novel taxonomy of LLM4graph methods from the view of the transformation. Specifically, existing methods can be divided into two paradigms: Graph2text and Graph2token, which transform graphs into texts or tokens as the input of LLMs, respectively. We point out four challenges during the transformation to systematically present existing methods from a problem-oriented perspective. For practical concerns, we provide a guideline for researchers on selecting appropriate models and LLMs for different graphs and hardware constraints. To empirically evaluate our taxonomy and different technical choices, we conduct experiments with representative methods in Graph2text and Graph2token. We also identify five future research directions for LLM4graph.

Pre-training graph foundation models (GFMs) on text-attributed graphs (TAGs) is important for web-scale retrieval and recommendation, where graph entities are matched with textual descriptions. Existing CLIP-style graph-text aligners typically assume one-to-one correspondence: each node is pulled close only to its paired text, and all other pairs are treated as negatives. This overlooks the many-to-many relations common in real TAGs, where a node and its local neighborhood can be semantically related to multiple texts, and vice versa. Meanwhile, TAG supervision is often imperfect: noisy or weak node-text links introduce false-positive pairs, causing contrastive learning to align mismatched semantics. These limitations reveal a fundamental trade-off: leveraging expressive many-to-many signals increases semantic coverage but may propagate errors under noise, whereas strict one-to-one training is more conservative yet still suffers when mismatched pairs remain in the training set. Therefore, we propose ADAligner, a quality-aware graph–text alignment framework that adapts between expressive many-to-many and conservative one-to-one objectives based on estimated alignment reliability. ADAligner tracks batch-level reliability online and adjusts optimization accordingly—promoting soft, subgraph-level alignment when supervision is clean while emphasizing reliable one-to-one alignment by filtering low-confidence pairs under noise. We provide theoretical analysis showing that this closed-loop adaptation is stable and convergent. Experiments on nine TAG benchmarks show that, under 30% mismatched node-text supervision, ADAligner consistently improves cross-modal retrieval by 144.70% on average, zero-/few-shot node classification by 26.13%, and link prediction by 4.70% over the strongest multimodal baseline, demonstrating strong robustness to alignment noise across both unsupervised and transfer settings. Our code is available at https://github.com/karmaisacat-13/ADAligner.

Incomplete node features are ubiquitous in real-world scenarios such as user profiling and cold-start recommendation, which severely hinders the practical deployment of graph learning systems (e.g., GNNs). Existing solutions typically rely on diffusion-based structural smoothing (e.g., feature propagation) to impute missing values. However, we find that these approaches suffer from structural overfitting, leading to three progressive challenges: 1) performance degradation on disjoint graphs, 2) loss of semantic diversity due to over-smoothing, and 3) feature distribution shift when generalizing to unseen graph structures (inductive tasks). To address these challenges, we introduce the DART framework. It begins by employing Global Structural Augmentation (GSA), which establishes global correlations to bridge disjoint components and extend diffusion coverage. Building upon this, we design a semantic rectifier based on masked autoencoding. This module learns the latent feature manifold to recover natural semantic details. Crucially, we introduce a test-time distribution rectification mechanism that projects structurally biased features back onto the learned manifold during inference, effectively bridging the inductive distribution gap. Furthermore, considering that synthetic masking fails to reflect realworld sparsity, we present a new dataset Sailing collected from voyage records with naturally missing attributes. Extensive experiments on six public datasets and Sailing demonstrate that DART significantly outperforms state-of-the-art methods in both transductive and inductive settings. Our code and dataset are available at https://github.com/yfsong00/DART.

Existing works in graph active learning (GAL) mostly assume a closed-world setting where all classes are known in advance, while in practice, we need to deal with the open-world setting where novel classes are encountered during learning. Apart from selecting informative nodes for refining the current classifier, the open-world GAL algorithms are also expected to discover novel classes. Motivated by the observation that identifying an informative region can be easier than finding the most informative example, we propose a novel hierarchical cluster-based algorithm for open-world GAL. Our algorithm performs clustering in the feature space to identify an informative region which either contains informative examples for known classes or novel classes, then performs a novel semi-supervised sub-clustering on the selected cluster to identify the most informative example. To facilitate the discovery of informative regions, we introduce a cluster-based self-distillation loss between ego and final embeddings to learn well-clustered node embeddings that are more likely to align with the classes. We also introduce a novel cluster informativeness score for the clusters, which measures not only how uncertain the cluster is (similar to standard active learning algorithms), but whether the cluster is likely to contain novel classes. Our semi-supervised sub-clustering algorithm partitions the selected cluster into subregions for known classes and an additional uncertain subregion, then we query the label for a representative node for the uncertain subregion. Extensive experiments on five benchmark datasets demonstrate the effectiveness of the proposed method with a more balanced improvement over novel classes. Ablation study shows that our cluster-based self-distillation loss, informativeness score, and the semi-supervised sub-clustering strategy are all beneficial. Empirical analysis also reveals how our informativeness score is effective for novel class discovery and identifying informative examples for refining the decision boundary.

Heterophily is a prevalent property of real-world graphs and is well known to impair the performance of homophilic Graph Neural Networks (GNNs). Prior work has attempted to adapt GNNs to heterophilic graphs through non-local neighbor extension or architecture refinement. However, the fundamental reasons behind misclassifications remain poorly understood. In this work, we take a novel perspective by examining recurring inductive subgraphs, empirically and theoretically showing that they act as spurious shortcuts that mislead GNNs and reinforce non-causal correlations in heterophilic graphs. To address this, we adopt a causal inference perspective to analyze and correct the biased learning behavior induced by shortcut inductive subgraphs. We propose a debiased causal graph that explicitly blocks confounding and spillover paths responsible for these shortcuts. Guided by this causal graph, we introduce Causal Disentangled GNN (CD-GNN), a principled framework that disentangles spurious inductive subgraphs from true causal subgraphs by explicitly blocking non-causal paths. By focusing on genuine causal signals, CD-GNN substantially improves the robustness and accuracy of node classification in heterophilic graphs. Extensive experiments on real-world datasets not only validate our theoretical findings but also demonstrate that our proposed CD-GNN outperforms state-of-the-art heterophily-aware baselines.

Large Language Models (LLMs) have fundamentally transformed dense retrieval, upgrading backbones from discriminative encoders to generative architectures. However, a critical disconnect remains: while LLMs possess strong reasoning capabilities, current retrievers predominantly utilize them as static encoders, leaving their potential for complex reasoning unexplored. To address this, existing approaches typically adopt ''rewrite-then-retrieve'' pipelines to generate explicit Chain-of-Thought (CoT) rationales before retrieval. However, this incurs prohibitive latency. Conversely, implicit reasoning methods utilizing latent tokens offer efficiency but often suffer from semantic degeneration due to the lack of explicit supervision. In this paper, we propose LaSER, a novel self-distillation framework that internalizes explicit reasoning into the latent space of dense retrievers. Operating on a shared LLM backbone, LaSER introduces a dual-view training mechanism: an Explicit view that explicitly encodes ground-truth reasoning paths, and a Latent view that performs implicit latent thinking. To bridge the gap between these views, we design a multi-grained alignment strategy. Beyond standard output alignment, we introduce a trajectory alignment mechanism that synchronizes the intermediate latent states of the latent path with the semantic progression of the explicit reasoning segments. This allows the retriever to ''think'' silently and effectively without autoregressive text generation. Extensive experiments on both in-domain and out-of-domain reasoning-intensive benchmarks demonstrate that LaSER significantly outperforms state-of-the-art baselines. Furthermore, analyses across diverse backbones and model scales validate the robustness of our approach, confirming that our unified learning framework is essential for eliciting effective latent thinking. Our method successfully combines the reasoning depth of explicit CoT pipelines with the inference efficiency of standard dense retrievers. The code, model, and training data are available at https://github.com/RUC-NLPIR/LaSER.

Query suggestion (QS) in web search must provide efficient and up-to-date suggestions that are relevant to the session context. Recent studies treat QS as generation, which captures session context well but suffers from high latency and costly retraining to maintain information freshness. In this study, we recast QS as dense retrieval: given a search session, the next query is retrieved from a large index of historical queries using efficient approximate nearest neighbor search. We propose Context-aware Asymmetric Dual Encoder for QS (CADE-QS), which uses an asymmetric dual encoder to model query -to- suggestion dependency and incorporates session history into the query encoder via context-aware contrastive learning. We evaluate our method on two real-world search-log datasets: a recent Japanese web search log and the public AOL log. CADE-QS rivals strong generative baselines in quality while reducing end-to-end latency to around 30 ms on CPU, representing an order-of-magnitude improvement. Detailed analyses confirm CADE-QS's robust context awareness, unidirectional modeling, and practicality for reflecting evolving information via index refreshes, as well as the effectiveness of an adaptive hybrid strategy for low-coverage scenarios. Our code is publicly available at https://github.com/lycorp-jp/cadeqs.

Dense retrieval, which encodes queries and documents into a single dense vector, has become the dominant neural retrieval approach due to its simplicity and compatibility with fast approximate nearest neighbor algorithms. As the tasks dense retrieval performs grow in complexity, the fundamental limitations of the underlying data structure and similarity metric---namely vectors and inner-products---become more apparent. Recent work has shown theoretical limitations inherent to single vectors and inner-products, which are generally tied to the embedding dimension. Given the importance of embedding dimension for retrieval capacity, understanding how dense retrieval performance changes as embedding dimension is scaled is fundamental to building next-generation retrieval models that balance effectiveness and efficiency. In this work, we conduct a comprehensive analysis of the relationship between embedding dimension and retrieval performance. Our experiments include a range of model sizes from two model families to construct a detailed picture of embedding scaling behavior. We find that the scaling behavior fits a power law, allowing us to derive scaling laws for performance given only the embedding dimension, as well as a joint law accounting for embedding dimension and model size. Our analysis shows that for evaluation tasks aligned with the training task, performance continues to improve as embedding size increases, though with diminishing returns. For evaluation data that are less aligned with the training task, we find that performance is less predictable, with performance degrading with larger embedding dimensions for certain tasks. We hope our work provides additional insight into the limitations of embeddings and their behavior as well as offers a practical guide for selecting model and embedding dimension to achieve optimal performance with reduced storage and compute costs.

Dense retrieval uses embedded dense representations for documents to be obtained from language models ingested in vector stores. To reduce large space requirements and/or efficiency concerns, approximate nearest neighbours algorithms can be deployed, typically at the cost of effectiveness. Among existing approaches, Product Quantisation (PQ) is one well-known algorithm, where embeddings are split along dimensions, and then assigned to similar cluster centroids on a per-split basis. Joint Product Quantisation (JPQ) improves over PQ by further training the obtained centroid embeddings, as well as the query encoder, for enhanced effectiveness. In this replicability study, we reimplement JPQ, comparing its performance on the STAR biencoder model with that reported in the original paper, as well as considering its generalisability in several dimensions: to other biencoders such as TCT and TAS-B; to larger models such as RepLLama; and beyond document retrieval, to a retrieval augmented generation task. Among many interesting results, we find that JPQ does work on more modern biencoder models such as TCT but not on larger models such as RepLLama; however, it does scale to large corpora, allowing dense retrieval for RAG that is as effective from an index that is 32x larger.

Retrieval benchmarks for large language models (LLMs) should reflect the long, reasoning-intensive queries typical of retrieval-augmented generation (RAG). We present a systematic study of BRIGHT, a reasoning-focused retrieval benchmark, along with strong, reproducible reference methods integrated into Anserini, Pyserini, and RankLLM. We evaluate lexical, sparse, dense, and fusion-based retrievers, as well as LLM rerankers, under long-query settings. In reproducing BRIGHT's lexical baseline, we identify a key under-documented detail: query-side BM25 (BM25Q), which applies BM25 weighting to the query itself. On long, multi-sentence queries, BM25Q consistently outperforms standard BM25, making it the strongest lexical baseline for reasoning-oriented retrieval. We further audit the BRIGHT corpus, uncovering data quality issues that impact evaluation, and offer mitigation. Finally, we study the generalizability of BM25Q across five additional benchmarks, finding its gains largely specific to BRIGHT, while fusion with standard BM25 provides the most consistent improvements across datasets.

The proliferation of social bots poses a persistent threat to online social platforms, making accurate and efficient detection increasingly critical. Although recent graph-based methods have achieved notable progress by modeling user interactions, many of them implicitly assume neighborhood consistency and degrade in sparse, directed, and heterophilic social graphs, where a user's neighborhood may poorly reflect its own attributes. In this work, we propose SNBot, a novel social bot detection framework that explicitly models the discrepancy between node self-representations and their neighborhood embeddings. SNBot first learns unified self-representations from multi-modal user attributes through a lightweight type-aware encoding scheme. It then performs message passing on an augmented directed graph using a bidirectional aggregation mechanism, which separately captures incoming and outgoing interactions to better characterize asymmetric behaviors. By preserving and exploiting self–neighborhood representation discrepancies rather than over-smoothing them, SNBot produces more discriminative node representations. Extensive experiments on multiple real-world benchmark datasets demonstrate that SNBot consistently outperforms state-of-the-art methods.

Depression has become a critical global public health challenge, creating an urgent need for automated and scalable screening solutions. Social media platforms, which capture rich and spontaneous multi-modal behavioral data, offer a promising avenue for detecting early signs of mental distress. However, existing depression detection methods predominantly rely on static multi-modal fusion strategies and frequently fail to effectively tackle cross-modal semantic gaps. To address these limitations, we propose a Mutual Guidance Multi-modal Network with Complementary Graph Learning (MGMN) for depression detection by observing individuals' behavioral performance on social media. Specifically, a cross-modal mutual guidance mechanism is designed to dynamically construct a complementary graph by using mutual similarities within and across visual and acoustic modalities common in social media. More specifically, based on this complementary graph, a modality-specific adaptive residual learning module is applied to each modality to stabilize deep feature learning and preserve modality-specific and complementary information via graph-conditioned adaptive residual fusion. Furthermore, the refined uni-modal features are subsequently fed into a joint-modal fusion and prediction module to output the final disease prediction probability. Extensive experiments on the MUD3, LMVD, and D-vlog datasets demonstrate our proposed method's superiority over state-of-the-art methods, confirming that the proposed framework provides a robust and effective solution for mental health monitoring. Codes are available at https://github.com/Petofi-romance/MGMN

The rise of live streaming has transformed online interaction, enabling massive real-time engagement but also exposing platforms to complex risks such as scams and coordinated malicious behaviors. Detecting these risks is challenging because harmful actions often accumulate gradually and recur across seemingly unrelated streams. To address this, we propose CS-VAR (Cross-Session Evidence-Aware Retrieval-Augmented Detector) for live streaming risk assessment. In CS-VAR, a lightweight, domain-specific model performs fast session-level risk inference, guided during training by a Large Language Model (LLM) that reasons over retrieved cross-session behavioral evidence and transfers its local-to-global insights to the small model. This design enables the small model to recognize recurring patterns across streams, perform structured risk assessment, and maintain efficiency for real-time deployment. Extensive offline experiments on large-scale industrial datasets, combined with online validation, demonstrate the state-of-the-art performance of CS-VAR. Furthermore, CS-VAR provides interpretable, localized signals that effectively empower real-world moderation for live streaming.

Emotion-Cause Pair Extraction in Conversations (ECPEC) aims to identify speakers' emotions and the corresponding causes within conversation contexts. This task has gained increasing attention in recent years. Existing methods typically operate from the perspective of text object, inputting the entire conversation text and relying on neural networks to highlight utterance features as a means of emotion perception, while employing pair concatenation as a single path cause reasoning strategy. However, in conversation scenarios, humans are the true origin of emotions, while language text merely serves as the medium of expression. Consequently, these methods lack the modeling of the cognition and reasoning processes behind the human's behavior from the perspective of the speaker subject. To address this issue, we propose a novel Hierarchical Cognition–Perception–Reasoning Agent (HCPRA) framework. The framework is oriented around the speaker, simulating their personality cognition and conversation scenario cognition. Moreover, it utilizes hybrid memory and implicit emotion enhancement to simulate the human emotion perception. Additionally, it employs a hierarchical emotion cause reasoning mechanism to extract interpretable relationships between the speaker's emotions and their causes through emotion awareness and multi-path cause reasoning. Experimental results demonstrate that our approach achieves state-of-the-art performance on three datasets, with F1 improvement up to 15.61%.

Hateful videos have become prevalent on online platforms, highlighting an urgent need for effective detection. However, existing studies primarily focus on binary classification and fail to provide contextual rationales that reveal the implicit meanings behind these judgments, significantly undermining model explainability. To fill this gap, we aim to achieve explainable hateful video detection, enabling models to provide contextual rationales that integrate relevant evidence and logical reasoning alongside decisions. This approach can comprehensively enhance the understanding of video content and the explainability of the decision-making process. We first introduce two datasets, Ex-HateMM and Ex-ImpliHateVid, for explainable hateful video detection. Each dataset provides fine-grained annotations of multimodal harmful elements, along with contextual rationales. We then propose an Information Augmentation and Reasoning Enhancement (IARE) framework designed for explainable detection. The framework employs an information augmentation phase that leverages the multimodal chain-of-thought to integrate harmful elements, thereby enriching rationale evidence. Additionally, IARE incorporates a reasoning enhancement phase, in which Direct Preference Optimization guides the model toward correct reasoning paths and away from incorrect ones, thereby improving the logical coherence of its justifications. We conduct extensive experiments on the two datasets, comparing multiple baselines with our proposed IARE framework. The results demonstrate that IARE achieves state-of-the-art performance while also generating accurate rationales.

Provider exposure fairness is crucial for sustaining a healthy content ecosystem and preventing monopolization in recommender systems. Yet, most existing methods either incorporate fairness constraints during model training, requiring expensive retraining when fairness objectives change, or rely on post-hoc reranking with fixed criteria, which lacks adaptability to diverse fairness requirements. To overcome these limitations, we propose Post-hoc Fairness Adaptation (PFA), a lightweight framework that equips a frozen recommender with a fairness adapter, enabling flexible fairness control without retraining the backbone model. Specifically, the fairness adapter learns personalized additive score adjustments from user–item embeddings, which are injected into the original ranking scores to steer provider exposure toward fairness. To train the adapter, we minimize the Kullback-Leibler (KL) divergence between the actual and the target fair exposure distributions. However, this global objective implicitly treats all providers equally, ignoring structural disparities such as imbalanced provider group sizes and heterogeneous exposure within groups. Consequently, fairness may appear satisfied at an aggregate level while severe inter-group and intra-group exposure imbalances persist, undermining practical fairness. To address this, we design Hierarchical Exposure Fairness Alignment (HEFA), which explicitly balances inter- and intra-group provider exposure disparities, enabling flexible adaptation to diverse fairness requirements. To mitigate potential accuracy degradation, PFA jointly optimizes HEFA with a differentiable NDCG loss, enabling end-to-end fairness optimization while preserving ranking quality. Extensive experiments on three public datasets demonstrate that PFA achieves substantial fairness gains with negligible accuracy loss, consistently outperforming strong baselines. Code is available at https://github.com/Tam-JQK/Post-train.

Ranking plays a central role in connecting users and providers in Information Retrieval (IR) systems, making provider-side fairness an important challenge. While recent research has begun to address fairness in ranking, most existing approaches adopt an equality-based perspective, aiming to ensure that providers with similar content receive similar exposure. However, it overlooks the diverse needs of real-world providers, whose utility from ranking may depend not only on exposure but also on outcomes like sales or engagement. Consequently, exposure-based fairness may not accurately capture the true utility perceived by different providers with varying priorities. To this end, we introduce an equity-oriented fairness framework that explicitly models each provider's preferences over key outcomes such as exposure and sales, thus evaluating whether a ranking algorithm can fulfill these individualized goals while maintaining overall fairness across providers. Based on this framework, we develop EquityRank, a gradient-based algorithm that jointly optimizes user-side effectiveness and provider-side equity. Extensive offline and online simulations demonstrate that EquityRank offers improved trade-offs between effectiveness and fairness and adapts to heterogeneous provider needs.

The integration of Large Language Models (LLMs) into recommender systems has introduced a powerful paradigm for personalized recommedation. However, due to biases stemming from sensitive user attributes (e.g., gender, age), LLMs may generate discriminatory and unfair recommendations, even for users with similar preferences. Existing methods to improve LLM-based recommendation fairness, like prompt engineering or data-processing, are often unstable and fail to address unfairness rooted in pre-training data and model parameters. While fine-tuning approaches show promise, improving fairness typically comes at the cost of degraded recommendation performance, as the model’s parameters conflates user preferences with sensitive attributes. To address this, we propose \textbf{FairSpec}, a novel lightweight fine-tuning framework designed to enhance fairness while preserving recommendation quality. FairSpec has a Parameter-Efficient Mixture-of-Experts architecture (PE-MoE) with a fairness-aware router. The experts are specialized into utility experts and fairness experts, and the latter are adversarially trained to suppress sensitive information. Furthermore, FairSpec introduces Fairness-Aware Direct Preference Optimization (FA-DPO), which employs a composite reward function to jointly optimize recommendation performance and fairness. Results on three public benchmarks with four LLM backbones demonstrate that FairSpec achieves the best overall performance-fairness balance, attaining the top average rank across all metrics. Furthermore, visualizations of routing weights illustrate the strategic behavior and effectiveness of the proposed fairness-aware routing mechanism. Our code can be found at \url{https://github.com/qin2mu4/FairSpec}.

Fair re-ranking aims to promote long-tail items and enhance diversity within groups in information retrieval. While previous research on online fairness-aware re-ranking has shown promising outcomes, our comprehensive evaluation of online fair re-ranking methods over 20 settings reveals significant performance disparities among existing methods. To uncover the root causes of these inconsistencies, we reformulate fair re-ranking within an attentional market framework governed by a Walrasian Equilibrium, where the fairness is treated as a taxation cost. This market-based formulation is then coupled with manifold optimization, demonstrating that seeking this equilibrium is equivalent to performing gradient descent on a specific ranking manifold constructed by the market. Different re-ranking settings induce distinct manifold geometries, and these intrinsic geometric differences dictate the gradient landscapes and optimization trajectories. We propose ManifoldRank, an efficient online fair re-ranking algorithm. ManifoldRank adjusts gradients to align with the ranking manifold, considering various contextual settings. On the supply side, it incorporates a gradient adjustment based on different fairness requirements, accounting for associated costs. On the demand side, it empirically predicts an additional gradient adjustment term derived from the ranking scores. By integrating these two gradient adjustments, ManifoldRank effectively balances fairness and accuracy. Experimental results across multiple datasets confirm ManifoldRank's effectiveness.

Current Information Retrieval (IR) systems, including conversational IR paradigms enhanced by Retrieval-Augmented Generation (RAG), assume that users can formulate coherent queries and reliably interpret retrieved information. In dementia care contexts, these assumptions fail: queries are unstable, feedback signals are unreliable, relevance fluctuates with cognitive state, and multi-user interaction is essential. Dementia impacts over 50 million people globally, yet current RAG research has yet to consider use cases where users have dementia or Mild Cognitive Impairment (MCI), and so RAG systems continue to be designed without considering people with dementia, MCI, or their caregivers as users, creating a fundamental mismatch between system capabilities and this population's needs. Our perspective outlines functional requirements, not optional features, for RAG systems serving cognitively diverse populations, including temporal user modelling that tracks cognitive trajectories over time, caregiver-in-the-loop retrieval that enables verification and oversight, consent-aware evidence access that respects fluctuating capacity, and cognitive-load-aware presentation that adapts complexity to comprehension level. These requirements would collectively transform RAG from a cognitively-stable user paradigm into a cognitively-adaptive one. Drawing on participatory workshops with 41 members of the public and dementia-care community, our perspective was shaped through empirical evidence from those with lived experience of dementia or MCI. We advocate for inclusive IR through our cognitively-adaptive RAG that supports cognitive independence in dementia care, and call on the IR community to advance accessibility and equity through inclusive retrieval system design.

Large language models (LLMs) frequently prioritize parametric world knowledge over provided context -- a failure mode that is particularly catastrophic in enterprise or counterfactual settings where local facts contradict web-scale training data. While modern reasoning models improve general response quality, they fail to resolve these underlying prior knowledge biases even when generating a high volume of costly thinking tokens. We propose a context-editing framework that addresses this by performing selective abstraction over entities that appear in both the context and the question. Our approach replaces these overlapping entities with typed, indexed identifiers (e.g., Paris with City łangle1\rangle), suppressing spurious lexical associations while preserving the relational structure required for grounded reasoning. This mechanism is decoupled into an offline preprocessing stage and a lightweight inference-time substitution, requiring no model retraining. Experiments across counterfactual and multi-hop knowledge-conflict benchmarks demonstrate consistent accuracy gains across many model families, open- and closed-sourced, especially for small to medium sized LLMs (0.5B to 18B). Notably, our framework yields up to a 25% improvement over state-of-the-art instruct models and 13% over reasoning models, establishing symbolic abstraction as a highly cost-efficient solution for ensuring context fidelity in LLMs.

Large language models (LLMs) have achieved remarkable success across a wide range of applications especially when augmented by external knowledge through retrieval-augmented generation (RAG). Despite their widespread adoption, recent studies have shown that LLMs often struggle to perform faithful reasoning when conflicting knowledge is retrieved. However, existing work primarily focuses on conflicts between external knowledge and the parametric knowledge of LLMs, leaving conflicts across external knowledge largely unexplored. Meanwhile, modern RAG systems increasingly emphasize the integration of unstructured text and (semi-)structured data like knowledge graphs (KGs) to improve knowledge completeness and reasoning faithfulness. To address this gap, we introduce ConflictQA, a novel benchmark that systematically instantiates conflicts between textual evidence and KG evidence. Extensive evaluations across representative LLMs reveal that, facing such cross-source conflicts, LLMs often fail to identify reliable evidence for correct reasoning. Instead, LLMs become more sensitive to prompting choices and tend to rely exclusively on either KG or textual evidence, resulting in incorrect responses. Based on these findings, we further propose XoT, a two-stage explanation-based thinking framework tailored for reasoning over heterogeneous conflicting evidence, and verify its effectiveness with extensive experiments.

Standard Retrieval-Augmented Generation (RAG) systems predominantly rely on semantic relevance as a proxy for utility. However, this assumption collapses in realistic decision-making scenarios where user queries are laden with cognitive biases, such as false premises or confirmation bias. In such cases, maximizing relevance paradoxically promotes the retrieval of sycophantic evidence that reinforces hallucinations, a critical failure we term the ``Relevance-Robustness Gap''. To bridge this gap, we propose CoRM-RAG (Counterfactual Risk Minimization for RAG), a framework that aligns retrieval with decision safety rather than mere similarity. Grounded in causal intervention, we introduce a Cognitive Perturbation Protocol to simulate user biases during training, which is then distilled into a lightweight Evidence Critic. This scoring module learns to identify documents that possess sufficient evidential strength to steer the model toward correctness despite adversarial query perturbations. Extensive experiments on decision-making benchmarks demonstrate that CoRM-RAG significantly outperforms strong dense retrievers and LLM-based rerankers in adversarial settings, while enabling effective risk-aware abstention through reliable robustness scoring. Our code is available at https://github.com/PeiYangLiu/CoRM-RAG.git.

Modern information retrieval (IR) is no longer consumed primarily by humans but increasingly by large language models (LLMs) via retrieval-augmented generation (RAG) and agentic search. Unlike human users, LLMs are constrained by limited attention budgets and are uniquely vulnerable to noise; misleading or irrelevant information is no longer just a nuisance, but a direct cause of hallucinations and reasoning failures. In this perspective paper, we argue that denoising-maximizing usable evidence density and verifiability within a context window-is becoming the primary bottleneck across the full information access pipeline. We conceptualize this paradigm shift through a four-stage framework of IR challenges: from inaccessible to undiscoverable, to misaligned, and finally to unverifiable. Furthermore, we provide a pipeline-organized taxonomy of signal-to-noise optimization techniques, spanning indexing, retrieval, context engineering, verification, and agentic workflow. We also present research works on information denoising in domains that rely heavily on retrieval such as lifelong assistant, coding agent, deep research, and multimodal understanding.

Retrieval-Augmented Generation (RAG) systems rely on retrieved documents being concatenated into a model's input context, making both document ordering and context size critical yet controversial design choices. Prior work reports position-based effects such as lost in the middle and related long-context phenomena. However, empirical findings remain inconsistent and hard to reproduce across models, datasets, and evaluation protocols. In this paper, we present a systematic reproducibility study that revisits these claims and examines how they evolve with contemporary LLMs under a controlled evaluation framework. We first show that topic sampling is a major source of variance: small topic sets can mask or exaggerate ordering effects. Based on repeated subset sampling across multiple topic budgets, we provide a practical calibration procedure that identifies topic counts yielding stable trends at feasible cost. Using these fixed topic sets, we then reproduce and extend results on position sensitivity, re-evaluating lost in the middle and positional biases in modern LLMs. Then, we also study a more realistic RAG scenario in which relevance is mediated by a retriever rather than oracle access to ground-truth documents. In this setting, we re-examine a recent industry study and identify discrepancies to evaluation choices such as limited topic coverage and reliance on LLM-based judges. Finally, we conduct an analysis of how retrieval order and context size affect downstream LLM performance under imperfect retrieval. Our results demonstrate that both factors interact strongly with retrieval quality and model choice, and that conclusions drawn from idealised setups do not always transfer to real-world RAG pipelines. We release all code and configurations to support reproducibility and future work on robust RAG evaluation.

Multi-behavior recommendation enriches user preference modeling by incorporating diverse auxiliary interactions. However, most existing methods simply treat interactions without target behaviors as absolute negative feedback. This strategy ignores an important intermediate state known as user hesitation, where users exhibit strong intent but fail to complete the final conversion due to various reasons. Consequently, models cannot distinguish true disinterest from intended but hesitant behavior, which introduces substantial noise into preference modeling. To address this issue, we propose a novel framework named Calibrated Hesitation Analysis for Multi-Behavior Recommendation via Mutual Information Maximization (CHARM). Specifically, we aggregate auxiliary behaviors that lead to successful conversions into latent intent representations and train an inference network by maximizing the mutual information between these intents and observed target behaviors. We then apply this network to auxiliary behaviors without conversion, under the assumption that the conversion had occurred, in order to infer latent conversion probabilities and identify high-intent hesitation candidates. Furthermore, to distinguish genuine hesitation from interaction termination caused by competing item choices, we design a competitor substitution penalty strategy to refine hesitation confidence scores. Finally, the calibrated hesitation set is incorporated into the recommendation process to improve ranking quality. Extensive experiments on three real-world datasets demonstrate that CHARM consistently outperforms existing state-of-the-art methods. The source code is available at https://github.com/city59/CHARM.

Multi-behavior recommendation, unlike traditional single-behavior recommendation focusing solely on the target behavior (e.g., purchase), exploits multiple types of user-item interactions (e.g., view, favorite, add-to-cart, purchase) to mitigate data sparsity and learn more comprehensive user preferences. However, real-world data is inherently transmitted in a streaming manner, and most existing multi-behavior recommendation models are trained on static offline datasets, which cannot efficiently handle continuously arriving data streams. Full retraining incurs high computational cost and long training time. In contrast, fine-tuning is efficient but prone to ''catastrophic forgetting'', resulting in performance degradation. Incremental recommendation techniques offer a potential solution, but the existing methods are mostly designed for cases with one single behavior, ignoring valuable information contained in heterogeneous behaviors. To fill this gap, we introduce a new and important problem, i.e., incremental multi-behavior recommendation (IMBR), requiring a model to handle streaming data while utilizing valuable heterogeneous behaviors. To this end, we propose a novel solution called Behavior-aware Incremental Graph Convolution Network with Multi-Task Learning (BIGCN-MTL) for IMBR. Specifically, our BIGCN-MTL contains two key modules: behavior-aware incremental graph convolution network (BIGCN), which employs incremental cascading graph convolution to preserve both historical neighborhood information and behavior cascading dependency, and historical interest distillation (HID), which maintains users' preference stability and prevents behavioral semantic forgetting via non-interest consistency constraint and preference contrastive distillation. Extensive experiments on four real-world datasets demonstrate that our BIGCN-MTL achieves superior recommendation performance and improves training efficiency compared to full retraining. The source code of our BIGCN-MTL is available at https://github.com/codingPeter1/BIGCN-MTL.

Modern applications provided personalized recommendations across diverse scenarios, including the homepage, local pages, and live streams on platforms like TikTok. These scenarios exhibit varying user behavior patterns, resulting in heterogeneous yet interrelated distributions. Existing Multi-Scenario Recommendation (MSR) methods usually use parameter-sharing networks for shared features and scenario-specific networks for unique features. However, these methods fail to handle different distribution across scenarios, resulting in representation entanglement and localization, which hinder effective knowledge transfer and compromise performance. In this paper, we propose a Distribution-aware Re-representations (DAR) method for MSR. Its core idea is to construct distribution-aware prototype spaces and learn disentangled re-representations around global prototypes. Specifically, DAR employs a Multi-gate Mixture of Experts (MMoE) to obtain scenario-shared representations, and uses independent networks to learn scenario-specific representations. These representations are then projected into scenario-shared and scenario-specific prototype spaces, producing scenario-shared re-representations (capturing global information) and scenario-specific re-representations (focusing on distributional differences). During this process, DAR utilizes Unbalanced Optimal Transport (UOT) to compute the transport relationships between representations and global prototypes, taking these as pseudo-labels for re-representation learning. Moreover, to prevent prototype entanglement, a matrix orthogonalization constraint ensures independence among global prototypes. The effectiveness of DAR is demonstrated through extensive offline experiments conducted on four datasets, as well as online A/B tests on a video platform.

Search and recommendation (S&R) are fundamental components of modern commercial platforms, enabling users to access and explore information efficiently. User behaviors in these scenarios reflect different aspects of user intent, providing an opportunity for joint modeling of S&R. However, effectively leveraging search logs to enhance recommendation remains a challenging task. Existing methods often encode S&R histories either jointly or separately; however, they tend to regard all search signals as equally informative, thereby neglecting that many search behaviors can be irrelevant or even detrimental to recommendation performance. In practice, however, search histories frequently contain noisy or outdated behaviors that may introduce spurious correlations and degrade recommendation performance. Motivated by the human decision-making process, where one first identifies recommendation intent and then selectively reasons about relevant search signals, we propose LCR-SER, a latent cross reasoning method for search-enhanced recommendation. LCR-SER first encodes the user's S&R history into a unified latent representation that captures users' global interests. It then performs iterative reasoning in the latent space to dynamically identify informative search signals that are most relevant to the recommendation. To further guide this reasoning process, we introduce contrastive learning to align the reasoning states with the target items. In addition, we employ reinforcement learning to directly optimize ranking-oriented metrics, enabling LCR-SER to refine its reasoning strategy toward improved recommendation performance. Experiments on public datasets demonstrate that LCR-SER consistently outperforms strong baselines, validating the effectiveness of latent reasoning in enhancing search-aware recommendation.

Price is a key determinant of user purchase behavior in e-commerce. As traditional recommendation techniques typically fail to account for price information, Price-Aware Recommender Systems have emerged in response, aiming to explicitly incorporate price as a core feature to enhance the overall quality and relevance of recommendations. Recent advances include session-based approaches such as CoHHN (Heterogeneous Hypergraphs), PASBR (Graph Neural Network), and the collaborative-filtering approach PUP (Graph Convolutional Networks). All reporting improved performance on their respective evaluation datasets. However, these methods have never been compared under controlled experimental conditions, limiting our understanding of their relative performance. This work addresses this gap through a reproducibility study over CoHHN, PASBR, and PUP. We successfully conducted the first cross-evaluation across six public datasets using standardized metrics (HR@20, MRR@20, NDCG@20). Our findings indicate that CoHHN provides the strongest and most consistent performance across evaluation metrics and datasets, while PASBR shows domain-limited competitiveness and PUP exhibits the weakest generalization, highlighting CoHHN's heterogeneous hypergraph architecture as the most robust framework for Price-Aware Recommendation. Source code available at https://anonymous.4open.science/r/reproducibility-sigir2026-E2CE.

Next point-of-interest (POI) recommendation aims to predict users' next destinations from their mobility trajectories. Nevertheless, accurately capturing user preferences from check-in data remains challenging, as user behaviors are shaped by complex contextual factors and varying visit intentions. Despite substantial progress, existing methods still face two critical limitations. (i) They typically assume that all check-in behaviors faithfully reflect users' intrinsic preferences, overlooking behaviors driven by external contextual factors (e.g., work requirements) that misalign with users' authentic preference space, limiting the ability to understand dynamic user behavior patterns. (ii) Some of the POIs that users check in at may just be the closest available options to their expected visit intentions, rather than places they genuinely prefer. Existing methods neglect users' expected intentions, leading to inaccurate identification of users' authentic preferences. In this paper, we propose a novel next POI recommendation method (MPDC) that leverages Multi-Perspective modeling and Diffusion-based Contrastive learning. From the behavior-aware perspective, a behavior-aware graph prompt module mitigates external contextual influences via bidirectional sequence modeling and a preference filter, while capturing visit preferences. From the spatial-aware perspective, a spatial-aware hyperbolic graph module leverages hyperbolic space to model higher-order, nonlinear geographic patterns, effectively capturing the influence of spatial factors. In addition, a diffusion-based preference contrast module leverages the generative ability of diffusion models to learn latent expected location preferences that reflect user intent under different contextual conditions. Comprehensive experiments across five real-world datasets demonstrate that MPDC significantly outperforms state-of-the-art algorithms. Our code is available at https://github.com/LanPangxiang/LaMDA2026.

The post-click Conversion Rate (CVR) prediction plays a pivotal role in industrial recommender systems, directly serving as a decisive factor for ranking strategies and revenue optimization. While Entire Space Multi-Task Models (e.g., ESMM) have become the dominant framework by effectively addressing sample selection bias and data sparsity, they rely heavily on the assumption that the observed click and conversion labels perfectly reflect user preferences. In real-world scenarios, however, observed data is inevitably contaminated by label noise. Crucially, within the entire space modeling framework, this noise exhibits a complex hierarchical structure, where the superimposition of noise from antecedent click labels and subsequent conversion labels results in conventional single-task denoising methods being ineffective. To bridge this gap, we propose HiDe-ESMM (Hierarchical Denoising Entire Space Multi-Task Model), a unified framework designed to systematically mitigate hierarchical label noise. First, adopting the loss correction paradigm, we derive a statistically unbiased CTCVR loss via backward correction and identify the requisite noise rates based on the weak separability assumption. Second, to alleviate the numerical instability inherent in the unbiased loss approach, we introduce a robust forward correction strategy as an efficient alternative and theoretically prove the consistency of its optimal solution with the optimal solution on clean data. Extensive experiments on one semi-synthetic dataset and two real-world industrial datasets demonstrate that HiDe-ESMM significantly outperforms state-of-the-art baselines, validating its robustness in handling hierarchical noisy labels.

Click-through Rate (CTR) prediction is a crucial task in online advertising and recommender systems. Theoretically, proper ensemble of multiple different CTR prediction models can improve the prediction effectiveness. Unfortunately, most of the existing ensemble learning methods for CTR prediction are designed for cross-sectional data and neglect user historical behavior sequence which is important for predicting user's future click behavior. To address the above issues, we propose a Hypergraph Diffusion-based Sequential Ensemble framework for CTR prediction (HDSE). Specifically, considering the inherent capability of diffusion models in space exploration, we design a generalized conditional diffusion model, which adaptively identifies critical information from diverse sequential models, to capture user's dynamic interest evolution across diverse contexts. To avoid corrupting the item dependencies caused by isotropic Gaussian noise used in traditional diffusion models, we construct a behavior hypergraph and design an anisotropic hypergraph-based smoothing operator to inject structure-aware noise for better exploring the user's interest space. For large-scale application scenarios, we propose a computationally efficient approximation method for estimating hypergraph propagation matrix in the smoothing operator. Extensive experiments on three real-world datasets demonstrate the effectiveness of the proposed model.

Online platforms such as Amazon and Netflix serve users across multiple countries and regions, underscoring the importance of multi-market recommendation (MMR). Most MMR methods adopt a pre-training and fine-tuning paradigm, in which a unified model is first trained on centralized, global data and subsequently adapted to specific markets. However, this approach ignores the privacy of market data. While traditional federated learning preserves privacy, it typically aims to obtain a global model by aggregating model parameters and does not account for significant market heterogeneity. Additionally, because ID spaces are disjoint across markets, embedding-based aggregation strategies become ineffective. To overcome these challenges, we propose a federated collaborative signal quantization (FedMM) method for multi-market click-through rate (CTR) prediction. Our core idea leverages a discrete codebook mechanism to achieve privacy-preserving transmission and align disjoint ID spaces. We further employ a hierarchical codebook structure to capture cross-market shared patterns and market-specific characteristics. Specifically, we deploy a residual quantized variational autoencoder (RQ-VAE) with a dual-layer codebook mechanism for each market to quantize collaborative embeddings. The first layer utilizes a global federated codebook, updated via aggregation to capture universally shared collaborative patterns, while the second layer maintains a local codebook to learn market-specific semantics. Finally, the learned discrete codes, which integrate both general and specific collaborative signals, are incorporated into downstream CTR models to enhance prediction accuracy across all markets. Extensive experiments on benchmark datasets demonstrate that FedMM significantly improves recommendation performance with privacy guarantees.

Inspired by the success of language models (LM), scaling up deep learning recommendation systems (DLRS) has become a recent trend in the community.All previous methods tend to scale up the model parameters during training time.However, how to efficiently utilize and scale up computational resources during test time remains underexplored, which can prove to be a scaling-efficient approach and bring orthogonal improvements in LM domains.The key point in applying test-time scaling to DLRS lies in effectively generating diverse yet meaningful outputs for the same instance.We propose two ways: One is to explore the heterogeneity of different model architectures.The other is to utilize the randomness of model initialization under a homogeneous architecture.The evaluation is conducted across eight models, including both classic and SOTA models, on three benchmarks. Sufficient evidence proves the effectiveness of both solutions.We further prove that under the same inference budget, test-time scaling can outperform parameter scaling.Our test-time scaling can also be seamlessly accelerated with the increase in parallel servers when deployed online, without affecting the inference time on the user side.Code is available here\footnote{https://github.com/aTitye/TTS4CTR}.

Industrial large-scale recommendation models (LRMs) face the challenge of jointly modeling long-range user behavior sequences and heterogeneous non-sequential features under strict efficiency constraints. However, most existing architectures employ a decoupled pipeline: long sequences are first compressed with a query-token based sequence compressor like LONGER, followed by fusion with dense features through token-mixing modules like RankMixer, which thereby limits both the representation capacity and the interaction flexibility. This paper presents HyFormer, a unified hybrid transformer architecture that tightly integrates long-sequence modeling and feature interaction into a single backbone. From the perspective of sequence modeling, we revisit and redesign query tokens in LRMs, and frame the LRM modeling task as an alternating optimization process that integrates two core components: Query Decoding which expands non-sequential features into Global Tokens and performs long sequence decoding over layer-wise key-value representations of long behavioral sequences; and Query Boosting which enhances cross-query and cross-sequence heterogeneous interactions via efficient token mixing. The two complementary mechanisms are performed iteratively to refine semantic representations across layers. Extensive experiments on billion-scale industrial datasets demonstrate that HyFormer consistently outperforms strong LONGER and RankMixer baselines under comparable parameter and FLOPs budgets, while exhibiting superior scaling behavior with increasing parameters and FLOPs. Large-scale online A/B tests in high-traffic production systems further validate its effectiveness, showing significant gains over deployed state-of-the-art models. These results highlight the practicality and scalability of HyFormer as a unified modeling framework for industrial LRMs.

Recent progress in scaling large models has motivated recommender systems to increase model depth and capacity to better leverage massive behavioral data. However, recommendation inputs are high-dimensional and extremely sparse, and simply scaling dense backbones (e.g., deep MLPs) often yields diminishing returns or even performance degradation. Our analysis of industrial CTR models reveals a phenomenon of implicit connection sparsity: most learned connection weights tend towards zero, while only a small fraction remain prominent. This indicates a structural mismatch between dense connectivity and sparse recommendation data; by compelling the model to process vast low-utility connections instead of valid signals, the dense architecture itself becomes the primary bottleneck to effective pattern modeling. We propose SSR (Explicit Sparsity for Scalable Recommendation), a framework that incorporates sparsity explicitly into the architecture. SSR employs a multi-view ''filter-then-fuse'' mechanism, decomposing inputs into parallel views for dimension-level sparse filtering followed by dense fusion. Specifically, we realize the sparsity via two strategies: a Static Random Filter that achieves efficient structural sparsity via fixed dimension subsets, and Iterative Competitive Sparse (ICS), a differentiable dynamic mechanism that employs bio-inspired competition to adaptively retain high-response dimensions. Experiments on three public datasets and a billion-scale industrial dataset from AliExpress (a global e-commerce platform) show that SSR outperforms state-of-the-art baselines under similar budgets. Crucially, SSR exhibits superior scalability, delivering continuous performance gains where dense models saturate. The code is available at https://github.com/Atticus666/SSRNet.

Recent advances in generative recommenders adopt a two-stage paradigm: items are first tokenized into semantic IDs using a pretrained tokenizer, and then large language models (LLMs) are trained to generate the next item via sequence-to-sequence modeling. However, these two stages are optimized for different objectives: semantic reconstruction during tokenizer pretraining versus user interaction modeling during recommender training. This objective misalignment leads to two key limitations: (i) suboptimal static tokenization, where fixed token assignments fail to reflect diverse usage contexts; and (ii) discarded pretrained semantics, where pretrained knowledge—typically from language model embeddings—is overwritten during recommender training on user interactions. To address these limitations, we propose to learn DEcomposed COntextual Token Representations (DECOR), a unified framework that preserves pretrained semantics while enhancing the adaptability of token embeddings. DECOR introduces contextualized token composition to refine token embeddings based on user interaction context, and decomposed embedding fusion that integrates pretrained codebook embeddings with newly learned collaborative embeddings. Experiments on three real-world datasets demonstrate that DECOR consistently outperforms state-of-the-art baselines in recommendation performance. Our code is available at https://github.com/yliuaa/DECOR.git.

Generative recommendation is emerging as a transformative paradigm by directly generating recommended items, rather than relying on matching. Building such a system typically involves two key components: (1) optimizing the tokenizer to derive suitable item identifiers, and (2) training the recommender based on those identifiers. Existing approaches often treat these components separately—either sequentially or in alternation—overlooking their interdependence. This separation can lead to misalignment: the tokenizer is trained without direct guidance from the recommendation objective, potentially yielding suboptimal identifiers that degrade recommendation performance. To address this, we propose BLOGER, a Bi-Level Optimization for GEnerative Recommendation framework, which explicitly models the interdependence between the tokenizer and the recommender in a unified optimization process. The lower level trains the recommender using tokenized sequences, while the upper level optimizes the tokenizer based on both the tokenization loss and recommendation loss. We adopt a meta-learning approach to solve this bi-level optimization efficiently, and introduce gradient surgery to mitigate gradient conflicts in the upper-level updates, thereby ensuring that item identifiers are both informative and recommendation-aligned. Extensive experiments on multiple real-world datasets demonstrate that BLOGER consistently outperforms state-of-the-art generative recommendation methods while maintaining practical efficiency with no significant additional computational overhead, effectively bridging the gap between item tokenization and autoregressive generation. We release our code at https://github.com/Ten-Mao/BLOGER.

Generative recommendation provides a novel paradigm in which each item is represented by a discrete semantic ID (SID) learned from rich content. Most methods treat SIDs as predefined and train recommenders under static indexing. In practice, SIDs are optimized only for content reconstruction rather than recommendation accuracy. This leads to an objective mismatch : the system optimizes an indexing loss to learn the SID, and a recommendation loss for interaction prediction, but because the tokenizer is trained independently, the recommendation loss cannot update it. A natural approach is to make semantic indexing differentiable so recommendation gradients can directly influence SID learning, but this often causes codebook collapse with only a few codes used. We attribute this to early deterministic assignments that limit codebook exploration, leading to imbalance and unstable optimization. In this paper, we therefore propose DIGER (Differentiable Semantic ID for GEnerative Recommendation). DIGER is a first step towards an effective differentiable semantic ID for generative recommendation. The Gumbel noise explicitly encourages early-stage exploration over codes, mitigating collapse and improving code utilization. To better balance exploration and convergence, we introduce two uncertainty decay strategies that reduce the Gumbel noise, enabling a gradual shift from early-stage exploration to the exploitation of learned SIDs. Extensive experiments across multiple public datasets demonstrate consistent improvements from differentiable semantic ID. These results confirm the effectiveness of aligning indexing and recommendation objectives through differentiable SIDs. This identifies differentiable SID as a promising area of study. Our code is released under https://github.com/junchen-fu/DIGER.

Generative recommendation frameworks typically represent items as discrete Semantic IDs (SIDs). While existing studies have sought to enhance SID construction by incorporating multimodal content, collaborative signals, or more advanced quantization techniques, learning high-quality SIDs still faces two key challenges: (1) The two-stage generative recommendation paradigm (SID construction and autoregressive generation) provides insufficient supervision for heterogeneous fusion, which hinders learning high-quality SIDs, and (2) non-uniform embeddings lead to codeword imbalance and generation bias. To address these challenges, we propose a novel generative recommendation framework, called CARD. CARD introduces a visual semantic unit that unifies textual, visual, and collaborative signals into a structured visual representation prior to encoding, enabling holistic semantic modeling and effectively alleviating the semantic gap, thereby reducing the reliance on supervision signals during SID learning. Furthermore, to deal with the highly non-uniform distribution of item semantic embeddings in recommendation scenarios, we develop a non-uniform quantization framework (NU-RQ-VAE), which incorporates a learnable and invertible non-uniform transformation into the quantization process to map skewed semantic distributions into a more balanced latent space, thereby significantly improving codebook utilization and quantization accuracy. Experiments on multiple datasets show that CARD consistently outperforms baseline methods under various settings; meanwhile, the proposed non-uniform transformation module is plug-and-play and remains robust across different quantization schemes. Code is available at https://github.com/HAI-UESTC/CARD.

Recently, generative recommendation has emerged as a promising paradigm, attracting significant research attention. The basic framework involves an item tokenizer, which represents each item as a sequence of codes serving as its identifier, and a generative recommender that predicts the next item by autoregressively generating the target item identifier. However, in existing methods, both the tokenizer and the recommender are typically domain-specific, limiting their ability for effective transfer or adaptation to new domains. To this end, we propose UTGRec, a Universal item Tokenization approach for transferable Generative Recommendation. Specifically, we design a universal item tokenizer for encoding rich item semantics by adapting a multimodal large language model (MLLM). By devising tree-structured codebooks, we discretize content representations into corresponding codes for item tokenization. To effectively learn the universal item tokenizer on multiple domains, we introduce two key techniques in our approach. For raw content reconstruction, we employ dual lightweight decoders to reconstruct item text and images from discrete representations to capture general knowledge embedded in the content. For collaborative knowledge integration, we assume that co-occurring items are similar and integrate collaborative signals through co-occurrence alignment and reconstruction. Finally, we present a joint learning framework to pre-train and adapt the transferable generative recommender across multiple domains. Extensive experiments on four public datasets demonstrate the superiority of UTGRec compared to both traditional and generative recommendation baselines.

Generative recommendation commonly adopts a two-stage pipeline in which a learnable tokenizer maps items to discrete token sequences (i.e., identifiers) and an autoregressive generative recommender model (GRM) performs prediction based on these identifiers. Recent tokenizers further incorporate collaborative signals so that items with similar user-behavior patterns receive similar codes, substantially improving recommendation quality. However, real-world environments evolve continuously: new items cause identifier collision and shifts, while new interactions induce collaborative drift in existing items (e.g., changing co-occurrence patterns and popularity). Fully retraining both tokenizer and GRM is often prohibitively expensive, yet naively fine-tuning the tokenizer can alter token sequences for the majority of existing items, undermining the GRM's learned token–embedding alignment. To balance plasticity and stability for collaborative tokenizers, we propose DACT, a Drift-Aware Continual Tokenization framework with two stages: (i) tokenizer fine-tuning, augmented with a jointly trained Collaborative Drift Identification Module (CDIM) that outputs item-level drift confidence and enables differentiated optimization for drifting and stationary items; and (ii) hierarchical code reassignment using a relaxed-to-strict strategy to update token sequences while limiting unnecessary changes. Experiments on three real-world datasets with two representative GRMs show that DACT consistently achieves better performance than baselines, demonstrating effective adaptation to collaborative evolution with reduced disruption to prior knowledge. Our implementation is publicly available at https://github.com/HomesAmaranta/DACT for reproducibility.

In recent years, multi-modal multi-domain fake news detection has garnered increasing attention. Nevertheless, this direction presents two significant challenges: (1) Failure to Capture Cross-Instance Narrative Consistency: existing models usually evaluate each news in isolation, fail to capture cross-instance narrative consistency, and thus struggle to address the spread of cluster-based fake news driven by social media; (2) Lack of Domain-Specific Knowledge for Reasoning: conventional models, which rely solely on knowledge encoded in their parameters during training, struggle to generalize to new or data-scarce domains (e.g., emerging events or niche topics). To tackle these challenges, we introduce Retrieval-Augmented Multimodal Model for Fake News Detection (RAMM). First, RAMM employs a Multimodal Large Language Model (MLLM) as its backbone to capture cross-modal semantic information from news samples. Second, RAMM incorporates an Abstract Narrative Alignment Module. This component adaptively extracts abstract narrative consistency from diverse instances across distinct domains, aggregates relevant knowledge, and thereby enables the modeling of high-level narrative information. Finally, RAMM introduces a Semantic Representation Alignment Module, which aligns the model's decision-making paradigm with that of humans—specifically, it shifts the model's reasoning process from direct inference on multimodal features to an instance-based analogical reasoning process. Extensive experimental results on three public datasets validate the efficacy of our proposed approach. Our code is available at the following link: \hrefhttps://github.com/li-yiheng/RAMM https://github.com/li-yiheng/RAMM

Propagation-based fake news detectors primarily extract structural patterns from news propagation trees via graph neural networks (GNNs), which are crucial for trustworthy information access on social platforms. However, these systems remain vulnerable to adversarial message injection, increasingly enabled by large language models (LLMs). Such attacks pollute both semantic and structural signals, causing GNN-based aggregators to fuse logically conflicting content and yield unreliable representations. To address this, we propose LLM-TKT, a novel framework that distills LLM-based narrative reasoning into lightweight GNNs for robust fake news detection. The framework operates in two stages. First, we construct an offline LLM-driven narrative hub to synthesize global propagation narratives and diagnose local node-level coherence. Second, we design a dual-level narrative alignment to learn the semantic invariance of propagation with the guidance of propagation narratives. It filters unreliable neighbor nodes via local consistency and optimizes graph representations via global anchoring. Experiments on three real-world datasets demonstrate that LLM-TKT significantly outperforms existing methods, particularly in defending against sophisticated LLM-driven injection attacks without incurring runtime LLM inference costs.

The rapid spread of multimodal fake news poses a serious societal threat, as its evolving nature and reliance on timely factual details challenge existing detection methods. Dynamic Retrieval-Augmented Generation provides a promising solution by triggering keyword-based retrieval and incorporating external knowledge, thus enabling both efficient and accurate evidence selection. However, it still faces challenges in addressing issues such as redundant retrieval, coarse similarity, and irrelevant evidence when applied to deceptive content. In this paper, we propose ExDR—an Explanation-driven Dynamic Retrieval-Augmented Generation framework for Multimodal Fake News Detection. Our framework systematically leverages model-generated explanations in both the retrieval triggering and evidence retrieval modules. It assesses triggering confidence from three complementary dimensions, constructs entity-aware indices by fusing deceptive entities, and retrieves contrastive evidence based on deception-specific features to challenge the initial claim and enhance the final prediction. Experiments on two benchmark datasets, AMG and MR2, demonstrate that ExDR consistently outperforms previous methods in retrieval triggering accuracy, retrieval quality, and overall detection performance, highlighting its effectiveness and generalization capability.

Retrieval-Augmented Generation (RAG) systems can amplify political bias from underlying web corpora. To empirically demonstrate this amplification, we first analyze 16,254 documents from the C4 dataset and 24,300 LLM-generated responses, revealing significant left-leaning and supportive stance bias that can propagate strongly from retrieval to generation. To mitigate this amplification of political bias, we propose FAIR-RAG, an end-to-end framework integrating (1) multi-LLM persona-based annotation, (2) a vector database with political-stance metadata, and (3) a multi-stage fairness engine designed for each of the three stages in RAG systems. FAIR-RAG achieves Attention Weighted Rank Fairness of 97.51 (82.1% improvement) and Perspective Balance of 51.01/82.37 (average 5.6% improvement over state-of-the-art) while maintaining high output quality (Context Precision: 0.974/0.975, Faithfulness: 0.994/0.996). Ablation studies confirm that all three components must operate collaboratively for optimal bias mitigation. This work provides a foundational framework for developing trustworthy and equitable AI information systems. All source code and experimental scripts are publicly available at: https://github.com/bigbases/FAIR-RAG.

As Large Language Models (LLMs) increasingly serve as interfaces for proprietary data (e.g., enterprise knowledge bases, legal statutes), ensuring their fidelity to trusted internal information is paramount. While integrating real-time web search can enhance model utility, it introduces a critical vulnerability: the ingestion of conflicting, misleading, or hallucinated content from the open web can override the model's adherence to its verified internal knowledge. We define this failure mode as search-induced distortion, a significant risk in high-stakes domains where the internal knowledge base serves as the absolute ground truth. To address this challenge, we present PurifAI, a proactive, model-agnostic, cache-level purification system designed for safety- and compliance-sensitive deployments. Rather than serving as a general fact-checking engine, PurifAI is explicitly designed to preserve knowledge alignment with a pre-defined trusted knowledge core. It automatically generates diagnostic probes from trusted documents to identify and neutralize searched web content that conflicts with the canonical internal source before such content distorts the LLM's responses. Extensive evaluations on news, encyclopedic, and legal domains show that PurifAI effectively improves alignment with the trusted core, identifying and blocking distortive content at the source and achieving repair success rates above 70% across mainstream LLMs. Our work offers a practical safeguard for enterprises and other high-stakes applications seeking to integrate web-augmented LLMs without compromising policy consistency and trusted knowledge integrity.

Recent advancements in Large Language Models (LLMs) have significantly extended context windows, igniting discussions about the necessity of Retrieval-Augmented Generation (RAG). U-NIAH, a unified Needle-In-A-Haystack (NIAH) framework, systematically evaluates LLMs and RAG methods in controlled long-context settings. It extends beyond traditional NIAH by incorporating more practical and complex scenarios like multi-needle, long-needle, and needle-in-needle configurations and leveraging the synthetic dataset to mitigate LLM biases. The experiments aim to address three research questions in long-context scenarios: (1) performance trade-offs between LLMs and RAG, (2) error patterns in RAG, and (3) RAG’s limitations in complex settings. Results show that smaller LLMs benefit more from RAG. In all settings, RAG achieves a win rate of 82.58% over direct answers. Additionally, it is found that retrieval noise and chunk ordering degrade RAG performance, and we further summarized typical error patterns, including omissions due to noise, hallucinations under high noise critical conditions, and self-doubt behaviors, as well as how these phenomena vary with context length. Finally, in some challenging scenarios, experiments show that deep reasoning models are more easily affected by distractors. These findings highlight the complementary roles of RAG and LLMs and offer actionable insights for optimizing deployment strategies.

High-quality relevance judgements over large query sets are essential for evaluating Information Retrieval (IR) systems, yet manual annotation remains costly and time-consuming. Large Language Models (LLMs) have recently shown promise as automatic relevance assessors, but their reliability is still limited. Most existing approaches rely on zero-shot prompting or in-context learning (ICL) with a small number of labelled examples. However, standard ICL treats examples as independent instances and fails to explicitly capture the underlying relevance criteria of a topic, restricting its ability to generalise to unseen query-document pairs. To address this limitation, we introduce Relevance Context Learning (RCL), a novel framework that leverages human relevance judgements to explicitly model topic-specific relevance criteria. Rather than directly using labelled examples for in-context prediction, RCL first prompts an LLM (Instructor LLM) to analyse sets of judged query-document pairs and generate explicit narratives that describe what constitutes relevance for a given topic. These relevance narratives are then used as structured prompts to guide a second LLM (Assessor LLM) in producing relevance judgements. To evaluate RCL in a realistic data collection setting, we propose a hybrid pooling strategy in which a shallow depth-k pool from participating systems is judged by human assessors, while the remaining documents are labelled by LLMs. Experimental results demonstrate that RCL substantially outperforms zero-shot prompting and consistently improves over standard ICL. Overall, our findings indicate that transforming relevance examples into explicit, context-aware relevance narratives is a more effective way of exploiting human judgements for LLM-based IR dataset construction.

Large Language Models (LLMs) are increasingly used as automated judges (LLM judges) to evaluate Information Retrieval (IR) systems, offering a cost-effective complement to human assessments. However, most prior work treats evaluation mainly as a tool for comparison rather than as a signal for improving the retrieval system. We study whether criterion grades from Multi-Criteria LLM-Judge relevance labeling, which decomposes relevance into Exactness, Coverage, Topicality, and Contextual Fit, can serve as effective ranking features for learning-to-rank (L2R). We evaluate this approach using manual relevance labels from TREC TREC DL 2019, DL 2020, and DL 2023. We then analyze how the Multi-Criteria LLM-Judge feature importance varies relative to retrieval scores across IR system performance levels. The results show that although weaker IR systems benefit the most from this treatment, we see improvements ranging from 14% to 36% in NDCG@20 across diverse IR systems. Contextual Fit often becomes the primary discriminator, exceeding the importance of the retrieval scores. As system quality increases, retrieval scores become the dominant feature, with Multi-Criteria LLM-Judge grades providing complementary, dataset-dependent signals. These findings show that LLM-based relevance labeling can go beyond comparison by turning Multi-Criteria LLM-Judge grades into effective ranking features for reranking.

The unjudged document problem, where systems that did not contribute to the original judgement pool may retrieve documents without a relevance judgement, is a key obstacle to the reuseability of test collections in information retrieval. This is commonly dealt with by treating unjudged documents as non-relevant, and recently, the use of large language models (LLMs) as a relevance judge (LLM-as-a-judge) emerged. However, this has been criticized, among other things, as circular, since the same LLM can be used as the ranker and the judge. We propose to train topic-specific relevance classifiers instead: By finetuning monoT5 with independent LoRA weight adaptation on the judgments of a single assessor for a single topic's pool, we align it to that assessor's notion of relevance for that topic. The system rankings obtained through our classifier's relevance judgments achieve a Spearmans' ρ correlation of >0.94 with ground truth system rankings. As little as 128 initial human judgments per topic suffice to improve the comparability of models, while achieving more reliability than existing LLM-as-a-judge approaches, and maintaining human judgments as the gold standard for retrieval evaluation. Code, models, and data are available.

Podcasts have continued to grow in popularity over the last two decades, with more than 4.52 million podcasts and 584 million listeners across the globe in 2025. Developing effective search systems for web-scale podcast corpora is of vital importance. Previous research has approached the search task primarily through text representation using a single transcription of the audio content via automatic speech recognition (ASR) models. However, there is currently limited understanding about how variation in podcast representations influences ranking, retrieval, and relevance assessment. In this paper, we provide a comprehensive analysis of the effectiveness of the podcast search problem from three different sources of variation. First, we create a set of 5,745 unique, automatically generated query variations derived from the TREC 2020 and 2021 podcast track topic descriptions. Second, to enable a comprehensive and reproducible analysis, we build a diverse pool composed of a set of 17 different ranked retrieval runs (system variations) using state-of-the-art (SOTA) information retrieval (IR) systems for all of the query variations over a set of corpus (transcript) variations. Third, we use five large language models (LLMs) to automatically assess the relevance of query-document pairs resulting from every query-system-transcription combination, yielding over 1.8 million judged pairs. Our analysis studies the impact of these variations on both search and judgment effectiveness, resulting in a series of practical recommendations for podcast and audio retrieval tasks.

Cranfield-style retrieval evaluations with too few or too many relevant documents or with low inter-assessor agreement on relevance can reduce the reliability of observations. In evaluations with human assessors, information needs are often formalized as retrieval topics to avoid an excessive number of relevant documents while maintaining good agreement. However, emerging evaluation setups that use Large Language Models (LLMs) as relevance assessors often use only queries, potentially decreasing the reliability. To study whether LLM relevance assessors benefit from formalized information needs, we synthetically formalize information needs with LLMs into topics that follow the established structure from previous human relevance assessments (i.e., descriptions and narratives). We compare assessors using synthetically formalized topics against the LLM-default query-only assessor on the~2019/2020~editions of TREC Deep Learning and Robust04. We find that assessors without formalization judge many more documents relevant and have a lower agreement, leading to reduced reliability in retrieval evaluations. Furthermore, we show that the formalized topics improve agreement between human and LLM relevance judgments, even when the topics are not highly similar to their human counterparts. Our findings indicate that LLM relevance assessors should use formalized information needs, as is standard for human assessment, and synthetically formalize topics when no human formalization exists to improve evaluation reliability.

Neural retrievers are often trained on large-scale triplet data comprising a query, a positive passage, and a set of hard negatives. In practice, hard-negative mining can introduce false negatives and other ambiguous negatives, including passages that are relevant or contain partial answers to the query. Such label noise yields inconsistent supervision and can degrade retrieval effectiveness. We propose ARHN (Answer-centric Relabeling of Hard Negatives), a two-stage framework that leverages open-source LLMs to refine hard negative samples using answer-centric relevance signals. In the first stage, for each query–passage pair, ARHN prompts the LLM to generate a passage-grounded answer snippet or to indicate that the passage does not support an answer. In the second stage, ARHN applies an LLM-based listwise ranking over the candidate set to order passages by direct answerability to the query. Passages ranked above the original positive are relabeled to additional positives. Among passages ranked below the positive, ARHN exclude any that contain an answer snippet from the negative set to avoid ambiguous supervision. We evaluated ARHN on the BEIR benchmark under three configurations: relabeling only, filtering only, and their combination. Across datasets, the combined strategy consistently improves over either step in isolation, indicating that jointly relabeling false negatives and filtering ambiguous negatives yields cleaner supervision for training neural retrieval models. By relying strictly on open-source models, ARHN establishes a cost-effective and scalable refinement pipeline suitable for large-scale training.

Dense retrieval usually fails in two ways: ranking non-relevant documents too high, or ranking relevant documents too low. We focus on the second case from the query side: documents that are genuinely relevant but receive low retrieval scores under certain query formulations, forming hard-positive query–document pairs. To study this failure mode systematically, we build on recent token-alignment work, which analyzes vocabulary logits obtained by projecting retriever representations through an architecture-specific token prediction head and shows that low overlap among top-ranked logit tokens can cause relevant documents to be scored low. This suggests a practical route to hard positives: generate relevant queries that rely on document phrases that receive low ranks under these logits. We therefore define model-specific Token and Phrase Representativeness Scores (TRS/PRS) to discover tokens and key phrases that appear in a document but are poorly expressed by its embedding. Using high-PRS phrases as anchors, we automatically construct challenging yet relevant queries, yielding hard-positive query–document pairs. Experiments across multiple retrievers and datasets show that PRS-anchored queries induce substantially larger drops for dense retrievers than other query construction strategies, revealing a dense-specific weakness. Moreover, when mixed with standard hard negatives during contrastive fine-tuning, these hard positives improve retrieval on our constructed hard-positive benchmark and can also improve standard in-domain retrieval benchmarks. Our code is publicly available at https://github.com/wzy2001wzy/HardPositive.

Complex information needs may involve set-compositional queries using conjunction, disjunction, and exclusion, yet it remains unclear whether current retrieval paradigms genuinely satisfy such constraints or exploit `semantic shortcuts'. We conduct a reproducibility study to benchmark major retrieval families and reasoning-targeted methods on QUEST and QUEST+Variants, and introduce LIMIT+, a controlled benchmark where relevance depends on arbitrary attribute predicates and constraint satisfaction, and less on pretrained knowledge. Our findings show that (i) on QUEST, the best neural retrievers achieve an effectiveness that is more than double what can be achieved with BM25 (Recall@100 ${>}$0.41 vs.\ 0.20), but reasoning-targeted methods like ReasonIR and Search-R1 do not outperform general-purpose retrievers uniformly; (ii) on LIMIT+, gains fail to transfer, where the strongest QUEST method collapses from Recall@100${\approx}$0.42 to below 0.02, while classic lexical retrieval gains to ${\sim}$0.96. Lastly, (iii) stratifying by compositional depth reveals a consistent degradation across all methods, where algebraic sparse and lexical methods show more stable performance while dense approaches collapse.We release code and LIMIT+ data generation scripts to support future reproducibility and controlled evaluation.

The Hypencoder, proposed by Killingback et al., is a retrieval framework that replaces the fixed inner-product scoring function used in standard bi-encoders with a query-specific neural network (the $q$-net), whose weights are generated by a hypernetwork from the contextualized query embeddings. This design enables more expressive relevance estimation while preserving independent query and document encoding. In this work, we conduct a reproducibility study of the Hypencoder and extend the original analysis in three directions. Our reproduction confirms that the Hypencoder outperforms a similarly trained bi-encoder baseline on in-domain and out-of-domain benchmarks, and that the proposed efficient search algorithm substantially reduces query latency with minimal performance loss. On hard retrieval tasks, we find partial support: the Hypencoder outperforms the baseline on DL-Hard and FollowIR, but not on TREC TOT, where checkpoint incompatibility and fine-tuning sensitivity complicate full verification. Beyond reproduction, we investigate three extensions: (i)~integrating alternative pre-trained encoders into the Hypencoder framework, where we find that performance gains depend on the encoder and fine-tuning strategy; (ii)~comparing query latency against a Faiss-based bi-encoder pipeline, revealing that standard bi-encoder retrieval remains faster under both exhaustive and efficient search settings; and (iii)~evaluating adversarial robustness, where we find that the $q$-net's non-linear scoring does not provide a consistent robustness disadvantage over inner-product scoring. Our code is publicly available at \url{https://github.com/arneeichholtz/Hypencoder-reprod}.

Advances in embedding models for text, image, audio, and video drive progress across multiple domains, including retrieval-augmented generation, recommendation systems, and others. Many of these applications require an efficient method to retrieve items that are close to a given query in the embedding space while satisfying a filter condition based on the item's attributes, a problem known as filtered approximate nearest neighbor search (FANNS). By performing an in-depth literature analysis on FANNS, we identify a key gap in the research landscape: publicly available datasets with embedding vectors from state-of-the-art transformer-based text embedding models that contain abundant real-world attributes covering a broad spectrum of attribute types and value distributions. To fill this gap, we introduce the arxiv-for-fanns dataset of transformer-based embedding vectors for the abstracts of over 2.7 million arXiv papers, enriched with 11 real-world attributes such as authors and categories. We benchmark eleven different FANNS methods on our new dataset to evaluate their performance across different filter types, numbers of retrieved neighbors, dataset scales, and query selectivities. We distill our findings into eight key observations that guide users in selecting the most suitable FANNS method for their specific use cases.

Multi-vector retrieval models (e.g., ColBERTv2) offer high retrieval accuracy but suffer from efficiency problems at scale. Recently, several methods have been developed to enhance the efficiency of multi-vector retrieval. On one hand, the state-of-the-art GPU-based method PLAID-GPU exploits the massive parallelism of the GPU to accelerate computation, but it needs to process a considerable amount (e.g., ten thousand) of document candidates. On the other hand, the state-of-the-art (SOTA) CPU-based method IGP employs a more effective strategy to reduce the number of candidates, but fails to utilize the massive parallelism of the GPU. To get the best of both worlds, we propose GIGP+, a GPU-based method designed to achieve high parallelism and low computational overhead. Our contributions are: (1) an efficient candidate generation kernel that enjoys parallelism while retaining the effectiveness of IGP, (2) a score reordering mechanism that reduces the synchronization overhead and (3) a scheduling strategy for efficient batch processing. Our experiments demonstrate that GIGP+ achieves a 11.0× improvement in query per second (QPS) and reduces latency by 7.6× compared to PLAID-GPU, while maintaining equivalent retrieval accuracy. As for cloud pricing, GIGP+ delivers a 2.3× improvement in queries per dollar over SOTA CPU-based solutions.

Modern large-scale recommender systems rely on dense user and item embeddings, but using a full embedding table requires a large number of parameters. To reduce this cost, hashing tricks are commonly used to compress embedding tables. However, collisions and hard bucket assignments can degrade both performance and personalization. We introduce Graph Diffusion Gated Embeddings (GDE), which precompute multi-scale, type-separated heat-kernel diffusions from a small set of user/item seeds (via HK-relax) and convert them into degree-corrected probabilistic gating distributions. These gates mix a compact set of learnable seed-cluster embeddings to form node representations. Ablations show that the cross-only variant (cGDE), which uses only cross-type gating, achieves the strongest performance. Across both Top-K retrieval and click through rate prediction tasks, cGDE outperforms hashing-based baseline methods. Our code is available at https://github.com/latent-x/Graph-Diffusion-Gated-Embeddings-for-Recommender-Systems.

Question and answer (Q&A) platforms usually recommend question-answer pairs to meet users' knowledge acquisition needs, unlike traditional recommendations that recommend only one item. This makes user behaviors more complex, and presents two challenges for Q&A recommendation, including: the collaborative information entanglement, which means user feedback is influenced by either the question or the answer; and the semantic information entanglement, where questions are correlated with their corresponding answers, and correlations also exist among different question-answer pairs. Traditional recommendation methods treat the question-answer pair as a whole or only consider the answer as a single item, which overlooks the two challenges and cannot effectively model user interests. To address these challenges, we introduce a graph neural network model named Question & Answer Graph Collaborative Filtering (QAGCF). QAGCF creates graphs separately from collaborative and semantic views to disentangle the collaborative and semantic information of question-answer pairs. The collaborative view disentangles questions and answers to individually model collaborative information, while the semantic view captures the semantic information both within and between question-answer pairs. These views are further merged into a global graph to integrate the collaborative and semantic information. Polynomial-based graph filters are used to address the high heterophily issues of the global graph. Additionally, contrastive learning is utilized to obtain robust embeddings during training. Extensive experiments on industrial and public datasets demonstrate that QAGCF consistently outperforms baselines and achieves state-of-the-art results.

Graph models that manipulate the frequency spectrum of user-item interactions to separate preference signals from noise often report significant improvements, but concerns about evaluation rigor and reproducibility persist. We conduct a reproducibility and replicability study that examines three major families: (i) spectral denoising methods, (ii) graph signal processing (GSP) models, and (iii) spectral propagation approaches. Reproducing published pipelines reveals a polarized landscape: while several works are fully reproducible, others rely on flawed metric implementations and incomplete hyperparameter disclosures. In particular, we observe systematic inflation of Recall in the spectral denoising methods due to an implementation error, and theoretically invalid ranking metrics in GSP models due to unordered prediction lists; conversely, the graph filtering models are consistently reproducible. Beyond reproduction, we establish a unified evaluation protocol on four datasets with consistent splits and hyperparameter optimization for all baselines, showing that strong classical methods (e.g., SLIM, Item-kNN) remain highly competitive and that no single spectral model dominates across domains. We further analyze robustness under varying data sparsity and assess beyond-accuracy properties, finding that spectral filtering often improves catalog exploration even when accuracy gains are marginal. Our code is available at https://github.com/sisinflab/Mind_the_Metric_SIGIR-26.

In this work, we propose VerbaLightGCN, a novel LLM-based recommendation framework that integrates the semantic understanding of LLMs with user-item interaction modeling. Traditional collaborative filtering (CF) models typically embed user and item IDs into a latent space to capture interaction signals. However, pretrained LLMs cannot natively interpret these learned embeddings. To bridge this gap, VerbaLightGCN adopts a CF-as-text paradigm, in which collaborative signals are encoded in textual form and directly learned from the user–item interaction graph, and are then combined with semantic information to construct user and item profiles that function as latent embeddings. Inspired by LightGCN, our method retains its message-passing design but replaces numerical embedding computations with a Chain-of-Thought prompting mechanism. This enables LLMs to simulate the LightGCN aggregation process through natural language. The result is a recommendation framework that unifies semantic understanding with collaborative signals in a fully language-native form. Experiments show that VerbaLightGCN achieves superior performance to both zero-shot LLM-based and traditional CF-based baselines. Further analysis reveals that the user and item profiles generated by VerbaLightGCN effectively capture both semantic preferences and collaborative filtering signals.

Energy efficiency has become a critical requirement, driving recommendation systems for resource-constrained environments such as edge devices. Model quantization offers an effective way to build low-bitwidth models while preserving accuracy. However, user–item interaction graphs contain numerous nodes and complex topological structures, leading nodes to exhibit unique similarities and differences. Existing quantization methods uniformly process parameters in high-dimensional DNN layers (e.g., linear, convolutional, or attention layers), while inadequately capturing such similarities among node embeddings. This paper proposes GraphQ, a chunk-wise quantization framework for graph collaborative filtering that supports both the training and post-training phases in a unified perspective. Our core idea is to adaptively partition node embeddings into multiple chunks based on the distribution of embedding values, and then apply chunk-wise quantization. Specifically, for quantization-aware training (QAT), we introduce learnable low-precision quantization factors that partition node embeddings into multiple chunks and are dynamically updated following message passing. For post-training quantization (PTQ), we first cluster nodes and then partition their dimensions into chunks for weight clipping. Extensive experiments on four real-world datasets show that GraphQ outperforms state-of-the-art QAT methods by an average of 27.49% in Recall@10 under the 256-dimensional embedding and 2-bit settings, and surpasses PTQ methods by 78.64% on average under 4-bit settings.

Recommendation plays a crucial role in the modern Web ecosystem, powering personalized services across e-commerce, social platforms, and online content networks. To model complex user–item interactions in such Web environments, Graph Neural Networks (GNNs) have become a popular and effective approach due to their ability to capture relational dependencies. However, existing GNN-based methods face some challenges, such as limited capacity for nonlinear representation, inability to capture global structural information, and susceptibility to noise in user interaction data. Although self-supervised learning methods have been introduced to address these issues, these methods often overlook the intricate dependencies between users and items and fail to effectively utilize high-order global information. To address these challenges, we propose Fourier Kolmogorov-Arnold Network and Hypergraph Enhanced Contrastive Learning (FHCL) for recommendation. Our method constructs two complementary views: a graph generative view and a denoising view. In the graph generative view, we use the Fourier Kolmogorov-Arnold Network (Fourier KAN) to enhance the nonlinear representation capabilities by decomposing complex user-item interactions. Subsequently, we employ Variational Graph Auto-Encoders (VGAE) to reconstruct the graph structure, extracting meaningful structural information while mitigating the impact of noise. Then we use hypergraph learning to capture high-order global dependencies. In the denoising view, we introduce a denoising matrix to filter noisy edges and further combine hypergraph learning to improve user preferences. Finally, we integrate these views through contrastive learning to generate robust and accurate recommendations. Extensive experiments on two public datasets demonstrate the superior performance of FHCL, while comprehensive ablation studies validate the necessity and effectiveness of each component.

Negative sampling is significant for training sequential recommendation models under implicit feedback. The predominant strategy, self-guided hard negative sampling, selects negatives based on the model’s current state but suffers from three limitations: (1) the coupling between sampling and model updates triggers a vicious cycle that drives the model into local optima; (2) relying on current model parameters narrows sampling to a small region of the item space, reducing diversity and harming generalization; (3) identifying a hard negative requires scoring the entire candidate pool, causing substantial computational overhead with minimal information gain. To address these challenges, we propose MDCNS (Multi-source Divergence-Consensus for Negative Sampling), a novel ``Teacher-Peer-Self'' framework inspired by Vygotsky’s Zone of Proximal Development (ZPD) theory. The proposed method comprises three components, including multi-source scoring, divergence re-ranking, and consensus distillation. Firstly, multi-source scoring incorporates peer and ensemble teacher models to inject external negative signals and break the self-reinforcement loop. Then, divergence re-ranking exploits prediction discrepancy between self and peer models to enhance sampling diversity. Finally, consensus distillation aligns the self-model with the teacher via KL divergence, simultaneously improving computational cost utilization. Extensive experiments on six real-world datasets and five backbone models show that MDCNS consistently outperforms state-of-the-art negative sampling methods, demonstrating strong effectiveness and generalization.

Modern sequential recommendation systems often rely on negative sampling to efficiently train models over vast item corpora. However, common strategies such as uniform, popularity-based sampling, or hard-negative mining, often yield uninformative negatives, introduce popularity bias, or suffer from false negatives that hinder model learning. While several studies focus on identifying true negatives, none explore latent item representations to mitigate false negatives issue. We propose ‍CATS‍ (Cluster-Aware Thompson Sampling),an adaptive negative sampling method that balances exploration and exploitation by leveraging unsupervised item clustering and adaptive multi-armed bandit principles. The key insight behind CATS is that false negatives tend to lie in the same cluster as the positive item, whereas true hard negatives are more likely found in nearest clusters. CATS leverages this by adaptively sampling negatives from the positive item's cluster, its neighboring clusters, and the rest of the item space, and decaying the sampling probability from the positive item's cluster over time, reducing the risk of accumulating false negatives during training. Evaluation using state-of-the-art sequential models (SASRec, BERT4Rec, BSARec) and three benchmark datasets (MovieLens-1M, Amazon Beauty and BeerAdvocate) demonstrates that CATS consistently outperforms standard and advanced sampling baselines. Notably, CATS achieves improvements of over 20% in NDCG@10 compared to vanilla sampling methods, and of over 16% compared to more advanced methods. Our findings suggest that incorporating the latent item structure through adaptive sampling strategies like CATS can significantly enhance the performance of sequential recommendation systems. The complete code needed to reproduce the results presented in this paper can be found in our GitHub repository https://github.com/gditeodoro/CATS.

Frequency-domain sequential recommendation models enhance sequence representation capacity through spectral transformations. However, existing methods typically adopt coarse-grained spectrum reweighting strategies that strengthen high-frequency components while amplifying random noise within frequency bands. Moreover, they generally rely on a coupled modeling mechanism that handles both long-term and short-term preferences within a single backbone network, lacking dedicated modeling paths tailored to their distinct temporal characteristics. To address these challenges, we propose a Balanced Frequency Decoupling Sequential Recommendation model (BFDRec). Specifically, we design an energy-aware spectrum denoising mechanism to adaptively suppress low-energy noises according to the energy distribution within each frequency band while preserving salient behavioral fluctuation signals. Additionally, we construct a multi-scale decoupled architecture to model users' multi-scale preferences and adaptively integrate them through a dynamic gating mechanism, aligning the sequence modeling process with the distinct temporal characteristics of different frequency bands. Extensive experiments across five real-world datasets demonstrate that BFDRec effectively achieves noise suppression and accurate multi-scale preference modeling, with average improvements of 6.67% and 5.05% in HR and NDCG, respectively, over advanced baseline models. Our code is available at https://anonymous.4open.science/r/BFDRec.

Sub-sequence splitting (SSS) has been demonstrated as an effective approach to mitigate data sparsity in sequential recommendation (SR) by splitting a raw user interaction sequence into multiple sub-sequences. Previous studies have demonstrated its ability to enhance the performance of SR models significantly. However, in this work, we discover that (i). SSS may interfere with the evaluation of the model's actual performance. We observed that many recent state-of-the-art SR models employ SSS during the data reading stage (not mentioned in the papers). When we removed this operation, performance significantly declined, even falling below that of earlier classical SR models. The varying improvements achieved by SSS and different splitting methods across different models prompt us to analyze further when SSS proves effective. We find that (ii). SSS demonstrates strong capabilities only when specific splitting methods, target strategies, and loss functions are used together. Inappropriate combinations may even harm performance. Furthermore, we analyze why sub-sequence splitting yields such remarkable performance gains and find that (iii). it evens out the distribution of training data while increasing the likelihood that different items are targeted. Finally, we provide suggestions for overcoming SSS interference, along with a discussion on data augmentation methods and future directions. We hope this work will prompt the broader community to re-examine the impact of data splitting on SR and promote fairer, more rigorous model evaluation. All analysis code and data are available at https://github.com/KingGugu/SSS4SR.

Accurately modeling the continuous evolution of user preferences remains a fundamental challenge in sequential recommendation. While existing models have made significant strides in capturing sequential dependencies, they predominantly encode historical behaviors via discrete aggregation, often failing to explicitly model the continuous evolution trends and intrinsic momentum underlying preference shifts. To address this limitation, we propose Kinematic Evolution for Sequential Recommendation (KERec), which models user preference evolution from a continuous kinematic perspective. KERec adopts a dual-branch encoding architecture: the global context branch captures stable long-term preference structures, while the kinematic evolution branch explicitly models the first-order velocity and second-order acceleration of preference evolution based on Taylor series expansion and multi-scale difference operators. To effectively integrate these heterogeneous signals, we introduce an adaptive feature fusion mechanism that balances steady-state consistency with dynamic sensitivity. Furthermore, we design a momentum-guided contrastive learning paradigm that constrains augmentation views via evolutionary momentum, suppressing random noise while reinforcing the perception of genuine evolutionary trajectories. Extensive experiments on multiple public benchmark datasets demonstrate that KERec significantly outperforms state-of-the-art methods in recommendation accuracy and exhibits superior robustness under varying sequence lengths and noisy levels. The code of our method is available at https://github.com/yyt-1219/KERec.

Next-basket recommendation (NBR) is a type of recommendation that aims to predict a set of items a user will purchase based on their historical transaction basket sequences. It is governed by a dynamic interplay between two distinct user intents: habitual repurchase, which involves repeating past behaviors, and exploratory interest, which involves discovering new items. However, existing NBR methods generally suffer from two limitations: (1) they often entangle these conflicting motives within a single representation, causing habits to overshadow discovery, and (2) they rely on discrete sequential modeling that ignores continuous-time intervals and item-specific periodicities. In this paper, we propose a novel solution named Time-Interval Disentangled Experts (TIDE) to address these challenges. TIDE incorporates a Hawkes-enhanced Fourier Time Encoding to capture item-specific temporal periodicities and dynamic decay. To decouple user intentions, TIDE utilizes a dual-expert architecture that integrates a Habit Expert for recurring needs and a Pattern-Guided Exploration Expert for discovery. Combined with an item-aware gating mechanism, TIDE adaptively balances repurchase and exploration. Extensive experiments on four diverse real-world datasets demonstrate that TIDE consistently outperforms representative state-of-the-art NBR methods.

Large Language Models (LLMs) have achieved remarkable success in temporal reasoning. However, existing benchmarks predominantly adopt a ''single-answer'' paradigm, focusing on verifying the existence of a specific fact while overlooking the challenge of exhaustiveness. In real-world scenarios, entities often simultaneously play multiple roles or exist in multiple states within the same timeframe. To bridge this gap, we introduce MulTR, a comprehensive benchmark designed for multi-answer temporal reasoning from long unstructured contexts. Specifically, MulTR integrates structured temporal facts from Wikidata and natural language text from Wikipedia MulTR integrates structured temporal facts from Wikidata and natural language text from Wikipedia through a logic-driven synthesis process. Notably, we formulate two distinct settings, question-dependent and document-dependent, based on the presence of cue words in the question. It is designed to systematically decouple temporal reasoning capabilities from the uncertainty of the number of answers. Experiment results demonstrate that state-of-the-art models suffer from retrieval laziness, terminating the search process prematurely after locating the first valid piece of evidence. Consequently, their performance drops sharply when evaluated on strict exact match metrics. MulTR, as a diagnostic testing platform, reveal these defects and establish the rigorous standard for future research in dynamic knowledge processing. The MulTR benchmark and evaluation prompt are publicly available at https://github.com/TemporalNLP/MulTR.

Multi-hop question answering (MHQA) enables accurate answers to complex queries by retrieving and reasoning over evidence dispersed across multiple documents. Existing MHQA approaches mainly rely on iterative retrieval-augmented generation, which suffer from the following two major issues. On one hand, existing methods prematurely commit to surface-level entities rather than underlying reasoning structures, making question decomposition highly vulnerable to lexical ambiguity. On the other hand, existing methods overlook the logical dependencies among reasoning steps, resulting in uncoordinated execution. To address these issues, we propose STRIDE, a framework that separates strategic planning, dynamic control, and grounded execution. At its core, a Meta-Planner first constructs an entity-agnostic reasoning skeleton to capture the abstract logic of the query, thereby deferring entity grounding until after the reasoning structure is established, which mitigates disambiguation errors caused by premature lexical commitment. A Supervisor then orchestrates sub-question execution in a dependency-aware manner, enabling efficient parallelization where possible and sequential coordination when necessary. By dynamically deciding whether to retrieve new evidence or infer from existing facts, it avoids redundant queries and error propagation, while fusing cross-branch information and reformulating failed queries to enhance robustness. Grounded fact extraction and logical inference are delegated to specialized execution modules, ensuring faithfulness through explicit separation of retrieval and reasoning. While STRIDE is compatible with any large language models (LLMs), off-the-shelf open-source LLMs underperform closed-source counterparts in its structured reasoning pipeline. To close this gap, we further propose STRIDE-FT, a modular fine-tuning framework that uses self-generated execution trajectories from STRIDE, requiring neither human annotations nor stronger teacher models. Experiments show that STRIDE achieves robust and accurate reasoning on MHQA benchmarks, while STRIDE-FT effectively enhances open-source LLMs.

Reasoning is fundamental to human intelligence and critical for problem-solving. In practice, answering complex questions often requires abundant reasoning steps involving long reasoning paths. Existing multi-hop reasoning methods mainly focus on short-hop questions with only short reasoning paths, whose distribution is different from that of long paths. While training on long-hop data can improve corresponding performance, collecting it at scale is difficult and expensive. Hence, we explore the short-to-long generalization scenario for the first time, which aims to enhance the model's capability to handle long-hop questions using easily collected short-hop reasoning data. In this scenario, the reasoning model can only learn the features of short-hop questions due to the lack of long-hop data, and thus two challenges emerge. (1) Struggling to determine where to go next : as the number of steps increases, the model needs to reason about the next step based on a long context composed of more evidences. This reasoning process will confuse the reasoning model, which can only learn the pattern from a short context to the next step. (2) Hard to determine when to stop reasoning : too many reasoning hops may introduce excessive noise, while too few may fail to gather sufficient supporting information. The path lengths of long-hop questions vary significantly. Thus, a fixed stopping threshold, commonly used for short-hop questions with similar path lengths, cannot handle them effectively. Inspired by the classic recursive algorithm, we propose a novel Recursion-based Short-to-Long Generalization (RSLG) reasoning framework, which recursively decomposes long-hop questions into multiple short-hop questions that can be handled by a short-hop reasoning model. In this way, the above two challenges are transformed into how to decompose and when to stop decomposition. For how to decompose, we introduce the parallel and sequential recursion patterns to guide the recursive decomposition processes. A local-to-global recursive demonstration construction strategy is proposed to obtain the corresponding recursive demonstrations from the perspective of certainty, complexity, and diversity, respectively. For when to stop decomposition, we propose three recursive termination criteria in view of decomposability, redundancy, and relevance. Extensive experiments on six short-to-long settings across diverse domains demonstrate RSLG's superior performance, robustness, and efficiency in this challenging scenario. The code and data link: https://github.com/NLPGM/RSLG

Analytical information needs, such as trend analysis and causal impact assessment, are prevalent across various domains including law, finance, science, and much more. However, existing information retrieval paradigms, whether based on relevance-oriented document ranking or retrieval-augmented generation (RAG) with large language models (LLMs), often struggle to meet the end-to-end requirements of such tasks at the corpus scale. They either emphasize information finding rather than end-to-end problem solving, or simply treat everything as question answering, offering limited control over reasoning, evidence usage, and verifiability. As a result, they struggle to support analytical queries that have diverse utility concepts and high accountability requirements. In this paper, we propose analytical search as a distinct and emerging search paradigm designed to fulfill these analytical information needs. Analytical search extends search as an evidence-governed, process-oriented analytical workflow that explicitly models analytical intent, retrieves evidence for fusion, and produces verifiable conclusions through structured, multi-step inference. We analyze the limitations of existing search systems in supporting analytical information needs and present a unified example framework that integrates query understanding, recall-oriented retrieval, reasoning-aware fusion, and adaptive verification. We also discuss potential research directions for the construction of analytical search engines. In this way, we highlight the conceptual significance and practical importance of analytical search and call on efforts toward the next generation of search engines that support analytical information needs.

The proliferation of online news poses a challenge to extracting structured timelines from unstructured content. While recent studies have shown that Large Language Models (LLMs) can assist Timeline Summarization (TLS), these approaches primarily treat models as passive generators. The emergence of Large Reasoning Models (LRMs) presents an opportunity to reason over events actively, enabling iterative evidence acquisition, the detection of missing events, and the validation of temporal consistency. To leverage the reasoning capabilities of LRMs, we propose TimelineReasoner, a novel framework that shifts TLS from static generation to an active, reasoning-driven process. TimelineReasoner adopts a two-stage framework: Global Cognition, which tracks events at a macroscopic level and continuously updates a global event memory, and Detail Exploration, which identifies informational gaps and refines the timeline via targeted document retrieval. To support this, TimelineReasoner incorporates several specialized mechanisms, including an Event Scraper for retrieving temporal event descriptions, a Timeline Updater for refining the timeline, and a Supervisor for detecting gaps in the timeline and guiding retrieval. Experimental results on open-domain TLS datasets demonstrate that TimelineReasoner significantly outperforms existing LLM-based TLS methods in terms of timeline accuracy, coverage, and coherence. On closed-domain TLS datasets, our method performs on par with or exceeds state-of-the-art approaches. This work not only pushes the boundaries of TLS but also highlights the broader potential of LRM-based reasoning frameworks for timeline summarization.

Large language models (LLMs) are widely used for question answering (QA) but can generate biased or stereotype-driven answers due to demographic associations learned during pre-training. Existing mitigation strategies often rely on model access or fine-tuning, which limits their applicability to closed-source LLMs. We propose a Causal Front Door Prompting framework (CFDP) that reduces demographic influence by intervening on the chain of thought reasoning, which is treated as an observable mediator. CFDP samples and clusters multiple reasoning traces and estimates answer probabilities through weighted aggregation. Experiments on two widely used bias-sensitive QA benchmarks, BBQ and Stereotype, across major LLMs show that CFDP consistently improves fairness metrics without sacrificing QA accuracy. Ablation and sensitivity analyses confirm the value of each component, indicating that causal intervention on reasoning provides an effective and practical approach for bias mitigation in LLM-based QA.

Large language models (LLMs) excel when user goals are clearly specified, yet real-world queries are often vague and evolving, forcing LLMs to guess and leading to misaligned responses. Existing approaches attempt to clarify user intents through iterative questioning. While effective in alleviating disambiguation, this paradigm tends to merely provide standard and normal responses, which fails to meet the growing demand for diverse personalized expression of users. Based on the observation, we first identify its problem as the ignore of users' mental states, in which the complicated mental elements, unclear functional rules, and evolving mental states pose obstacles to approach the problem. Therefore, in this paper, we introduce the theory of the mentalizing network in the human brain and propose a neurobiologically inspired framework, i.e., NeurPIU, that endows LLMs with human-like mentalizing capabilities for personalized intent understanding. NeurPIU constructs an intent neural network that organizes users' long-term mental states into a three-layer graph; retrieves query-relevant states via spreading activation mechanism with temporal decay; and injects an encoded cognitive prefix into a frozen LLM through a lightweight LoRA-based cognitive model to guide response generation. The network is incrementally updated after each interaction to track evolving user cognition. Extensive experiments on four benchmarks show that NeurPIU consistently improves long-term dialogue quality, especially its plug-and-play feature, and generalizes to conversational recommendation and mental health counseling. A user study with physiological measurements further indicates that NeurPIU reduces interaction time by 53.95% while improving user experience ratings by 22.38%. All data and code are released.

Intent identification serves as the foundation for generating appropriate responses in personalized question answering (PQA). However, existing benchmarks evaluate only response quality or retrieval performance without directly measuring intent identification capabilities. This gap is critical because without understanding which intents users prioritize, systems cannot generate responses satisfying individual information needs. To address this, we introduce the concept of core intents: intents users prioritize when selecting answers to satisfy their information needs. To evaluate these core intents, we propose IPQA, a benchmark for core Intent identification in Personalized Question Answering. Since users do not explicitly state their prioritized intents, we derive core intents from observable behavior patterns in answer selection, grounded in satisficing theory where users choose answers meeting their acceptance thresholds. We construct a dataset with various domains through systematic filtering, LLM-based annotation, and rigorous quality control combining automated verification with human validation. Experimental evaluations across state-of-the-art language models reveal that current systems struggle with core intent identification in personalized contexts. Models fail to identify core intents from user histories, with performance degrading as question complexity increases.

Conversational search systems are typically evaluated using a fixed reference collection of conversations or through user studies with a live system. However, fixed-reference conversations can cover only a few plausible conversations, and user studies are costly, time-consuming, and often hard to reproduce. A promising alternative that avoids coverage and cost issues is user simulation, in which a computer program takes on the role of a user and interacts with the system under evaluation. But the complexity of human search behavior raises the question of how ''realistic'' the simulations actually need to be for reliable evaluations of conversational search systems. In this paper, we ask: Do simulated users need to remember? While real users may learn and forget information during conversational search sessions, which inspired previous research to also model memory capabilities in simulations, it remains unclear whether this actually influences the results of system evaluations. To investigate the impact of memory modeling, we analyze conversations of simulated users and of humans with four conversational search systems. Our results suggest that incorporating long-term memory into simulators can help reproduce system effectiveness rankings obtained from human conversations, whereas incorporating short-term memory can diminish the reproduction. We also find that simulators are generally valid and reproducible---and memory modeling even increases run-to-run reproducibility of system rankings---but overall, simulations approximate human evaluation scores better when ''helpful'' assistants are evaluated than when assistants with deteriorated response quality are assessed. Our code and data are available at https://github.com/webis-de/SIGIR-26.

While large language models (LLMs) have made significant progress in expanding their context windows, they still face great challenges in effectively organizing and utilizing long-term memory to maintain conversation consistency and coherence. Summarizing historical conversations has achieved remarkable performance, which, however, loses conversational trajectory and association, making it difficult to precisely combine memories from different sessions in response to current queries. To address this, we propose the Dynamic Memory Forest (DMF), a novel Consolidation-then-Growth framework for long-term open-domain conversation, which simulates the consolidation and growth processes of human memory by dynamically organizing long-term conversation histories into a memory forest of memory trees. To be specific, inspired by the principles of synaptic consolidation and plasticity from Cognitive Science, we first consolidate each session into memory units that preserve thematic coherence ("Consolidation"). Then, we first structure these units into memory trees and then grow the forest by dynamically connecting them through an evolutionary grafting mechanism, called Group Relative Voting Optimization, which mimics synaptic connection to decide whether a new memory tree should be grafted onto the existing forest or grow independently ("Growth"). For retrieval, we design an Entropy-Driven Memory Walk, constructing a logically coherent memory path via a navigation policy that prioritizes exploring high-entropy nodes. Experiments on three long-term conversation datasets show that our DMF significantly outperforms baselines in enhancing response generation for LLMs.

Despite their common goal of assisting users in information access, current Information Retrieval (IR) systems, such as search engines and recommender systems, are studied and deployed separately across application contexts, resulting in scattered user information and fragmented support for a user’s task. Can we develop a single general intelligent system to unify those specific systems for assisting users with information access using multiple modes (e.g., search, recommendation, and conversation)? In this perspective paper, we present a vision for developing a general intelligent information agent (GIANT) that not only unifies the current specific systems for supporting information access but also goes beyond information access to provide personalized task completion for users. We propose to formalize GIANT generally as an agent interacting with its users to minimize their effort on finishing a task as well as the agent's resource overhead. We propose a probabilistic modeling framework for optimizing GIANT's interactions with its users and discuss how to estimate its four component models, including Situation Model, Content Model, Task Model, and User Model. We discuss how the framework can be refined to derive specific interaction strategies and implemented with a general Markov Decision Process (MDP) architecture. We further discuss how to evaluate GIANT using user simulation and conclude with an outline of some promising directions for future research.

Retrieval-augmented generation (RAG) systems are rapidly becoming a primary entry point to information, replacing ranked lists with synthesized, citation-backed answers. While recent work has focused on improving RAG architectures and measuring answer correctness, much less is known about how users conceptualize these systems and how such mental models reshape information seeking. In this Perspectives Paper, we argue that RAG interfaces compress the visible stages of the Information Search Process: exploration, comparison, and synthesis migrate inside the system, while the user sees only a conversational request-response loop. We ground this argument in a qualitative mixed-methods study of nine participants using an LLM-based ''chat with documents'' interface for exploratory, known-item, list-based, and absent-item tasks. Most participants did not describe the system in terms of selective retrieval and ranking. Instead, they imported an analyst-centric mental model, assuming exhaustive reading, persistent memory, and near-perfect recall. These assumptions systematically reconfigured behavior: participants engaged in minimal query reformulation, delegated sense-making to the system, rarely inspected sources, and treated ''not found'' responses as strong evidence of nonexistence. A minority of participants articulated IR-consistent models and exhibited more iterative, verification-oriented behavior. We use these findings to question implicit behavioral assumptions in interactive IR and RAG evaluation and to outline design implications for handling missing information, retrieval visibility, and verification scaffolding. Our perspective is that understanding users' mental models of both system operation and their own role is central to assessing the epistemic consequences of generative search.

Unsupervised 2D image-based 3D model retrieval (IBMR) aims to retrieve semantically relevant 3D shapes for a given 2D image query when 3D annotations are unavailable. This setting is challenging due to severe modality gaps, category-imbalanced mini-batches, inconsistent cross-domain decision boundaries, and mismatched semantic neighborhood structures. In this paper, we propose a unified framework that integrates Category-Aligned Sampling (CAS), Decision Boundary Alignment (DBA), and Graph Semantic Propagation (GSP) into a single optimization paradigm. CAS constructs category-consistent mini-batches to stabilize crossmodal learning. Built upon CAS, DBA leverages a masked Margin Disparity Discrepancy to regularize cross-domain class decision boundaries via an adversarial min-max objective, encouraging discriminative separation beyond marginal feature matching. To complement boundary-level regularization, GSP builds a crossdomain affinity graph over 2D and 3D samples and propagates supervision-induced relational structure through semantic message passing, explicitly preserving instance-level neighborhood consistency that is critical for retrieval. Extensive experiments on MI3DOR and MI3DOR-2 demonstrate consistent improvements over representative unsupervised IBMR baselines.

Unsupervised cross-domain image retrieval aims to retrieve semantically consistent images across domains with significant domain gaps, which poses substantial challenges under the absence of annotations in both domains. Existing approaches primarily rely on internally derived supervision signals for representation learning and cross-domain alignment. However, such internally induced supervision tends to impose an upper bound on achievable retrieval performance, as it lacks stable semantic references to support reliable category-level correspondence across domains. To address these limitations, we propose Selective Constraint Learning (SCL), a framework that introduces external semantic guidance as a stable prior for unsupervised cross-domain image retrieval. Leveraging a pre-trained foundation model, SCL constructs a dual-scope constraint bank to capture high-confidence positive and negative semantic relations within and across domains. Based on this, we design a generic constraint loss to jointly facilitate intra-domain compactness and inter-domain alignment. In addition, prototypical geometry regularization is designed to enhance in-domain structural stability through prototype-centered pull-and-push forces. Extensive experiments on multiple benchmarks demonstrate that SCL consistently outperforms state-of-the-art methods.

Text-to-3D Scene Retrieval (T3SR) aims to retrieve 3D scenes that match users' linguistic queries, enabling intuitive access to 3D scene repositories. Existing approaches rely on joint embedding learning with large amounts of paired text–scene data, which is expensive to collect and often fails to generalize under open-vocabulary queries and diverse scene distributions. In this paper, we propose Struct-Align, a foundation-model-driven framework for zero-shot T3SR that eliminates the need for paired training data. Our key insight is to reformulate T3SR as a single-modality structural alignment problem by converting both 3D scenes and textual queries into a shared, schema-aligned textual representation compatible with pretrained text embedding models. To reliably derive such representations from complex 3D environments, we introduce a role-decomposed scene structuring pipeline that mitigates generative instability and produces semantically consistent scene depictions. To address the inherent semantic asymmetry between query and scene representations, we further propose a locality-aware structural matching strategy that explicitly localizes query intent and performs instance- and relation-level alignment within query-relevant substructures. Extensive experiments on multiple benchmarks demonstrate that Struct-Align outperforms both training-based and zero-shot baselines while exhibiting strong robustness to domain shift.

Text-based person retrieval (TPR) has become a crucial task in cross-modal retrieval due to its broad application in fields such as public safety and criminal investigation. Existing TPR methods typically rely on fully fine-tuning large-scale pretrained vision-language models like CLIP, which incurs high computational costs and tends to exhibit poor generalization in unseen domains due to overfitting. Fortunately, Parameter-Efficient Transfer Learning (PETL) has emerged as a lightweight alternative. However, applying PETL to TPR remains challenging, as its limited adaptation capacity struggles to capture intricate identity cues and becomes highly susceptible to gradient interference from unreliable image-text pairs. To address these challenges, we present a PETL-based framework named UniGR that unifies granularity and reliability for robust and efficient TPR. Specifically, we design a multi-granularity relational adapter (MRA) to capture both coarse-grained global and fine-grained local relational features among tokens, equipping the generic backbone with the task-specific, precise understanding needed for TPR. To combat the noise sensitivity of PETL, a reliability-aware reweighting strategy (RRS) is introduced to adaptively down-weight unreliable samples during training. Furthermore, we propose a parameter-free cross-modal cyclic verification (CMCV) module to mitigate ambiguities in cross-modal matching computations and refine retrieval ranking further. Experiments on benchmarks corroborate the superiority of UniGR among parameter-efficient methods. Remarkably, with only 4.5% of trainable parameters, UniGR outperforms most fully fine-tuned methods while maintaining strong generalization.

Text-based person search faces inherent limitations due to data scarcity, driven by stringent privacy constraints and the high cost of manual annotation. To mitigate this, existing methods usually rely on a Pretrain-then-Finetune paradigm, where models are first pretrained on synthetic person-caption data to establish cross-modal alignment, followed by fine-tuning on labeled real-world datasets. However, this paradigm lacks practicality in real-world deployment scenarios, where large-scale annotated target-domain data is typically inaccessible. In this work, we propose a new Pretrain-then-Adapt paradigm that eliminates reliance on extensive target-domain supervision through an offline test-time adaptation manner, enabling dynamic model adaptation using only unlabeled test data with minimal post-train time cost. To mitigate overconfidence with false positives of previous entropy-based test-time adaptation, we propose an Uncertainty-Aware Test-Time Adaptation (UATTA) framework, which introduces a bidirectional retrieval disagreement mechanism to estimate uncertainty, i.e., low uncertainty is assigned when an image-text pair ranks highly in both image-to-text and text-to-image retrieval, indicating high alignment; otherwise, high uncertainty is detected. This indicator drives offline test-time model recalibration without labels, effectively mitigating domain shift. We validate UATTA on four benchmarks, i.e., CUHK-PEDES, ICFG-PEDES, RSTPReid, and PAB, showing consistent improvements across both CLIP-based (one-stage) and XVLM-based (two-stage) frameworks. Ablation studies confirm that UATTA outperforms existing offline test-time adaptation strategies, establishing a new benchmark for label-efficient, deployable person search systems. Our code is available at https://github.com/nkuzjh/UATTA.

Continual Text-to-Video Retrieval (CTVR) is a challenging multimodal continual learning setting, where models must incrementally learn new semantic categories while maintaining accurate text-video alignment for previously learned ones, thus making it particularly prone to catastrophic forgetting. A key challenge in CTVR is feature drift, which manifests in two forms: intra-modal feature drift caused by continual learning within each modality, and non-cooperative feature drift across modalities that leads to modality misalignment. To mitigate these issues, we propose StructAlign, a structured cross-modal alignment method for CTVR. First, StructAlign introduces a simplex Equiangular Tight Frame (ETF) geometry as a unified geometric prior to mitigate modality misalignment. Building upon this geometric prior, we design a cross-modal ETF alignment loss that aligns text and video features with category-level ETF prototypes, encouraging the learned representations to form an approximate simplex ETF geometry. In addition, to suppress intra-modal feature drift, we design a Cross-modal Relation Preserving loss, which leverages complementary modalities to preserve cross-modal similarity relations, providing stable relational supervision for feature updates. By jointly addressing non-cooperative feature drift across modalities and intra-modal feature drift, StructAlign effectively alleviates catastrophic forgetting in CTVR. Extensive experiments on benchmark datasets demonstrate that our method shows competitive advantages over state-of-the-art continual retrieval approaches.

Novelty assessment is a critical yet complex task in the examination process for patent acceptance, requiring examiners to determine whether an invention is disclosed in a prior art document. The process involves intricate matching between specific features of a patent claim and passages in the prior art. While prior work has approached novelty prediction primarily as a binary classification task at the claim level, we argue that this formulation is susceptible to spurious correlations and lacks the granularity required for practical application. In this work, we introduce FiNE-Patents (Fine-grained Novelty Examination of Patents), a novel dataset comprising 3,658 first patent claims annotated with fine-grained, feature-level prior art references extracted from European Search Opinion (ESOP) documents. We propose shifting the evaluation paradigm from simple binary classification to a joint retrieval and abstract reasoning task at the feature level, requiring models to identify specific passages from a prior art document that disclose individual claim features, and to identify which features of a claim make it novel. We implement and evaluate LLM-based workflows that decompose claims into features, analyze each feature against prior art, and finally derive a claim-level novelty prediction. Our experiments demonstrate that these workflows outperform embedding-based baselines on passage retrieval and novel feature identification. Furthermore, we show that unlike trained classifiers, LLMs are robust against spurious correlations present in the claim-level novelty classification task. We release the dataset and code to foster further research into transparent and granular patent analysis.

Search and recommendation (S&R) are two integral components of modern online platforms, both aiming to model and satisfy user information needs. This shared objective motivates a unified modeling paradigm that enables richer user modeling and improves the effectiveness of both tasks. Recent attempts to unify S&R formulate item ranking in both tasks as conditional generation. While this paradigm is promising, existing methods rely on full fine-tuning, which is computationally expensive and limits scalability. Parameter-efficient fine-tuning (PEFT) offers a more practical alternative but faces two critical challenges in unifying S&R: (1) gradient conflicts across tasks due to divergent optimization objectives, and (2) shifts in user intent understanding caused by overfitting to fine-tuning data, which distort general-domain knowledge and weaken LLM reasoning. To address these issues, we propose Gradient Multi-Subspace Tuning (GEMS), a novel framework that unifies S&R with LLMs while alleviating gradient conflicts and preserving general-domain knowledge. GEMS introduces (1) Multi-Subspace Decomposition, which disentangles shared and task-specific optimization signals into complementary low-rank subspaces, thereby reducing destructive gradient interference, and (2) Null-Space Projection, which constrains parameter updates to a subspace orthogonal to the general-domain knowledge space, mitigating shifts in user intent understanding. Extensive experiments on benchmark datasets show that GEMS consistently outperforms the state-of-the-art baselines across both search and recommendation tasks, and the gains remain consistent when scaling to billion-parameter LLMs.

Large language model retrievers improve performance on complex queries, but their practical value depends on efficiency, robustness, and reliable confidence signals in addition to accuracy. We reproduce a reasoning-intensive retrieval benchmark (BRIGHT) across 12 tasks and 14 retrievers, and extend evaluation with cold-start indexing cost, query latency distributions and throughput, corpus scaling, robustness to controlled query perturbations, and confidence use (AUROC) for predicting query success. We also quantify reasoning overhead by comparing standard queries to five provided reasoning-augmented variants, measuring accuracy gains relative to added latency. We find that some reasoning-specialized retrievers achieve strong effectiveness while remaining competitive in throughput, whereas several large LLM-based bi-encoders incur substantial latency for modest gains. Reasoning augmentation incurs minimal latency for sub-1B encoders but exhibits diminishing returns for top retrievers and may reduce performance on formal math/code domains. Confidence calibration is consistently weak across model families, indicating that raw retrieval scores are unreliable for downstream routing without additional calibration. We release all code and artifacts for reproducibility.

Generative information retrieval (GenIR) consolidates retrieval into a single neural model that decodes document identifiers (docids) directly from queries. While this model-as-index paradigm offers architectural simplicity, it is poorly suited to dynamic document collections. Unlike modular systems, where indexes are easily updated, GenIR's knowledge is parametrically encoded in its weights; consequently, standard adaptation methods such as full and parameter-efficient fine-tuning can induce catastrophic forgetting. We show that sequential adaptation improves retrieval on newly added documents but substantially degrades performance on earlier slices, exposing a pronounced stability--plasticity trade-off. To address this, we propose post-adaptation memory tuning (PAMT), a memory-only stabilization stage that augments an adapted model with a modular parametric memory head (PMH). PAMT freezes the backbone and attaches a product-key memory with fixed addressing. During prefix-trie constrained decoding, decoder hidden states sparsely query PMH to produce residual corrections in hidden space; these corrections are mapped to score adjustments via the frozen output embedding matrix, computed only over trie-valid tokens. This guides docid generation while keeping routing and backbone parameters fixed. To limit cross-slice interference, PAMT updates only a fixed budget of memory values selected using decoding-time access statistics, prioritizing entries frequently activated by the current slice and rarely used in prior sessions. Experiments on MS~MARCO and Natural Questions under sequential, disjoint corpus increments show that PAMT substantially improves retention on earlier slices with minimal impact on retrieval performance for newly added documents, while modifying only a sparse subset of memory values per session.

Generative retrieval (GR) ranks documents by autoregressively generating document identifiers. Because many GR methods rely on trie-constrained beam search, they are vulnerable to early pruning of relevant prefixes under finite-beam decoding. Planning Ahead in Generative Retrieval (PAG) mitigates this failure mode by using simultaneous decoding to compute a document-level look-ahead prior that guides subsequent sequential decoding. We reproduce PAG at inference time and stress-test its decoding behavior. Using the authors' released checkpoint and identifier/trie artifacts under the reported decoding setup, we reproduce the main effectiveness results on MS~MARCO Dev and TREC-DL 2019/2020, and corroborate the reported beam-size--latency trade-off in our hardware setting. Beyond reproduction, we introduce plan drift diagnostics that quantify how intent-preserving query variations, including misspellings, reordering, synonym substitutions, paraphrases, and naturality shifts, alter the planner's top-n candidate set and highest-weight planner tokens, and how these changes affect guided decoding. We find that PAG's planning signal is brittle under lexical surface-form variation: intent-preserving typos can trigger plan collapse, where the planned candidate pool shifts enough that the look-ahead bonus provides little useful guidance, effectively reverting decoding toward weaker unguided search. We further evaluate fixed-index cross-lingual robustness using non-English mMARCO queries against an English index, and assess query-side mitigation strategies that require no re-indexing; query translation provides the strongest recovery in our setting. Overall, our results confirm PAG's reported effectiveness and the benefit of planning-guided decoding under the released inference setup, while showing that these gains depend on the stability of the planning signal under realistic query variation and query--document mismatch. Code available at https://github.com/kidist-amde/lost-in-decoding.

Document chunking is a critical preprocessing step in dense retrieval systems, yet the design space of chunking strategies remains poorly understood. Recent research has proposed several concurrent approaches, including LLM-guided methods (e.g., DenseX and LumberChunker) and contextualized strategies (e.g., Late Chunking), which generate embeddings before segmentation to preserve contextual information. However, these methods emerged independently and were evaluated on benchmarks with minimal overlap, making direct comparisons difficult. This paper reproduces prior studies in document chunking and presents a systematic framework that unifies existing strategies along two key dimensions: (1) segmentation methods, including structure-based methods (fixed-size, sentence-based, and paragraph-based) as well as semantically-informed and LLM-guided methods; and (2) embedding paradigms, which determine the timing of chunking relative to embedding (pre-embedding chunking vs. contextualized chunking). Our reproduction evaluates these approaches in two distinct retrieval settings established in previous work: in-document retrieval (needle-in-a-haystack) and in-corpus retrieval (the standard information retrieval task). Our comprehensive evaluation reveals that optimal chunking strategies are task-dependent: simple structure-based methods match or outperform LLM-guided alternatives for in-corpus retrieval, while LumberChunker performs best for in-document retrieval. Contextualized chunking improves in-corpus effectiveness but generally degrades in-document retrieval. We also find that chunk size correlates moderately with in-document but weakly with in-corpus effectiveness, suggesting segmentation method differences are not purely driven by chunk size. Our code and evaluation benchmarks are publicly available at https://github.com/ielab/Reproduce-chunking-2026.

Recent advances in Large Language Models (LLMs) have opened new avenues for sequential recommendation by enabling natural language reasoning over user behavior sequences. A common approach formulates recommendation as a language modeling task, where interaction histories are transformed into prompts and user preferences are learned via supervised fine-tuning. However, these methods operate solely in the textual modality and often miss users' fine-grained interests, especially when shaped by rich visual signals such as product images or movie posters. Multimodal Large Language Models (MLLMs) offer a promising alternative by aligning text and vision in a shared semantic space. A prevalent training paradigm applies Supervised Fine-Tuning (SFT) followed by Direct Preference Optimization (DPO) to model user preferences. Yet, two core challenges remain: 1) Imbalanced sample hardness, where random negative sampling causes overfitting on easy examples and under-training on hard ones; 2) Cross-modal semantic bias, where the fixed reference model in DPO prevents the policy model from correcting modality misalignments—especially over long sequences. To address these issues, we propose a Multimodal LLM framework that integrates Hardness-aware and Noise-regularized preference optimization for Recommendation (HaNoRec). Specifically, HaNoRec dynamically adjusts optimization weights based on both the estimated hardness of each training sample and the policy model's real-time responsiveness, prioritizing harder examples. It further introduces Gaussian-perturbed distribution optimization on output logits to enhance cross-modal semantic consistency and reduce modality bias inherited from the reference model. Experiments on three benchmarks demonstrate its superiority over state-of-the-art methods.

Large Language Models (LLMs) demonstrate significant potential in sequential recommendation, and leveraging their Chain-of-Thought (CoT) reasoning capabilities can further unlock profound user preference understanding. However, deploying explicit CoT reasoning in real-world systems faces prohibitive challenges: (i) the conflict between the large model scale required for high-fidelity reasoning and the resource constraints of online services, and (ii) the excessive latency introduced by auto-regressive rationale generation. To address these issues, we propose I Reasoning via Multi-Teacher Distillation (IRMD), a novel framework that 'compiles' the reasoning abilities of large teacher LLMs into a lightweight student Small Language Model (SLM). IRMD first employs a Multi-Teacher CoT Synthesis with Dual-Constraint Rejection Sampling module to generate a high-quality, diverse set of reasoning paths. Subsequently, our Annealing-Scheduled Reasoning Distillation strategy progressively trains the student to internalize this logic, transitioning from mimicking explicit CoT to performing purely implicit reasoning. Extensive experiments on multiple benchmark datasets demonstrate that IRMD significantly outperforms state-of-the-art baselines in both recommendation accuracy and inference efficiency. Our code is accessible at https://github.com/Cxx-0/IRMD.

Traditional sequential recommendation (SR) models learn low-dimensional item ID embeddings from user-item interactions, often overlooking textual information such as item titles or descriptions. Recent advances in Large Language Models (LLMs) have inspired a surge of research that encodes item textual information with high-dimensional semantic embeddings, and designs transformation methods to inject such embeddings into SR models. These embedding transformation strategies can be categorized into two types, both of which exhibits notable drawbacks: 1) adapter-based methods suffer from pronounced dimension collapse, concentrating information into a few dominant dimensions; 2) SVD-based methods are rigid and manual, considering only a few principal spectral components while discarding rich information in the remaining spectrum. To address these limitations, we propose SpecTran, a spectral-aware transformer-based adapter that operates in the spectral domain, attending to the full spectrum to select and aggregates informative components. A learnable spectral-position encoding injects singular-value cues as an inductive bias, guiding attention toward salient spectral components and promoting diversity across embedding dimensions. Across four real-world datasets and three SR backbones, it consistently outperforms strong baselines, achieving an average improvement of 9.17%.

Sequential Recommendation (SR) learns user preferences from their historical interaction sequences and provides personalized suggestions. In real-world scenarios, most items exhibit sparse interactions, known as the tail-item problem. This issue limits the model's ability to accurately capture item transition patterns. To tackle this, large language models (LLMs) offer a promising solution by capturing semantic relationships between items. Despite previous efforts to leverage LLM-derived embeddings for enriching tail items, they still face the following limitations: 1) They struggle to effectively fuse collaborative signals with semantic knowledge, leading to suboptimal item embedding quality. 2) Existing methods overlook the structural inconsistency between the ID and LLM embedding spaces, causing conflicting signals that degrade recommendation accuracy. In this work, we propose a \textbf{F}usion and \textbf{A}lignment \textbf{E}nhancement framework with LLMs for Tail-item Sequential \textbf{Rec}ommendation (\textbf{FAERec}), which improves item representations by generating coherently-fused and structurally consistent embeddings. For the information fusion challenge, we design an adaptive gating mechanism that dynamically fuses ID and LLM embeddings. Then, we propose a dual-level alignment approach to mitigate structural inconsistency. The item-level alignment establishes correspondences between ID and LLM embeddings of the same item through contrastive learning, while the feature-level alignment constrains the correlation patterns between corresponding dimensions across the two embedding spaces. Furthermore, the weights of the two alignments are adjusted by a curriculum learning scheduler to avoid premature optimization of the complex feature-level objective. Extensive experiments across three widely used datasets with multiple representative SR backbones demonstrate the effectiveness and generalizability of our framework. The codes are provided at \url{https://github.com/ZhifuWei/FAERec}.

Large language models (LLMs) have enhanced conventional recommendation models via user profiling, which generates representative textual profiles from users' historical interactions. However, their direct application to session-based recommendation (SBR) remains challenging due to severe session context scarcity and poor scalability. In this paper, we propose SPRINT, a scalable SBR framework that incorporates reliable and informative intents while ensuring high efficiency in both training and inference. SPRINT constrains LLM-based profiling with a global intent pool and validates inferred intents based on recommendation performance to mitigate noise and hallucinations under limited context. To ensure scalability, LLMs are selectively invoked only for uncertain sessions during training, while a lightweight intent predictor generalizes intent prediction to all sessions without LLM dependency at inference time. Experiments on real-world datasets show that SPRINT consistently outperforms state-of-the-art methods while providing more explainable recommendations.

With the rapid development of the Internet, users have increasingly higher expectations for the recommendation accuracy of online content consumption platforms (e.g., short video platforms). However, short videos often contain diverse segments, and users may not hold the same attitude toward all of them (e.g., music enthusiasts may not enjoy all songs in a medley). Traditional binary-classification recommendation models, which treat a video as a single holistic entity, face limitations in accurately capturing such nuanced preferences. Considering that user consumption is a temporal process, this paper demonstrates that the timing of user actions can represent diverse intentions through statistical analysis and examination of action patterns. Based on this insight, we propose a novel modeling paradigm: Action-Aware Gen erative Sequence Network (A2Gen ), which refines user actions (e.g., Like and Follow, etc.) along the temporal dimension and chains them into sequences for unified processing and prediction. First, we introduce the Context-aware Attention Module (CAM) to model action sequences enriched with item-specific contextual features. Building upon this, we develop the Hierarchical Sequence Encoder (HSE) to learn temporal action patterns from users' historical actions. Finally, through leveraging CAM, we design a module for action sequence generation: the Action-seq Autoregressive Generator (AAG). Extensive offline experiments on the Kuaishou's dataset and the Tmall public dataset demonstrate the superiority of our proposed model. Furthermore, through large-scale online A/B testing deployed on Kuaishou's platform, our model achieves significant improvements over baseline methods in multi-task prediction by leveraging sequential information. Specifically, it yields increases of 0.34% in user watch time, 8.1% in interaction rate, and 0.162% in overall user retention (LifeTime-7), leading to successful deployment across all traffic, serving over 400 million users every day.

Deep Research (DR) requires LLM agents to autonomously perform multi-step information seeking, processing, and reasoning to generate comprehensive reports. In contrast to existing studies that mainly focus on unstructured web content, a more challenging DR task should additionally utilize structured knowledge to provide a solid data foundation, facilitate quantitative computation, and lead to in-depth analyses.In this paper, we refer to this novel task as Knowledgeable Deep Research (KDR), which requires DR agents to generate reports with both structured and unstructured knowledge.Furthermore, we propose the Hybrid Knowledge Analysis framework (HKA), a multi-agent architecture that reasons over both kinds of knowledge and integrates the texts, figures, and tables into coherent multimodal reports.The key design is the Structured Knowledge Analyzer, which utilizes both coding and vision-language models to produce figures, tables, and corresponding insights.To support systematic evaluation, we construct KDR-Bench, which covers 9 domains, includes 41 expert-level questions, and incorporates a large number of structured knowledge resources (e.g., 1,252 tables). We further annotate the main conclusions and key points for each question and propose three categories of evaluation metrics including general-purpose, knowledge-centric, and vision-enhanced ones.Experimental results demonstrate that HKA consistently outperforms most existing DR agents on general-purpose and knowledge-centric metrics, and even surpasses the Gemini DR agent on vision-enhanced metrics, highlighting its effectiveness in deep, structure-aware knowledge analysis.Finally, we hope this work can serve as a new foundation for structured knowledge analysis in DR agents and facilitate future multimodal DR studies.

Deep search agents powered by large language models have demonstrated strong capabilities in multi-step retrieval, reasoning, and long-horizon task execution. However, their practical failures often stem from the lack of mechanisms to monitor and regulate reasoning and retrieval states as tasks evolve. Insights from cognitive neuroscience suggest that human metacognition is hierarchically organized, integrating fast anomaly detection with selectively triggered, experience-driven reflection. In this work, we propose Deep Search with Meta-Cognitive Monitoring (DS-MCM), a deep search framework augmented with an explicit hierarchical metacognitive monitoring mechanism. DS-MCM integrates a Fast Consistency Monitor, which performs lightweight checks on the alignment between external evidence and internal reasoning confidence, and a Slow Experience-Driven Monitor, which is selectively activated to guide corrective intervention based on experience memory from historical agent trajectories. By embedding monitoring directly into the reasoning–retrieval loop, DS-MCM determines both when intervention is warranted and how corrective actions should be informed by prior experience. Experiments across multiple deep search benchmarks and backbone models demonstrate that DS-MCM consistently improves performance and robustness.

Recently, Diffusion Large Language Models (dLLMs) have demonstrated unique efficiency advantages, enabled by their inherently parallel decoding mechanism and flexible generation paradigm. Meanwhile, despite the rapid advancement of Search Agents, their practical deployment is constrained by a fundamental limitation, termed as 1) Latency Challenge : the serial execution of multi-round reasoning, tool calling, and tool response waiting under the ReAct agent paradigm induces severe end-to-end latency. Intuitively, dLLMs can leverage their distinctive strengths to optimize the operational efficiency of agents under the ReAct agent paradigm. Practically, existing dLLM backbones face the 2) Agent Ability Challenge. That is, existing dLLMs exhibit remarkably weak reasoning and tool-calling capabilities, preventing these advantages from being effectively realized in practice. In this paper, we propose DLLM-Searcher, an optimization framework for dLLM-based Search Agents. To solve the Agent Ability Challenge, we design a two-stage post-training pipeline encompassing Agentic Supervised Fine-Tuning (Agentic SFT) and Agentic Variance-Reduced Preference Optimization (Agentic VRPO), which enhances the backbone dLLM's information seeking and reasoning capabilities. To mitigate the Latency Challenge, we leverage the flexible generation mechanism of dLLMs and propose a novel agent paradigm termed Parallel-Reasoning and Acting (P-ReAct). P-ReAct guides the model to prioritize decoding tool_call instructions, thereby allowing the model to keep thinking while waiting for the tool's return. Experimental results demonstrate that DLLM-Searcher achieves performance comparable to mainstream LLM-based search agents and P-ReAct delivers approximately 15% inference acceleration. Our code is available at https://github.com/bubble65/DLLM-Searcher

Information retrieval (IR) systems have traditionally been designed and trained for human users, with learning-to-rank methods relying heavily on large-scale human interaction logs such as clicks and dwell time. With the rapid emergence of large language model (LLM) powered search agents, however, retrieval is increasingly consumed by agents rather than human beings, and is embedded as a core component within multi-turn reasoning and action loops. In this setting, retrieval models trained under human-centric assumptions can be mismatched with the way agents issue intermediate queries and consume results. In this work, we argue that retrieval models for agentic search should be trained directly from agent interaction data. We study learning to retrieve from agent trajectories as a trajectory-supervised training setting, where supervision is derived from multi-step agent interactions. Through a systematic analysis of search agent trajectories, we identify key behavioral signals that reveal document utility, including browsing actions, unbrowsed rejections, and post-browse reasoning traces. Guided by these insights, we propose LRAT, a simple yet effective framework that mines high-quality retrieval supervision from agent trajectories and incorporates relevance intensity through weighted optimization. To instantiate this setting at scale, we deploy the Tongyi-DeepResearch-30B model on 10K InfoSeekQA queries with four retrievers, collecting 26,482 agent trajectories and constructing 91,713 training pairs. Extensive experiments on both in-domain and out-of-domain deep research benchmarks demonstrate that retrievers trained with LRAT consistently improve evidence recall, end-to-end task success, and execution efficiency across diverse agent architectures and scales. Our results highlight agent trajectories as a practical and scalable supervision source for retrieval in agentic search.

Large Language Model (LLM)-based search agents have proven promising for addressing knowledge-intensive problems by incorporating information retrieval capabilities. Existing works largely focus on optimizing the reasoning paradigms of search agents, yet the quality of intermediate search queries during reasoning remains overlooked. As a result, the generated queries often remain inaccurate, leading to unexpected retrieval results and ultimately limiting search agents' overall effectiveness. To mitigate this issue, we introduce SmartSearch, a framework built upon two key mechanisms: (1) Process Rewards, which provide fine-grained supervision for the quality of each intermediate search query through Dual-Level Credit Assessment. (2) Query Refinement, which promotes the optimization of query generation by selectively refining low-quality search queries and regenerating subsequent search rounds based on these refinements. To enable the search agent to progressively internalize the ability to improve query quality under the guidance of process rewards, we design a three-stage curriculum learning framework. This framework guides the agent through a progression from imitation, to alignment, and ultimately to generalization. Experimental results show that SmartSearch consistently surpasses existing baselines, and additional quantitative analyses further confirm its significant gains in both search efficiency and query quality. The code is available at https://github.com/RUC-NLPIR/SmartSearch.

Complex information needs in real-world search scenarios demand deep reasoning and knowledge synthesis across diverse sources, which traditional retrieval-augmented generation (RAG) pipelines struggle to address effectively. Current reasoning-based approaches face a key architectural challenge: they employ a single model to handle both high-level planning and detailed execution, resulting in inefficient reasoning and limited scalability. In this paper, we introduce HiRA, a hierarchical framework that separates strategic planning from specialized execution. Our approach decomposes complex search tasks into multiple subtasks, assigns each subtask to a domain-specific agent equipped with external tools and reasoning capabilities, and coordinates the results through a structured integration mechanism. This separation prevents execution details from disrupting high-level reasoning while enabling the system to leverage specialized expertise for different types of information processing. Experiments on four complex, cross-modal deep search benchmarks show that HiRA significantly outperforms state-of-the-art RAG and agent-based systems, highlighting the effectiveness of decoupled planning and execution for multi-step information seeking tasks. The code is available at https://github.com/RUC-NLPIR/HiRA.

Bid shading plays a crucial role in Real-Time Bidding (RTB) by adaptively adjusting the bid to avoid advertisers overspending. Existing mainstream two-stage methods, which first model bid landscapes and then optimize surplus using operations research techniques, are constrained by unimodal assumptions that fail to adapt for non-convex surplus curves and are vulnerable to cascading errors in sequential workflows. Additionally, existing discretization models of continuous values ignore the dependence between discrete intervals, reducing the model's error correction ability, while sample selection bias in bidding scenarios presents further challenges for prediction. To address these issues, this paper introduces Generative Bid Shading (GBS), which comprises two primary components: 1) an end-to-end generative model that utilizes an autoregressive approach to generate shading ratios by stepwise residuals, capturing complex value dependencies without relying on predefined priors; and 2) a reward preference alignment system, which incorporates a channel-aware hierarchical dynamic network (CHNet) as the reward model to extract fine-grained features, along with modules for surplus optimization and exploration utility reward alignment, ultimately optimizing both short-term and long-term surplus using group relative policy optimization (GRPO). Extensive experiments on both offline and online A/B tests validate GBS's effectiveness. Moreover, GBS has been deployed on the Meituan DSP platform, serving billions of bid requests daily.

Automated bidding has become a core component of modern digital advertising. Early methods were primarily rule-based, while easy to implement, they struggled to adapt to rapidly changing environments. Subsequent Reinforcement Learning methods modeled bidding as a Markov Decision Process but were limited in their ability to capture long-term dependencies. While recent generative models have shown encouraging progress, they generally lack explicit mechanisms to balance exploration and safety. They often rely solely on simple action perturbations or trajectory guidance to foster bidding exploration, and critically, they lack a safety fallback mechanism. This limitation leads to inefficient exploration and significantly increases the financial risk for advertising platforms. To bridge this gap, we propose a new framework named Generative Auto-Bidding with Uni fied Mod eling and Exploration (\baby), which synergistically integrates directed exploration with a safe fallback mechanism. \baby utilizes a Decision Transformer (DT) to jointly model historical bidding actions and environmental state transitions. A Q-value module guides the DT's exploration through regularization constraints. Concurrently, an Inverse Dynamics Module (IDM) leverages the future states predicted by the DT to infer robust and behaviorally consistent actions, thereby providing a safe policy fallback. The Q-value module then adaptively selects the final action from these two options, balancing exploration and safety. Together, these three components form an integrated ''explore--safeguard--select'' pipeline, unifying efficiency and safety. We conduct comprehensive experiments on public datasets, in simulated auction environments, and through a large-scale online deployment on Taobao, a leading advertising platform in China. The results demonstrate that \baby consistently outperforms state-of-the-art (SOTA) baseline methods across all scenarios. In real-world online deployment, \baby achieves remarkable improvements: +4.10% in ad GMV, +1.40% in ad clicks, +1.66% in ad cost, and +3.52% in ad ROI, demonstrating its effectiveness and strong industrial applicability.

Auto-bidding and budget pacing are formulated as sequential decision-making tasks. While flexible, such a framework faces a fundamental granularity mismatch: decisions are made at a fine temporal scale, while their performance feedback is fully and reliably observable at a much coarser resolution. This manifests as sparse reward signals and delayed feedback, forcing agents to learn from locally noisy and incomplete signals. We address this core challenge by introducing the \underline{T}rajectory \underline{A}uto-\underline{R}egressive Model (\textbf{TAR}), a generative framework that aligns planning resolution with feedback dynamics. Motivated by the insight that coarser temporal aggregation yields denser rewards and less scattered feedback, TAR generates trajectories in a coarse-to-fine manner. It incorporates three key innovations: (1) progressive trajectory generation across multiple temporal scales; (2) latent-space compression via a multi-scale VQVAE to handle heterogeneous feature types; and (3) state-action integration that captures long-term dependencies without auxiliary inverse models. Comprehensive experiments in both sparse-reward and delayed-feedback settings demonstrate that TAR consistently outperforms strong baselines in offline simulations and online deployment, validating its effectiveness in overcoming the granularity mismatch for more stable and robust advertising optimization.

In marketing, optimizing subsidy allocation to maximize overall profits is of substantial economic importance. Prior research has employed treatment effect estimation techniques to identify subsidy-sensitive items and design corresponding allocation strategies. However, more accurate treatment effect estimations do not necessarily lead to better allocations, underscoring the critical influence of decision boundaries in decision-making. This paper argues that optimal allocation fundamentally depends on predicting the expected optimal subsidy, a challenge distinct from conventional treatment effect estimation or causal decision-making, which existing approaches fail to address. To fill this gap, we introduce a two-stage Counterfactual optimal subsidy Learning method with an Asymmetric reward (CoLA). In the first stage, we derive a coarse estimate of the expected subsidy threshold by exploiting order information and the conditional independence between expected and observed subsidies. In the second stage, we refine these estimates using an asymmetric loss function, leading to more robust predictions. Under practical budget constraints, we prioritize candidates based on their Sharpe ratios to determine the final subsidy allocation strategy. Experiments on three public datasets and an online A/B test show that our method achieves significant performance improvements, yielding the highest total profit and incremental leverage ratios.

Recommender systems have advanced markedly over the past decade by transforming each user/item into a dense embedding vector with deep learning models. At industrial scale, embedding tables constituted by such vectors of all users/items demand a vast amount of parameters and impose heavy compute and memory overhead during training and inference, hindering model deployment under resource constraints. Existing solutions towards embedding compression either suffer from severely compromised recommendation accuracy or incur considerable computational costs. To mitigate these issues, this paper presents \algo, a fast and effective framework for compressing embedding tables. Unlike traditional ID hashing, \algo is built on the idea of exploiting collaborative signals in user-item interactions for user and item groupings, such that similar users/items share the same embeddings in the codebook. Specifically, we formulate a balanced co-clustering objective that maximizes intra-cluster connectivity while enforcing cluster-volume balance, and unify canonical graph clustering techniques into the framework through rigorous theoretical analyses. To produce effective groupings while averting codebook collapse, \algo instantiates this framework with a principled weighting scheme for users and items, an efficient label propagation solver, as well as secondary user clusters. Our extensive experiments comparing \algo against full models and 18 baselines over benchmark datasets demonstrate that \algo cuts embedding parameters by over 75% with a drop of at most 1.85% in recall, while surpassing the strongest baselines by being up to 346× faster.

Serving deep learning based recommendation models (DLRM) at scale is challenging. Existing approaches rely on dedicated ANN indexing and filtering services on CPUs, suffering from non-negligible costs and missing co-design opportunities. Such inefficiency makes them difficult to support complex model architectures, such as learned similarities and multi-task retrieval. In this paper, we present SilverTorch, a model-based serving system that brings all components into one unified model. It unifies model serving by replacing standalone indexing and filtering services with model layers. We propose a model-based GPU Bloom index for feature filtering and a fused Int8 ANN kernel for nearest neighbor search. Through co-design of the ANN search and feature filtering, we reduce GPU memory usage and eliminate computation. Benefiting from this design, we scale up retrieval by introducing an OverArch scoring layer and a multi-task retrieval with a Value Model to aggregate scores. These advancements improve the retrieval accuracy and enable future studies for serving more complex models. Our evaluation on industry-scale datasets shows that SilverTorch achieves up to 23.7× higher throughput compared to the state-of-the-art approaches. We also demonstrate that SilverTorch's solution is 13.35× more cost-efficient than CPU-based solution while improving accuracy via serving more complex models.

Retrieval-augmented code generation commonly conditions a decoder on large retrieved snippets, which couples online cost to repository size and introduces long-context noise. We present Hierarchical Embedding Fusion (HEF), a two-stage repository representation for code completion: (i) an offline cache that compresses repository chunks into a reusable hierarchy of dense vectors using a small fuser model, and (ii) an online interface that maps a small number of retrieved vectors into learned pseudo-tokens consumed by a code generator. This replaces thousands of retrieved tokens with a fixed pseudo-token budget while retaining access to repository-level information. On RepoBench and RepoEval, HEF with a 1.8B pipeline attains comparable exact-match accuracy to snippet-based retrieval baselines while operating at sub-second median latency on a single A100. Compared to graph- and iterative-retrieval systems in our setup, HEF reduces median end-to-end latency by 13×–26×. We additionally introduce a utility-weighted likelihood signal to filter training contexts and report ablations over pseudo-token budget, embedding models, and robustness to harmful retrieval. These results support hierarchical dense caching as an effective mechanism for low-latency repository-aware code completion.

Recently, Retrieval Augmented Generation (RAG) has shifted focus to multi-retrieval approaches to tackle complex tasks such as multi-hop question answering. However, these systems struggle to decide when to stop searching once enough information has been gathered. To address this, \citet{zhou2024metacognitive} introduced Metacognitive Retrieval Augmented Generation (MetaRAG), a framework inspired by metacognition that enables Large Language Models to critique and refine their reasoning. In this reproducibility paper, we reproduce MetaRAG following its original experimental setup and extend it in two directions: (i) by evaluating the effect of PointWise and ListWise rerankers, and (ii) by comparing with SIM-RAG, which employs a lightweight critic model to stop retrieval. Our results confirm MetaRAG's relative improvements over standard RAG and reasoning-based baselines, but also reveal lower absolute scores than reported, reflecting challenges with closed-source LLM updates, missing implementation details, and unreleased prompts. We show that MetaRAG is partially reproduced, gains substantially from reranking, and is more robust than SIM-RAG when extended with additional retrieval features.

Retrieval-augmented generation (RAG) has demonstrated its ability to enhance Large Language Models (LLMs) by integrating external knowledge sources. However, multi-hop questions, which require the identification of multiple knowledge targets to form a synthesized answer, raise new challenges for RAG systems. Under the multi-hop settings, existing methods often struggle to fully understand the questions with complex semantic structures and are susceptible to irrelevant noise during the retrieval of multiple information targets. To address these limitations, we propose a novel graph representation learning framework for multi-hop question retrieval. We first introduce a Multi-information Level Knowledge Graph (Multi-L KG) to model various information levels for a more comprehensive understanding of multi-hop questions. Based on this, we design a Question-Adaptive Graph Neural Network (Quest-GNN) for representation learning on the Multi-L KG. Quest-GNN employs intra/inter-level message passing mechanisms, and in each message passing the information aggregation is guided by the question, which not only facilitates multi-granular information aggregation but also significantly reduces the impact of noise. To enhance its ability to learn robust representations, we further propose two synthesized data generation strategies for pre-training the Quest-GNN. Extensive experimental results demonstrate the effectiveness of our framework in multi-hop scenarios, especially in high-hop questions the improvement can reach 33.8%. The code is available at: https://github.com/Jerry2398/QSGNN.

Retrieval-augmented reasoning (RAR) is a recent evolution of retrieval-augmented generation (RAG) that employs multiple reasoning steps for retrieval and generation. While effective for some complex queries, RAR remains vulnerable to errors and misleading outputs. Uncertainty quantification (UQ) offers methods to estimate the confidence of systems' outputs. These methods, however, often handle simple queries with no retrieval or single-step retrieval, without properly handling RAR setup. Accurate estimation of UQ for RAR requires accounting for all sources of uncertainty, including those arising from retrieval and generation. In this paper, we account for these sources and introduce Retrieval-Augmented Reasoning Consistency (R2C), a novel UQ method for RAR. The core idea of R2C is to perturb the multi-step reasoning process by applying various actions to reasoning steps. These perturbations alter the retriever's input, which shifts its output and consequently modifies the generator's input at the next step. Through this iterative feedback loop, the retriever and generator continuously reshape each other's inputs, enabling us to capture uncertainty arising from both components. Experiments on five popular RAR systems across diverse QA datasets show that R2C improves AUROC by over 5% on average compared to the state-of-the-art UQ baselines. Extrinsic evaluations using R2C as an external signal further confirm its effectiveness for two downstream tasks: in the Abstention task, it achieves ~5% gains in both F1Abstain and AccAbstain; in Model Selection, it improves exact match by ~7% over single models and ~3% over selection methods. Code is available on https://github.com/HeydarSoudani/R2C.

Long-form Retrieval-Augmented Generation (RAG) brings the challenge of coverage-based ranking, because ranking methods must ensure the inclusion of comprehensive relevant nuggets (i.e., facts), which can thereby be synthesized into a comprehensive output. In this work, we propose CoveR, a dense retrieval method optimized for coverage-aware retrieval scenarios. CoveR is a bi-encoder trained with the coverage-based contrastive and distillation objectives, which enables CoveR to capture diverse aspects of information needs. To train CoveR, we create the SCOPE dataset, which comprises 90K training pairs from Researchy Questions with synthetic coverage signals augmented from sub-question answerability judgments generated by LLMs. Our empirical experiments show that CoveR enhances nugget coverage by 10% over strong dense retrieval baselines without sacrificing its relevance-based retrieval capability. Further ablation studies validate the importance of our proposed learning method, showing that CoveR achieves a superior trade-off between relevance- and coverage-based ranking, which is essential for long-form RAG.

Learning ranking models that are both explainable and effective is an emerging topic within the research area of explainable AI. Several Learning-to-Rank (LtR) algorithms have been recently proposed that build models that are simple to explain and, at the same time, almost as effective as their state-of-the-art, black-box counterparts. In this work, we propose ILMART (Interpretable LambdaMART), a novel framework with different strategies to constrain the state-of-the-art LtR LambdaMART algorithm to generate interpretable models, i.e., ensembles whose trees can use either single features (main effects) or a limited number of interacting features (interaction effects). ILMART facilitates a straightforward trade-off between model explainability and effectiveness by precisely tuning the quantity of main and interaction effects during the learning phase. We show that slightly increasing their number allows ILMART models to reach ranking performances at par with full-complexity LambdaMART ones. Furthermore, reproducible experiments conducted on publicly available LtR datasets demonstrate that ILMART can improve nDCG@10 by up to 10% compared to state-of-the-art competitors while preserving an explainable structure. Finally, we explore the relationship between model explainability and inference efficiency by introducing a novel and easy-to-implement scoring algorithm for ILMART ranking models, achieving up to a 100× speedup compared to the baseline.

Learning-to-rank (LTR) is a fundamental component of modern large-scale information retrieval (IR) systems, playing an essential role across various stages of the ranking pipeline. Recently, differentiable sorting networks have attracted increasing attention for LTR as a permutation-level learning paradigm, enabling end-to-end optimization directly on ranking structure. However, existing approaches suffer from two critical limitations: (i) permutation-matrix fidelity, i.e., the predicted soft permutation matrix may deviate from the exact hard permutation matrix required by permutation-level objectives; and (ii) uncertainty in target ordering arising from coarse or tied relevance labels, where the ground-truth order is set-valued rather than unique. To address the first issue, we propose an Order-Preserving Sorting network (OPS), which combines a relaxation operator with random hard swaps to explicitly control the approximation error between the predicted soft permutation matrix and the exact hard permutation matrix. To handle target order uncertainty, we further introduce a novel Variational Order Uncertainty (VOU) loss that enables effective optimization under incomplete or partially ordered supervision. We evaluate our approach through comprehensive experiments on both public benchmarks and real-world systems. On public datasets, including a synthetic four-digit sorting task based on MNIST and LTR benchmarks such as WEB30K and Istella, our method consistently outperforms state-of-the-art baselines. Moreover, we deploy OPS in the advertising pre-ranking stage of a large-scale industrial retrieval system, where it achieves a statistically significant 1.74% revenue uplift in online A/B tests and is currently serving full production traffic. These results demonstrate the effectiveness and robustness of our approach across both academic benchmarks and real-world industrial scenarios. The experimental code is available at https://github.com/lx1563371/OPS_SIGIR2026.

Modern learning-to-rank methods often rely on listwise objectives that directly model and optimize relative document order over entire permutations. While these objectives improve ranking quality, they frequently produce models with highly inflated relevance scores whose magnitudes exceed what is necessary for meaningful document separation, leading to poor probabilistic calibration. In this work, we analyze this phenomenon in the context of ListMLE [34]. We show that score inflation is a structural consequence of the objective's overlapping softmax normalizers and shift invariance, which generate persistent gradient contributions even after the correct ordering is achieved. As a result, absolute logits drift to increasingly large values, producing saturated and overconfident predictions. We analyze this inflation at the gradient level and identify two complementary mechanisms to address it. Training-time temperature scaling bounds the effective logit scale allowed by shift invariance, while position-dependent discounting suppresses asymmetric tail gradients induced by overlapping suffix normalizers. Together, these modifications prevent persistent score inflation and yield bounded and well-separated scores. We evaluate our approach in a text-based retrieval setting using a state-of-the-art embedding model, and demonstrate substantial improvements in performance under calibration metrics alongside competitive ranking performance on the BEIR benchmark[30].

Query-Specific Document and Entity Representations (QDER) has shown state-of-the-art effectiveness on multiple information retrieval benchmarks. This work attempts to reproduce this promising approach. In an initial attempt, we were unable to find results that resemble the convincing effects measured in the original work. A deep dive into the published artifacts, plus a comparative study of similar work by the same authors show inconsistencies between the descriptions in the paper and their implementation in the published artifacts, and a leakage of relevance assessments due to oversampling entities that are known to appear in relevant documents only. Due to dependencies in the ranking pipeline upon relevance assessments, the final ranking models gain knowledge about the relevance of entities. We conclude that knowing the relevance of entities for an information need can be very valuable, but that the methods described are insufficient to determine this relevance information from the documents and queries alone. The resulting pipeline cannot be deployed effectively in practice, limiting the impact of the research results.

Two-tower models are a popular unbiased learning-to-rank (ULTR) approach for mitigating position bias in clicks. A major challenge in two-tower models is identifiability: whether relevance and position bias can be uniquely determined from clicks. Recent work shows that two-tower models can be identified when the same documents (or documents with similar features) are observed across positions. However, these conditions are defined in infinite data. In practice, we lack methods for diagnosing identifiability in real-world datasets with small sample sizes and limited positional variability. In this work, we present a practical identifiability diagnostic for two-tower models. We quantify whether individual bias parameters are uniquely determined by shifting them away from their optimal values, retraining parts of the model, and measuring whether click-prediction performance degrades significantly. Identified parameters exhibit measurable performance loss when shifted, while unidentified parameters can be freely adjusted without affecting model fit. Our method can distinguish between cases in which identifiability fails due to limitations in model assumptions or data collection and cases caused by a limited sample size. We demonstrate, through simulation, how non-deterministic logging policies, feature overlap, and increased sample size affect identifiability. Applying our method to the real-world Baidu-ULTR dataset, we find that despite large amounts of click data, some parameters of two-tower models remain unidentified, highlighting the practical need for diagnostic methods. All code, data, and results are available at: https://github.com/stanfris/practical-identifiability-ultr

Although attention-based sequential recommenders have achieved strong accuracy, they can exhibit a systematic failure mode, i.e., lost-in-the-middle, where attention disproportionately concentrates on sequence boundaries to informative middle behaviors. Through controlled perturbations, we find that this U-shaped bias persists even when collaborative signals are moved, indicating an architectural effect rather than a positional-encoding artifact. To mitigate this bias, we propose Disentangling and Calibrating Position Bias framework, which is a lightweight, plug-and-play calibration procedure that combines bias-adaptive masking during training with counterfactual logit calibration at inference to remove residual position-only effects. Experiments on diverse datasets demonstrate consistent improvements in ranking quality. Further ablations verify the contributions of both masking and calibration components.

Rayalaseema in southern India is one of the country's most drought-vulnerable regions, experiencing drought in 15 of 18 years between 2000-2018. Over 80% of its 2.7 million farmers are marginal or small landholders practicing rainfed agriculture. Despite pressing agricultural information needs around water management, crop selection, and drought adaptation, these farmers face compounding barriers: 36% rural internet penetration, 60% literacy rates, and information systems designed for English or Hindi speakers. We examine the current landscape of agricultural information retrieval resources for Telugu-speaking farmers and find a critical gap: while agricultural QA datasets exist for Hindi, Tamil, and Bengali, no Telugu agricultural QA benchmark exists, let alone one addressing drought-specific contexts. We document available information sources, identify where they fall short, and propose a methodology for constructing a Telugu Agricultural QA dataset using Kisan Call Center records and regional extension materials. This work charts a path toward inclusive agricultural IR for over 74 million Telugu speakers.

Despite being spoken by over 26 million people, Nepali lacks a standardized Information Retrieval (IR) benchmark, preventing systematic evaluation of retrieval models in this low-resource setting. While progress has been made in several Nepali NLP tasks, no publicly available test collection with queries and relevance judgments exists for ad-hoc retrieval task. We propose the creation of a first standardized Nepali IR benchmark supporting monolingual (Nepali queries \rightarrow Nepali documents), cross-lingual (English queries \rightarrow Nepali documents), and code-mixed (Nepali-English queries \rightarrow Nepali documents) retrieval. We argue that Nepali provides a compelling testbed for studying morphology-aware retrieval, code-mixed query processing, and multilingual embedding transfer.

Multi-document summarization for low-resource languages faces a critical trade-off: large language models are computationally prohibitive for most institutions, while smaller models suffer from severe hallucination in abstractive generation. We address this through extractive sentence retrieval, which guarantees faithfulness while operating within constrained computational budgets. Our approach combines language-adaptive mixture-of-experts embeddings with graph neural networks that model discourse structure, addressing linguistic challenges across typologically diverse low-resource languages. With only 3.2M trainable parameters, our model requires 28 times less training time than comparable transformer-based approaches, making it practical for single-GPU environments typical in resource-limited settings. We demonstrate applicability to some main Southeast Asia (SEA) countries including Vietnam, Thailand, Laos, Indonesia, and Malaysia, representing three distinct language families: Austroasiatic, Kra-Dai, and Austronesian.

Case-to-Statute Retrieval—identifying applicable statutes from case facts—is essential for judicial efficiency but remains underexplored in low-resource languages like Vietnamese, where annotated legal corpora are scarce and domain-specific challenges persist. To advance research in this setting, we introduce ViCSR, a new benchmark of 10,000 Vietnamese criminal judgments and 1,122 statutory articles with citation-based relevance labels. We further propose a lightweight two-stage framework that performs efficient candidate retrieval with a fine-tuned Vietnamese bi-encoder and leverages citation structure for reranking with a compact heterogeneous GNN. Experiments show clear improvements over strong sparse and dense baselines, highlighting the importance of both domain adaptation and structure-aware modeling in low-resource legal IR. We will publicly release the benchmark, code, and checkpoints.

Recent advances in Large Language Models (LLMs) have significantly improved Text-to-SQL performance in high-resource languages. However, their effectiveness in low-resource language settings remains largely underexplored. In this work, we investigate the adaptation of LLMs for Text-to-SQL generation in Chichewa, a low-resource Bantu language spoken by over 12 million people in Malawi and neighboring regions. We construct a structured Chichewa Text-to-SQL benchmark consisting of 400 manually curated natural language–SQL pairs grounded in a unified relational database covering agriculture, commodity prices, population statistics, market data, and food insecurity. We systematically evaluate five open-source LLMs under zero-shot, random 5-shot, and retrieval-augmented 5-shot prompting, in both English and Chichewa. We then apply parameter-efficient fine-tuning (QLoRA) to selected models and, crucially, evaluate the combined effect of QLoRA fine-tuning with retrieval-augmented prompting. QLoRA alone improves English execution accuracy to 78.3\% and Chichewa execution accuracy to 41.7\%. When combined with retrieval-augmented prompting, QLoRA achieves 53.3\% execution accuracy in Chichewa, representing the best reported result for this language on this benchmark and narrowing the English–Chichewa gap to 23.4 percentage points. Our findings offer practical guidance for deploying database interfaces in linguistically underserved environments.

With recent advancements, Small language models (SLMs) are increasingly used as preprocessors to handle query classification, routing, and candidate selection in retrieval pipelines, but they are nearly always prompted in English, even when users search in Hindi, Bengali, or code-mixed forms. We test whether prompting the same (frozen) SLM in three typologically diverse languages and aggregating the outputs can improve classification without retraining or translation. Nine decoder-only models (1B--9B parameters) evaluated on four public benchmarks show that confidence-weighted fusion of English, Hindi, and Bengali predictions raises macro-F1 by 3--5 points over English-only baselines, with the strongest gains on binary and coarse intent tasks. Parallel execution keeps latency within 1.2--1.4× of the single-language baseline. A paraphrase-only ensemble under identical conditions reaches only +1.4~F1 on average, suggesting that cross-lingual diversity rather than surface-level input variation drives the gain. Because no additional data, training, or translation services are required, our method may be useful when scaling to larger models is out of reach.

Sequential recommendation requires capturing diverse user behaviors, which a single network often fails to capture. While ensemble methods mitigate this, training multiple networks from scratch incurs high computational cost and instability from noisy mutual supervision. We propose Frozen and Learnable networks with Aligned Modular Ensemble (FLAME), a novel framework that condenses ensemble-level diversity into a single network for efficient sequential recommendation. During training, FLAME simulates exponential diversity using only two networks via modular ensemble, which dynamically combines sub-modules (e.g., layers) of each network to generate a rich space of diverse representation patterns. To stabilize training, FLAME pretrains and freezes one network as a semantic anchor and employs guided mutual learning to align diverse representations into the space of remaining learnable network. At inference, FLAME utilizes only the learnable network, achieving ensemble-level performance with zero overhead compared to a single network. Experiments on six datasets show that FLAME outperforms state-of-the-art baselines, achieving up to 7.69× faster convergence and 9.70% improvement in NDCG@20. Our code is available at https://github.com/woo-joo/FLAME_SIGIR26.

Cross-domain sequential recommendation (CDSR) alleviates interaction sparsity by jointly modeling user behaviors across multiple domains. While current studies have made some progresses, they still neglect two issues that severely impact recommendation performance: (i) ignoring domain-specific interaction frequencies and interest decay rates at identical time intervals; (ii) treating semantic preferences as time-invariant during cross-domain transfer. To address these, we propose a novel framework that bridges Behavior and Semantics for Time-aware Cross-Domain Sequential Recommendation (BST-CDSR ). Specifically, we design a behavioral preference evolution module that decouples long-term interests and short-term intentions, and models continuous-time preference via a neural ordinary differential equation (ODE) with event-driven updates. Additionally, to capture time-aware semantic preferences, we introduce a temporal counterfactual-enhanced semantic generator that discretizes temporal interval tokens and leverages large language models (LLMs) to extract robust temporal semantics, where counterfactual perturbations enhance the time sensitivity of semantic preferences. Furthermore, we propose a time-preference guided domain transfer module to adaptively control transfer weights and mitigate negative transfer. Extensive experiments on real-world datasets demonstrate that BST-CDSR consistently outperforms baselines.

Cross-domain Recommendation (CDR) exploits multi-domain correlations to alleviate data sparsity. As a core task within this field, inter-domain recommendation focuses on predicting preferences for users who interact in a source domain but lack behavioral records in a target domain. Existing approaches predominantly rely on overlapping users as anchors for knowledge transfer. In real-world scenarios, overlapping users are often scarce, leaving the vast majority of users with only single-domain interactions. For these users, the absence of explicit alignment signals makes fine-grained preference transfer intrinsically difficult. To address this challenge, this paper proposes Language-Guided Conditional Diffusion for CDR (LGCD), a novel framework that integrates Large Language Models (LLMs) and diffusion models for inter-domain sequential recommendation. Specifically, we leverage LLM reasoning to bridge the domain gap by inferring potential target preferences for single-domain users and mapping them to real items, thereby constructing pseudo-overlapping data. We distinguish between real and pseudo-interaction pathways and introduce additional supervision constraints to mitigate the semantic noise brought by pseudo-interaction. Furthermore, we design a conditional diffusion architecture to precisely guide the generation of target user representations based on source-domain patterns. Extensive experiments demonstrate that LGCD significantly outperforms state-of-the-art methods in inter-domain recommendation tasks.

Cross-domain Sequential Recommendation (CDSR) has been proposed to enrich user-item interactions by incorporating information from various domains. Despite current progress, the domain imbalance issue and domain transition issue hinder further development of CDSR. The former presents a phenomenon where interactions in one domain dominate the entire behavior, leading to difficulty in capturing domain-specific features in the other domain. The latter points to the difficulty in capturing users' cross-domain preferences within the mixed interaction sequence, resulting in poor next-item prediction performance for specific domains. With world knowledge and powerful reasoning abilities, Large Language Models (LLMs) partially alleviate the above issues by functioning as both a generator and an encoder. However, current LLMs-enhanced CDSR methods are still under exploration, which fail to recognize the irrelevant noise and rough profiling problems. Thus, to address the aforementioned challenges, we propose an LLMs Enhanced Cross-domain Sequential Recommendation with Dual-phase Training (LLM-EDT). To address the domain imbalance issue while minimizing irrelevant noise, we propose the transferable item augmenter to adaptively generate possible cross-domain behaviors for users. Then, to alleviate the domain transition issue, we introduce a dual-phase training strategy to empower the domain-specific thread with a domain-shared background. As for the rough profiling problem, we devise a domain-aware profiling module to summarize the user's preference in each domain and adaptively aggregate them to generate comprehensive user profiles. The experiments on three public datasets validate the effectiveness of our proposed LLM-EDT. To ease reproducibility, we have released the detailed code online. https://github.com/Applied-Machine-Learning-Lab/SIGIR26_LLM-EDT.

Cross-domain recommendation is a well-known technique for improving recommendations in a target domain, especially under sparse data or cold-start conditions. A common strategy is to train user embeddings separately in the source and target domains and learn a transfer function between them. In contrast, we propose SetCDR, which constructs more effective user representations in the target domain by directly incorporating each user’s source and target history. We additionally introduce a lightweight domain indicator that preserves data–domain relational information. These histories, composed of item–rating pairs, are represented as variable-length sets and processed using a permutation-invariant neural architecture. This differs from conventional neural networks, which do not naturally handle unordered inputs, and is well suited to recommender systems where user history sizes vary greatly among users. The use of a permutation-invariant architecture ensures consistent embeddings regardless of input order, improving robustness to real-world variability and training efficiency. We demonstrate SetCDR in two forms: a simple sum-pooling method and an extended multihead attention–based method that captures more complex dependencies within user histories. Moreover, because SetCDR operates directly on sets of user history records, it provides a natural way to examine how histories influence user representations. Finally, SetCDR adapts immediately to new interactions without additional retraining, enabling on-the-fly performance improvement. Experimental results across multiple cross-domain benchmarks confirm that SetCDR consistently outperforms strong baselines in recommendation quality.

Many internet companies operate multiple flagship applications, each of which can be regarded as a distinct business domain, covering areas such as video, reading, and gaming. Within each domain, diverse recommendation scenarios coexist, and users engage in various tasks with heterogeneous behaviors. In our industrial setting, we observe three key phenomena that existing methods rarely address: (i) users' Cross-Domain Interests (CDI) are weakly exploited, since behavior sequences are often pooled for efficiency, losing transferable cross-domain dependencies; (ii) multi-domain, scenario, and task variations are under-modeled, making it difficult to capture fine-grained complementarities; and (iii) multimodal features remain misaligned with ID features, especially when different domains emphasize different modalities. These gaps hinder cross-product collaboration. We propose STAR-CDR, a Scenario- and Task-Aware Cross-Domain Recommendation architecture. Following the modular decomposition of industrial recommenders, STAR-CDR introduces three innovations: (i) a CDI Extractor that enhances sequential modeling with domain indicators and other key features; (ii) a CDI Injector that recalibrates activations and enables domain-/scenario-/task-aware expert routing; and (iii) a CDI Multimodal Adapter that aligns image, text, and ID features under CDI-guided gating. Compared to the strongest baseline in our setting, STAR-CDR improves offline AUC 3.5% across five domains and yields +3.17% to +9.24% relative KPI gains in online A/B tests. The system now supports multiple domains (flagship applications), scenarios (recommendation scenarios), and tasks (user behaviors) at scale, serving over 280 million daily active users. We provide a GitHub repository. https://github.com/Rescomk/STAR-CDR with STAR-CDR's code, training scripts, and a synthetic data construction pipeline.

Multimodal data plays a critical role in web-based recommendation systems, where information from diverse modalities such as vision and text enhances representation learning. However, real-world multimodal datasets often suffer from modality incompleteness due to sensor failures, annotation scarcity, or privacy constraints, which substantially degrade model performance and reliability. One effective solution to address this issue is modality completion, which reconstructs missing features to provide modality-complete graphs for downstream tasks. Given a query node with missing multimodal features, existing modality completion methods typically infer information from the node itself or its neighbors to reconstruct the missing modality. However, these methods may overlook semantically relevant context in the graph, which contains valuable cues that are non-trivial to capture through simple methods like neighborhood aggregation. In this work, we propose GRE-MC, a Graph Retrieval–Enhanced Modality Completion framework, to overcome these limitations. By introducing a modality-aware subgraph retrieval mechanism, GRE-MC selects semantically relevant subgraphs from the entire graph, providing richer contextual information for completing missing modalities. Subsequently, a graph transformer jointly encodes the query node and the retrieved subgraph via global attention to complete the missing features, while a learnable sparse-routing codebook regularizes latent embeddings into compact bases for improved robustness. Extensive experiments on multimodal recommendation benchmarks demonstrate that GRE-MC consistently outperforms state-of-the-art methods, validating the effectiveness of subgraph retrieval and joint-encoding graph transformer for robust modality completion.

The emergence of generative models enables the creation of texts and images tailored to users' preferences. Existing personalized generative models have two critical limitations: lacking a dedicated paradigm for accurate preference modeling, and generating unimodal content despite real-world multimodal-driven user interactions. Therefore, we propose personalized multimodal generation, which captures modal-specific preferences via a dedicated preference model from multimodal interactions, and then feeds them into downstream generators for personalized multimodal content. However, this task presents two challenges: (1) Gap between continuous preferences from dedicated modeling and discrete token inputs intrinsic to generator architectures; (2) Potential inconsistency between generated images and texts. To tackle these, we present a two-stage framework called Discrete Preference learning for Personalized Multimodal Generation (DPPMG). In the first stage, to accurately learn discrete modal-specific preferences, we introduce a modal-specific graph neural network (a dedicated preference model) to learn users' modal-specific preferences, which preferences are then quantized into discrete preference tokens. In the second stage, the discrete modal-specific preference tokens are injected into downstream text and image generators. To further enhance cross-modal consistency while preserving personalization, we design a cross-modal consistent and personalized reward to fine-tune token-associated parameters. Extensive experiments on two real-world datasets demonstrate the effectiveness of our model in generating personalized and consistent multimodal content.

Multimodal embedding models, rooted in multimodal large language models (MLLMs), have yielded significant performance improvements across diverse tasks such as retrieval and classification. However, most existing approaches rely heavily on large-scale contrastive learning and offer limited exploration of how the architectural and training paradigms of MLLMs affect embedding quality. While effective for generation, the causal attention and next-token prediction paradigm of MLLMs does not explicitly encourage the formation of globally compact representations, limiting their effectiveness as multimodal embedding backbones. To address this, we propose CoCoA, a Content reconstruction pre-training paradigm based on Collaborative Attention for universal multimodal representation learning. Specifically, we restructure the attention flow and introduce an EOS-based reconstruction task, encouraging the model to reconstruct input from the corresponding (EOS) embeddings. This drives the multimodal model to compress the semantic information of the input into the (EOS) token, laying the foundations for subsequent contrastive learning. Extensive experiments on MMEB-V1 demonstrate that CoCoA built upon Qwen2-VL and Qwen2.5-VL significantly improves embedding quality. Results validate that content reconstruction serves as an effective strategy to maximize the value of existing data, enabling multimodal embedding models to generate compact and informative representations, raising their performance ceiling. Our project is available at https://github.com/Trustworthy-Information-Access/CoCoA.

Multimodal alignment is critical for bridging the semantic gap in information retrieval. However, traditional pairwise strategies introduce a geometric blind spot: while they align anchor modalities (e.g., text) with others, they lack constraints to enforce mutual consistency between peripheral modalities (e.g., video and audio). The TRIANGLE framework addresses this by minimizing the area of modality triplets on a hypersphere to enforce holistic alignment. In this reproducibility study, we verify the robustness of this geometric objective for retrieval tasks. We confirm that TRIANGLE outperforms pairwise baselines in zero-shot settings, achieving Recall@1 gains of up to +8.7 points, though benefits are domain-dependent. However, we fail to reproduce the reported learning-from-scratch results. Analysis using a synthetic toy dataset attributes this to instability when jointly optimizing geometric alignment with Data-Text Matching (DTM) loss. Furthermore, we find that cosine regularization primarily stabilizes text-to-video retrieval, and fine-tuning with domain supervision amplifies geometric benefits but reduces cross-dataset generalization. Our findings support the efficacy of geometric alignment while highlighting critical optimization sensitivities.

Multimodal knowledge editing (MKE) aims to efficiently rectify outdated or incorrect knowledge in multimodal large language models (MLLMs) while preserving reliability, generality, and locality. Recently, in-context editing (ICE) has emerged as a prevalent paradigm for MKE, focusing on inference-time context manipulation. While ICE mitigates the side effects of intrinsic interventions on MLLMs, it still suffers from an out-of-control limitation. Specifically, previous methods rely excessively on the implicit contextualization of MLLMs and regard the MKE process as a matter of chance. Based on this observation, we further identify the corresponding challenges as localizing the responsible key units and defining the triggering conditions within MLLMs. To address these challenges, we pursue a controllable ICE approach and refer to the multimodal neurons. Consequently, we propose a neuron-level controllable ICE framework for MKE, namely NeuCon-ICE. It consists of (1) a multimodal contextual neuron identification module that aims to determine where to edit by identifying the tightly coupled multimodal contextual neurons, and (2) a context-aware neuron editing module that aims to determine how to edit by selectively injecting context-aware updates into the identified neurons. Experiments on three representative MLLMs (BLIP-2, MiniGPT-4, and LLaVA 1.5) on ComprehendEdit and E-VQA demonstrate that NeuCon-ICE consistently achieves state-of-the-art overall performance, delivering an overall gain of at least 10.79% across baselines and datasets. The code is available at https://github.com/jc4357/NeuCon-ICE.

Weakly Supervised Object Detection (WSOD) is fundamentally limited by instance ambiguity, manifesting as either part domination (focusing on discriminative fragments) or merged detection (confusing objects with context). Unlike existing approaches that rely on object-level patterns, we draw inspiration from human cognition, where objects are perceived structurally by integrating constituent parts. Building on this perspective, we propose P2WDet (Part-to-Whole Detection), a novel framework that shifts WSOD from traditional instance selection to semantic reconstruction. P2WDet comprises three systematic stages: (1) constructing a comprehensive visually-grounded part vocabulary leveraging large language models; (2) training dedicated part detectors by mining consistent patterns from cross-image proposal clusters; and (3) assembling detected parts into complete objects to enforce structural consistency. Extensive experiments show that P2WDet not only outperforms state-of-the-art methods on standard benchmarks but also demonstrates superior generalization in open-world settings. Furthermore, by decoupling detection from fixed category labels, P2WDet enables the flexible detection of novel or ambiguous objects defined solely by their components.

Recommendation model interpretation aims to reveal the relationships between inputs, model internal representations, and outputs to enhance the transparency, interpretability, and trustworthiness of recommendation systems. However, the inherent complexity and opacity of deep learning models pose challenges for model-level interpretation. Moreover, most existing methods for interpreting recommendation models are tailored to specific architectures or model types, limiting their generalizability across different types of recommenders. In this article, we propose RecSAE, a generalizable probing framework that interprets Recommendation models with Sparse AutoEncoders. The framework extracts interpretable latents from the internal representations of recommendation models and links them to semantic concepts for interpretations. It does not alter original models during interpretations and also enables targeted tuning to models. Experiments on three types of recommendation models (general, graph-based, sequential) with four widely used public datasets demonstrate the effectiveness and generalization of the RecSAE framework. The interpreted concepts are further validated by human experts, showing strong alignment with human perception. Overall, RecSAE serves as a novel step in both model-level interpretations to various types of recommendation models without affecting their functions and offering potential for targeted tuning of models.

Short-video recommender systems typically optimize ranking models using dense user behavioral signals, such as clicks and watch time. However, these signals are only indirect proxies of user satisfaction and often suffer from noise and bias. Recently, explicit satisfaction feedback collected through questionnaires has emerged as a high-quality direct alignment supervision, but is extremely sparse and easily overwhelmed by abundant behavioral data, making it difficult to incorporate into online recommendation models. To address these challenges, we propose a novel framework which is towards End-to-End Alignment of user Satisfaction via Questionnaire, named EASQ, to enable real-time alignment of ranking models with true user satisfaction. Specifically, we first construct an independent parameter pathway for sparse questionnaire signals by combining a multi-task architecture and a lightweight LoRA module. The multi-task design separates sparse satisfaction supervision from dense behavioral signals, preventing the former from being overwhelmed. The LoRA module pre-inject these preferences in a parameter-isolated manner, ensuring stability in the backbone while optimizing user satisfaction. Furthermore, we employ a DPO-based optimization objective tailored for online learning, which aligns the main model outputs with sparse satisfaction signals in real time. This design enables end-to-end online learning, allowing the model to continuously adapt to new questionnaire feedback while maintaining the stability and effectiveness of the backbone. Extensive offline experiments and large-scale online A/B tests demonstrate that EASQ consistently improves user satisfaction metrics across multiple scenarios. EASQ has been successfully deployed in a production short-video recommendation system, delivering significant and stable business gains.

Counterfactual explanations (CEs) provide an intuitive way to understand recommender systems by identifying minimal modifications to user-item interactions that alter recommendation outcomes. Existing CE methods for recommender systems, however, have been evaluated under heterogeneous protocols, using different datasets, recommenders, metrics, and even explanation formats, which hampers reproducibility and fair comparison. In this paper, we systematically reproduce, re-implement, and re-evaluate eleven state-of-the-art CE methods for recommender systems, covering both native explainers (e.g., LIME-RS, SHAP, PRINCE, ACCENT, LXR, GREASE) and specific graph-based explainers originally proposed for GNNs. Building on these implementations, a unified benchmarking framework is proposed to assess explainers along three dimensions: explanation format (implicit vs. explicit), evaluation level (item-level vs. list-level), and perturbation scope (user interaction vectors vs. user-item interaction graphs). Our evaluation protocol includes effectiveness, sparsity, and computational complexity metrics, and extends existing item-level assessments to top-K list-level explanations. Through extensive experiments on three real-world datasets and six representative recommender models, we analyze how well previously reported strengths of CE methods generalize across diverse setups. We observe that the trade-off between effectiveness and sparsity depends strongly on the specific method and evaluation setting, particularly under the explicit format; in addition, explainer performance remains largely consistent across item level and list level evaluations, and several graph-based explainers exhibit notable scalability limitations on large recommender graphs. Our results refine and challenge earlier conclusions about the robustness and practicality of CE generation methods in recommender systems, and provide a reproducible reference point for future work: \href{https://github.com/L2R-UET/CFExpRec}{https://github.com/L2R-UET/CFExpRec}.

Post-click conversion rate (CVR) prediction is a central task in recommender systems, yet selection bias creates a severe distributional gap between the clicked training samples and the entire inference space. To address selection bias, propensity-based methods such as inverse propensity scoring (IPS) and doubly robust (DR) have been adopted, which aim to estimate the unbiased learning objective from biased training samples. However, these approaches assume strictly positive propensities, implying every user-item pair has a nonzero probability of interaction. In practice, such positivity assumption maybe violated, for example, in food-delivery platforms, some restaurants located more than 10 kilometers away will be blocked for recommendation. In this study, we theoretically show that when such zero-propensity samples, termed extrapolation samples exist, both IPS and DR estimators become biased. To overcome this limitation, we propose ExtraDebias method, which enables debiased recommendation in both non-extrapolation and extrapolation samples. Specifically, we first train a propensity model to identify extrapolation samples with extremely small propensity estimates, then estimate their pseudo-label intervals, and derive an upper bound of the learning objective for extrapolation samples. By minimizing the derived upper bound, debiased learning on extrapolation samples is ensured, while unbiased learning on non-extrapolation samples is achieved by standard IPS. Experiments on four real-world offline datasets and one online A/B test show that ExtraDebias effectively minimizes prediction errors on extrapolation samples and achieves optimal performance.

\begin{abstract}Multimodal personality understanding plays a critical role in human-centered artificial intelligence. Previous work mainly focus on learning rich multimodal representations for video personality understanding. However, they often suffer from potential harm caused by subject bias (e.g., observable age and unobservable mental states), as subjects originate from diverse demographic backgrounds. Learning such spurious associations between multimodal features and traits may lead to unfair personality understanding. In this work, we construct a Structural Causal Model (SCM) to analyze the impact of these biases from a causal perspective, and propose a novel \textbf{D}ual \textbf{C}ausal \textbf{A}djustment Network \textbf{(DCAN)} to mitigate the interference of subject attributes on personality understanding. Specifically, we design a Back-door Adjustment Causal Learning (BACL) module to block spurious correlations from observable demographic factors via a prototype-based confounder dictionary, and subsequently apply a Front-door Adjustment Causal Learning (FACL) module to address latent and unobservable biases through a learned mediator dictionary intervention, thereby achieving causal disentanglement of representations for deconfounded reasoning. Importantly, we construct a \textbf{D}emographic-annotated \textbf{M}ultimodal \textbf{S}tudent \textbf{P}ersonality \textbf{(DMSP)} dataset to support the analysis and discussion of fairness-related factors.Extensive experiments on the benchmark dataset CFI-V2 and our DMSP dataset demonstrate that DCAN consistently improves prediction accuracy, reaching 92.11\% and 92.90\%, respectively.Meanwhile, the improvements in the fairness metrics of equal opportunity and demographic parity are 6.57\% and 7.97\% on CFI-V2, and 15.38\% and 20.06\% on the DMSP dataset. Our code and DMSP dataset are availableat \url{https://github.com/Sabrina-han/DCAN}\end{abstract}

LLM-based reranking has been evaluated for some African languages, but whether LLM-based query expansion helps or hurts retrieval for low-resource African languages remains an open question. Adeyemi et al. evaluated cross-lingual LLM reranking for Hausa with English queries, yet to our knowledge no published work has evaluated LLM-based query expansion for Hausa or Fongbe specifically, and no IR evaluation resources were found for Fongbe. This study builds upon our prior work on LLM translation quality evaluation and data augmentation for corpus expansion in Hausa and Fongbe. We propose experiments that compare LLM reranking and query expansion against BM25 and multilingual dense retrieval baselines (mDPR, mContriever) for Hausa and Fongbe using three commercial LLMs. Our completed translation quality assessment confirms a large LLM-capability gap between the two languages (best BLEU: Hausa 15.75 vs. Fongbe 7.18; human scores 4.5/5 vs. 2.2/5), and our data augmentation experiments across three encoder models show that LLM-generated text consistently hurts downstream NER tasks while producing mixed effects on POS tagging, motivating careful language-specific IR evaluation. We plan to use the CIRAL test collection for Hausa and to construct a new cross-lingual test set derived from Fongbe Wikipedia data following the AfriCLIRMatrix methodology.

Research on Information Retrieval (IR) has historically prioritised high-resource languages such as English and Chinese, with less attention given to many low-resource languages. For example, Shona, a Bantu language spoken by approximately 12 million people in Zimbabwe and neighbouring countries, remains under-explored in IR research despite its widespread societal use in Southern Africa. In this work, we present a preliminary study of Shona IR using sparse and dense retrieval models, demonstrating significant performance limitations due to morphological complexity and data scarcity. Based on these findings, we propose to develop a framework to advance Shona IR by developing a large-scale benchmark dataset to support morphology-aware retrieval. We hypothesise that improving Shona IR supports equitable access to digital information and enables language-inclusive AI technologies aligned with global development priorities such as accessibility to education, dissemination of healthcare information, and digital inclusion.

Modern retrieval pipelines increasingly rely on query reformulation and neural reranking to improve effectiveness, but this comes at a significant computational cost and introduces a fundamental tradeoff between recall and query drift. Generating many reformulated queries can substantially increase recall, yet naïvely merging or exhaustively reranking their results is prohibitively expensive. In this work, we argue that the core challenge is not reformulation generation itself, but the adaptive selection of reformulations and their retrieved documents under a strict inference budget. We propose ReformIR, a budget-aware retrieval framework that treats query reformulations as first-class features and performs online relevance estimation using a strong reranker as a teacher. Given multiple reformulated queries, ReformIR constructs a large candidate pool and learns a lightweight surrogate model that estimates document utility from reformulation-specific retrieval signals. Under a fixed reranking budget, the surrogate adaptively prioritizes both reformulations and documents, selectively querying a teacher reranker anchored to the original query. This process increases recall while actively suppressing drift through online feature selection over reformulations. We conduct extensive experiments on the MSMARCO passage corpora and TREC Deep Learning benchmarks (DL19–DL22), evaluating effectiveness under realistic reranking budgets. Our results show that ReformIR consistently outperforms existing reformulation strategies, particularly as the number of reformulations increases, where prior methods suffer from severe quality degradation due to drift. Our findings also suggest a shift in retrieval system design: rather than using large language models as rerankers, their capacity is more effectively leveraged in the reformulation stage with feedback-driven optimization. Our code: This paragraph is centered. https://github.com/VenkteshV/ReformIR

Large Language Models (LLMs) are now widely used for query reformulation and expansion in Information Retrieval, with many studies reporting substantial effectiveness gains. However, these results are typically obtained under heterogeneous experimental conditions, making it difficult to assess which findings are reproducible and which depend on specific implementation choices. In this work, we present a systematic reproducibility and comparative study of ten representative LLM-based query reformulation methods under a unified and strictly controlled experimental framework. We evaluate methods across two architectural LLM families at two parameter scales, three retrieval paradigms (lexical, learned sparse, and dense), and nine benchmark datasets spanning TREC Deep Learning and BEIR. Our results show that reformulation gains are strongly conditioned on the retrieval paradigm, that improvements observed under lexical retrieval do not consistently transfer to neural retrievers, and that larger LLMs do not uniformly yield better downstream performance. These findings clarify the stability and limits of reported gains in prior work. To enable transparent replication and ongoing comparison, we release all prompts, configurations, evaluation scripts, and run files through QueryGym, an open-source reformulation toolkit with a public leaderboard. https://leaderboard.querygym.com

Recently, Large Language Models (LLMs) have demonstrated remarkable reasoning capabilities through reinforcement learning (RL). However, enabling LLM-based agents to effectively orchestrate multiple tools remains an open challenge. In this paper, we introduce Tool-Star, an end-to-end agentic post-training framework that empowers LLM-based web agents to strategically interact with external multi-tool environments. Tool-Star begins with a general tool-integrated data synthesis pipeline that combines two complementary sampling strategies to generate tool-use trajectories, followed by quality normalization and difficulty-aware curriculum construction to filter noisy samples and organize training data from easy to hard. We then introduce a two-stage training paradigm for multi-tool collaborative reasoning: (1) cold-start supervised fine-tuning with tool feedback to bootstrap long-horizon tool-augmented reasoning, and (2) a multi-tool self-critic RL algorithm with hierarchical rewards to reinforce effective tool coordination. Experiments across 13 benchmarks demonstrate Tool-Star's effectiveness. Further analyses provide practical insights for optimizing strategic tool use in web agents. The code is available at https://github.com/RUC-NLPIR/Tool-Star.

Deciding when to stop reviewing the results of a search is a common problem with multiple applications. Existing stopping rules developed within Technology-Assisted Review (TAR) aim to achieve a pre-specified recall target and do not take into account the reason for examining the results, potentially leading to sub-optimal recommendations. This paper applies decision theory to the problem and uses it to derive three practical stopping policies based on the Expected Value of Perfect Information. The approach is applied to two professional search tasks: patent examining and systematic reviewing. Experiments on CLEF-IP and medical systematic review datasets show that the proposed approach generally produces more appropriate stopping decisions than existing methods, as demonstrated by higher net utility under the evaluated cost and payoff settings.

Generative AI is being increasingly integrated into web search for the convenience it provides users. In this work, we aim to understand how generative AI disrupts web search by retrieving and presenting the information and sources differently from traditional search engines. We introduce a public benchmark dataset of 11,500 user queries to support our study and future research of generative search. We compare the search results returned by Google's search engine, the accompanying AI Overview (AIO), and Gemini Flash 2.5 for each query. We have made several key findings. First, we find that for 51.5% of representative, real-user queries, AIOs are generated, and are displayed above the organic search results. Controversial questions frequently result in an AIO. Second, we show that the retrieved sources are substantially different for each search engine (<0.2 average Jaccard similarity). Traditional Google search is significantly more likely to retrieve information from popular or institutional websites in government or education, while generative search engines are significantly more likely to retrieve Google-owned content. Third, we observe that websites that block Google's AI crawler are significantly less likely to be retrieved by AIOs, despite having access to the content. Finally, AIOs are less consistent when processing two runs of the same query, and are less robust to minor query edits. Our findings have important implications for understanding how generative search impacts website visibility, the effectiveness of generative engine optimization techniques, and the information users receive. We call for revenue frameworks to foster a sustainable and mutually beneficial ecosystem for publishers and generative search providers.

A systematic, reliable, and low-cost evaluation of Conversational Information Access (CIA) systems remains an open challenge. Existing reference-based evaluation methods have proven insufficient for evaluating the dynamic nature of information access conversations, while existing LLM-based reference-free methods suffer from evaluation bias and limited generalizability. We propose FACE: a Fine-grained, Aspect-based Conversation Evaluator that evaluates diverse turn- and dialogue-level aspects of conversations. FACE leverages beam search and bandit optimization to select optimized LLM instructions for each evaluation aspect. It assigns scores to atomic information units (particles) using the selected instructions and then aggregates them into a single score. We show that FACE achieves strong correlations with human judgments, outperforming state-of-the-art conversation evaluation methods by a large margin. We further demonstrate that its optimized instructions are transferable across various LLMs and datasets. Additionally, unlike existing LLM-based methods that provide single uninterpretable scores, FACE provides insights into the system performance and enables identifying and locating problems within conversations.

Information retrieval has long relied on evaluations that measure system performance. Improvements on standard evaluation protocols are interpreted as progress in system effectiveness, on the understanding that improved metrics indicate a better experience. However, most evaluations are a drastic abstraction and simplification of that experience. It is reasonable to inquire after the validity of our evaluations, or the degree to which they do in fact represent phenomena we care about. If a metric improves, can we be sure there is a corresponding improvement in real-world effectiveness? We discuss practical ways to discuss, measure, and improve the validity of evaluations in a range of settings. By considering validity, we can make better choices in evaluation protocols; we have a chance to make progress if and when evaluating and retrieving collapse into each other entirely, e.g., with LLM-as-judge; and we can optimise towards systems that people actually want.

Modern information access ecosystems consist of mixtures of systems, such as retrieval systems and large language models, and increasingly rely on marketplaces to mediate access to models, tools, and data, making competition between systems inherent to deployment. In such settings, outcomes are shaped not only by benchmark quality but also by competitive pressure, including user switching, routing decisions, and operational constraints. Yet evaluation is still largely conducted on static benchmarks with accuracy-focused measures that assume systems operate in isolation. This mismatch makes it difficult to predict post-deployment success and obscures competitive effects such as early-adoption advantages and market dominance. We introduce Marketplace Evaluation, a simulation-based paradigm that evaluates information access systems as participants in a competitive marketplace. By simulating repeated interactions and evolving user and agent preferences, the framework enables longitudinal evaluation and marketplace-level metrics, such as retention and market share, that complement and can extend beyond traditional accuracy-based metrics. We formalize the framework and outline a research agenda, motivated by business and economics, around marketplace simulation, metrics, optimization, and adoption in evaluation campaigns like TREC.

In benchmarking of Information Retrieval systems, the Wilcoxon signed-rank test is often treated as a safer alternative to the t-test. This belief is fueled by textbooks and recommendations that portray Wilcoxon as the proper non-parametric alternative because metric scores are not normally distributed. We argue that this narrative is misleading and harmful. A careful review of Statistics textbooks reveals inconsistencies and omissions in how the assumptions underlying these tests are presented, fostering confusion that has propagated into IR research. As a result, Wilcoxon has been routinely misapplied for decades, creating a false sense of safety against a threat that was never there to begin with, while introducing another one so severe that it virtually guarantees the test will break down and mislead researchers. Through a combination of systematic literature review, analysis and empirical demonstrations with TREC data, we show how and why the Wilcoxon test easily loses control of its Type I error rate in IR settings. We conclude that the continued use of Wilcoxon in IR evaluation is unjustified and that abandoning it would improve the methodological soundness of our field.

While deep learning-based approaches dominate the research on collaborative filtering (CF) in recommender systems, recent studies have re-examined this paradigm and found that deep models do not consistently outperform traditional CF alternatives on standard evaluation metrics. Notably, similarity-based methods exhibit substantial advantages. Inspired by these observations, we demonstrate that the performance of a minimalist similarity-based model can match or exceed leading CF approaches, highlighting its significant potential. Despite this, the expressiveness of high-performing similarity-based models is fundamentally confined to pairwise interactions within immediate neighborhoods, overlooking the combinatorial preference patterns in high-order connections. To overcome these limitations and fully unleash the potential of similarity-based models, we propose a novel bidirectional extension framework that systematically generalizes the modeling of similarity to group-level and multi-hop perspectives. However, directly computing these generalized similarities is prohibitively expensive due to their combinatorial complexity. To address this, we develop a MinHash-based estimation algorithm, which ensures the feasibility of the computation and reduces the complexity from exponential to linear. Extensive experiments on three public datasets show that our framework achieves competitive or superior performance over strong baselines, yielding relative improvements ranging from 4\% to 11\% on key metrics, underscoring the enduring power of similarity-based methods in recommender systems.

Flow Matching (FM) has recently been introduced into generative collaborative filtering due to its potential to capture user preferences. However, existing FM-based methods are often trained with only the standard regression objective in flow matching and lack explicit inter-user supervision. This can drive the model toward an averaged preference pattern and generate indistinguishable predictions across users. A natural remedy is to introduce contrastive learning for inter-user separation, yet naive designs face two challenges: (1) In FM-based collaborative filtering, the output is a user's predicted interaction scores over items. Applying contrastive learning directly on this output forces different users to be separated even when they share the same preferences, leading to false repulsion. (2) Flow matching is highly sensitive to its input, so input-level perturbations for view construction can change the generated outcome, making positive pairs semantically mismatched. To address these issues, we propose CoFlowCF, a contrastive flow matching framework for collaborative filtering. Specifically, CoFlowCF introduces hidden contrastive learning by applying inter-user constraints before the flow model's final outputs. We decompose the flow model into a flow encoder and a prediction head, and apply contrastive regularization to the encoder output representations to avoid false repulsion in the prediction space. To build semantically consistent views, we keep the input unchanged and use independent dropout inside the encoder instead of input-level perturbations. Extensive experiments and ablation studies on multiple real-world datasets demonstrate that CoFlowCF consistently improves recommendation performance and robustness. The code is available at https://github.com/DevByQian/CoFlowCF.

Recent research has shown that single-vector embedding retrieval models face a fundamental bottleneck: the finite dimensionality of single-vector representations limits their capacity to represent arbitrary top-k relevant item combinations, even with perfect training. This inherent bottleneck substantially limits the expressive capacity of such models and reduces their ability to capture complex user-item interaction patterns. To overcome this bottleneck, we propose Adaptive Rich-kernelized Contrastive Learning (ARC), which enhances model expressiveness while maintaining the computational efficiency of single-vector retrieval. ARC replaces the fixed inner product with a learnable spherical kernel family parameterized by a truncated Gegenbauer expansion, thereby increasing the model's effective dimensionality while preserving single-vector efficiency and low-pass inductive bias for generalization. Specifically, we construct a positive-definite kernel on the unit sphere using a positive combination of Gegenbauer polynomials, and adopt a contrastive learning objective to jointly learn the polynomials weights and user–item embeddings. Through data-driven optimization, the adaptive kernel induces a more expressive representation space, enabling the model to better capture complex preferences. From a theoretical perspective, the learned kernel implicitly maps embeddings into a higher-dimensional Reproducing Kernel Hilbert Space, allowing the model to capture a broader range of top-k item combinations before reaching the geometric limit imposed by the embedding dimensionality. Consequently, the proposed ARC framework effectively alleviates the inherent representational bottleneck in traditional single-vector embedding models. Extensive experiments on four real-world datasets demonstrate the effectiveness of our method, showing consistent improvements over strong baselines and state-of-the-art models.

Graph Contrastive Learning (GCL) methods for recommendation learn representations by propagating signals over user-item interaction graphs. However, modeling these graphs with homogeneous edges can lead to semantic-structural misalignment, where information is exchanged between structurally adjacent but semantically dissimilar items, adversely affecting retrieval quality. Existing solutions typically rely on auxiliary encoders or additional supervision, increasing model complexity and training cost. We propose ACE (Affective Contrastive Embeddings), a framework that improves representation alignment by incorporating affective semantics into contrastive learning. ACE encodes the affective dimensions of valence and arousal as a topological prior, encouraging consistency between learned embeddings and an affective semantic space distilled from large language models. To operationalize this alignment, we introduce a Semantically Weighted Noise Contrastive Estimation (SW-NCE) loss that modulates contrastive gradients according to users' affective preferences. Experiments on Amazon, Last.fm, and SiTunes demonstrate that ACE consistently improves top-K retrieval performance over 11 baselines while reducing computational overhead. These results indicate that affective geometric alignment is an effective and efficient mechanism for enhancing graph-based retrieval models.

General recommender systems deliver personalized services by learning user and item representations, with the central challenge being how to capture latent user preferences. However, representations derived from sparse interactions often fail to comprehensively characterize user behaviors, thereby limiting recommendation effectiveness. Recent studies attempt to enhance user representations through sophisticated modeling strategies (e.g., intent or language modeling). Nevertheless, most works primarily concentrate on model interpretability instead of representation optimization. This imbalance has led to limited progress, as representation optimization is crucial for recommendation quality by promoting the affinity between users and their interacted items in the feature space, yet remains largely overlooked. To overcome these limitations, we propose DIAURec, a novel representation learning framework that unifies intent and language modeling for recommendation. DIAURec reconstructs representations based on the prototype and distribution intent spaces formed by collaborative and language signals. Furthermore, we design a comprehensive representation optimization strategy. Specifically, we adopts alignment and uniformity as the primary optimization objectives, and incorporates both coarse- and fine-grained matching to achieve effective alignment across different spaces, thereby enhancing representational consistency. Additionally, we further introduce intra-space and interaction regularization to enhance model robustness and prevent representation collapse in reconstructed space representation. Experiments on three public datasets against fifteen baseline methods show that DIAURec consistently outperforms state-of-the-art baselines, fully validating its effectiveness and superiority.

Real-world user behaviors are often noisy due to factors such as human errors, uncertainty, and behavioral ambiguity, which can lead to degraded recommendation performance. To address this issue, recent approaches widely adopt self-supervised learning (SSL), particularly contrastive learning, by generating perturbed views of user interaction sequences and maximizing their mutual information to improve model robustness. However, these methods heavily rely on their pre-defined static augmentation strategies (where the augmentation type remains fixed once chosen) to construct augmented views, leading to two critical challenges: (1) the optimal augmentation type can vary significantly across different scenarios; (2) inappropriate augmentations may even degrade recommendation performance, limiting the effectiveness of SSL. To overcome these limitations, we propose an adaptive augmentation framework. We first unify existing basic augmentation operations into a unified formulation via structured transformation matrices. Building on this formulation, we introduce AsarRec, an Adaptive Sequential Augmentation for Robust Sequential Recommendation. To enable stable end-to-end optimization of discrete and strongly constrained augmentations, AsarRec learns to generate transformation matrices by encoding user sequences into probabilistic transition matrices and projecting them into hard semi-doubly stochastic matrices via a differentiable Semi-Sinkhorn algorithm. To ensure that the learned augmentations benefit downstream performance, we jointly optimize three objectives: diversity (encouraging distinct views), semantic invariance (preserving semantic consistency among views), and informativeness (identifying augmentations most beneficial to recommendation). Extensive experiments on four benchmarks under varying noise levels validate the effectiveness of AsarRec, demonstrating its superior robustness and consistent improvements.

Multimodal Entity Linking (MEL) aims to disambiguate mentions in multimodal contexts by grounding them to specific entities in a knowledge base. A pivotal challenge in MEL is capturing multi-level correspondences: the semantic consistency between mention-entity pairs and the complementary correlations across modalities. However, existing methods often suffer from element dominance due to their reliance on coupled interactions or coarse global aggregations. In response, we propose the Multi-Element Collaborative Interaction (MECI) framework. First, to capture multi-element mention-entity correspondences, we develop a Multi-view Experts Network that leverages a ''divide-and-conquer'' strategy for decoupled feature learning to mitigate element dominance, supported by a KL-guided routing mechanism that governs expert specialization and collaboration. Furthermore, to model cross-modal complementary correlations, we propose a Hierarchical Multimodal Interaction Module, where a dynamic modality-aware weighting network refines interactions across hierarchical semantic levels, thereby integrating multi-granular evidence to counteract element dominance. Finally, we incorporate a generative semantic refinement stage that utilizes large language models for zero-shot re-ranking. Extensive experiments on WikiDiverse, RichpediaMEL, and WikiMEL show that MECI consistently outperforms state-of-the-art baselines, improving Hits@1 by 1.95%, 7.30%, and 2.31%, respectively.

Temporal Knowledge Graph Question Answering (TKGQA) aims to answer natural language questions based on quadruple facts stored in Temporal Knowledge Graphs (TKGs). Recent studies have integrated Large Language Models (LLMs) to handle the complex semantic reasoning required by temporal questions. They typically linearize TKG quadruples into plain text, reducing TKGQA to a top-n text retrieval and context-based question answering problem. However, this textualization process destroys the original quadruple structure and entangles structural and temporal information with text semantics. Moreover, selecting top-n facts based on semantic similarity inevitably introduces a large amount of irrelevant noise into the LLM input, which degrades reasoning performance. To address these limitations, we propose SSR, a Structured Subgraph Retrieval framework for TKGQA with LLMs. In SSR, a Temporal Question Parser is first introduced to extract structured subgraph patterns and temporal constraints from the input questions, leveraging background context obtained via text retrieval. The Subgraph Retrieval module is then applied to directly filter a relevant subgraph from the TKG that contains the facts required to answer the question. The retrieved subgraph is further temporally compressed and subsequently incorporated into the LLM context for final question answering. Experimental results on two benchmark datasets demonstrate that SSR consistently outperforms strong baselines by a clear margin, achieving state-of-the-art performance. Our code is available at https://github.com/zhangli-coding/SSR.

While Knowledge Graphs (KGs) can ground Large Language Models (LLMs) in factual knowledge, existing LLM-KG integration methods for complex reasoning are plagued by computational inefficiency and semantic inconsistency. The tight coupling of LLM inference and KG traversal leads to prohibitive costs, while spurious reasoning paths often misguide learning-based agents, causing reward hacking. To this end, we propose EAT (Exploration-and-Thinking), a novel agentic framework that synergizes LLMs with Reinforcement Learning (RL) for efficient and faithful reasoning on KGs. EAT's core innovations are twofold: (1) an adaptive retrieval-augmented generation mechanism that decouples language comprehension from structured exploration, dramatically improving efficiency; and (2) an LLM-guided reward shaping strategy that explicitly penalizes semantically inconsistent paths and promotes logically valid trajectories grounded in the KG. Extensive experiments on benchmarks like WebQSP, CWQ, and GrailQA show that EAT achieves state-of-the-art performance. Notably, it surpasses the reasoning capability of GPT-4o while utilizing a significantly smaller 8B-parameter LLM.

Continual Knowledge Graph Embedding (CKGE) aims to continually learn embeddings for new knowledge, i.e., entities and relations, while retaining previously acquired knowledge. Most existing CKGE methods mitigate catastrophic forgetting via regularization or replaying old knowledge. They conflate new and old knowledge of an entity within the same embedding space to seek a balance between them. However, entities inherently exhibit multi-faceted semantics that evolve dynamically as their relational contexts change over time. A shared embedding fails to capture and distinguish these temporal semantic variations, degrading lifelong link prediction accuracy across snapshots. To address this, we propose a Multi-Faceted CKGE framework (MF-CKGE) for semantic-aware link prediction. During offline learning, MF-CKGE separates temporal old and new knowledge into distinct embedding spaces to prevent knowledge entanglement and employs semantic decoupling to reduce semantic redundancy, thereby improving space efficiency. During online inference, MF-CKGE adaptively identifies semantically query-relevant entity embeddings by quantifying their semantic importance, reducing interference from query-irrelevant noise. Experiments on eight datasets show that MF-CKGE achieves an average (maximum) improvement of 1.7% (2.7%) and 1.4% (3.8%) in MRR and Hits@10, respectively, over the best baseline. Our source code and datasets are available at: https://anonymous.4open.science/r/MF-CKGE-04E5.

Few-shot knowledge graph completion (FKGC) aims to enhance knowledge reasoning for newly emerging relations using only a limited number of supporting triples. Given the inherent growth and evolution of knowledge graphs (KGs), FKGC faces greater challenges in continual learning (CL) scenarios, as models tend to rely more heavily on historical knowledge in few-shot settings. Existing studies have mainly focused on preserving old knowledge to mitigate catastrophic forgetting. However, these strategies often overlook that not all knowledge is worth retaining. Some prior knowledge may conflict with newly acquired knowledge, either structurally or semantically, distorting entity representations and relation patterns and weakening few-shot inference for new relations. To bridge this research gap, we propose a novel FKGC framework named CSKU, which achieves dynamic knowledge adaptation by selectively unlearning structural and semantic conflicts. Specifically, to preserve entity representation consistency, we design a graph-structured dynamic progressive unlearning mechanism that adaptively mitigates structural conflicts during training. Meanwhile, we introduce a semantic-guided selective unlearning mechanism that dynamically attenuates the impact of conflicting relations. Experimental results on multiple real-world FKGC datasets demonstrate that CSKU significantly outperforms state-of-the-art baselines in both forgetting mitigation and few-shot completion, validating the effectiveness of the proposed selective knowledge unlearning strategy. Our code and datasets are publicly available. https://anonymous.4open.science/r/CSKU.

Knowledge base question answering (KBQA) aims to answer natural language questions using large-scale knowledge bases (KBs). Among various KBQA approaches, semantic parsing-based (SP-based) methods have demonstrated strong effectiveness by generating concise logical forms (LFs) that capture complex subgraph structures and semantic information. Recent research suggests that integrating large language models (LLMs) with SP can achieve significant improvements in the performance and efficiency of KBQA by facilitating the direct generation of LFs with minimal retrieval. However, generating complete LFs with LLMs continues to pose a challenge due to the complexity of the required graph structures and constraints, leading to the significant issue of non-executability. To address these challenges, we propose GCA-KBQA, a step-wise fine-tuned LLM-based framework that employs hop-wise generation, knowledge-assisted calibration, and path-level assembly to construct complete LFs for KBQA. Specifically, we decompose the complex SP process into manageable steps: first, we iteratively generate LFs for each topic entity one hop at a time using a fine-tuned LLM, leveraging KB knowledge to calibrate intermediate outputs and mitigate error propagation. Subsequently, we guide the LLM in assembling path-level LFs from different topic entities, resulting in optimized final LF. We evaluate the proposed method on four KBQA benchmarks spanning two distinct KBs, demonstrating its superior performance compared to state-of-the-art baselines. The code is available at https://github.com/pvfeldt/GCA-KBQA.

Large vision-language models (LVLMs) have shown significant progress in video understanding, but they struggle to scale to long videos due to limited context windows. Existing methods reduce input dimensionality via frame sampling and feature compression, yet discard details and incur high computational cost for post-training. In contrast, retrieval-augmented generation (RAG) that indexes long videos for query retrieval and memory-based methods that maintain evolving long-term stores, offer a lighter and deployment-friendly solution. Nevertheless, they rely on shallow retrieval that selects only top-ranked segments and fails to integrate information across multiple relevant video episodes. Inspired by Multiple-Trace Theory in cognitive psychology, we revisit long video understanding from a probe-echo perspective, in which human episodic memories are activated and integrated in parallel. Building on this insight, we propose ProEchoMem, a cognitive-inspired framework that simulates the probe-echo mechanism: (1) Incremental Episodic Memory Construction builds structured knowledge graphs from video streams; (2) Probe-Driven Memory Activation generates probe signals from user queries to activate all stored traces simultaneously; (3) Memory Echo Synthesis integrates activated traces into a coherent and structured memory echo. Experiments on LongerVideos, LVBench, and cross-domain settings demonstrate the effectiveness of ProEchoMem, with multi-trace probing achieving up to 14.2% higher relevance and ablation studies validating the contribution of each module. The code is available at https://github.com/Applied-Machine-Learning-Lab/SIGIR26_ProEchoMem

Visual document retrieval aims to retrieve a set of document pages relevant to a query from visually rich collections. Existing methods often employ Vision-Language Models (VLMs) to encode queries and visual pages into a shared embedding space, which is then optimized via contrastive training. However, during visual document representation, localized evidence is usually scattered across complex document layouts, making it difficult for retrieval models to capture crucial cues for effective embedding learning. In this paper, we propose Reasoning-Guided Alignment (ReAlign), a method that enhances visual document retrieval by leveraging the reasoning capability of VLMs to provide fine-grained visual document descriptions as supervision signals for training. Specifically, ReAlign employs a superior VLM to identify query-related regions on a page and then generates a query-aware description grounding the cropped visual regions. The retriever is then trained using these region-focused descriptions to align the semantics between queries and visual documents by encouraging the document ranking distribution induced by the region-focused descriptions to match that induced by the original query. Experiments on diverse visually rich document retrieval benchmarks demonstrate that ReAlign consistently improves visual document retrieval performance on both in-domain and out-of-domain datasets, achieving up to 2% relative improvements. Moreover, the advantages of ReAlign generalize across different VLM backbones by guiding models to better focus their attention on critical visual cues for document representation. All code and datasets are available at https://github.com/NEUIR/ReAlign.

Visual document retrieval requires understanding heterogeneous and multi-modal content to satisfy implicit information needs. Recent advances use screenshot-based document encoding with fine-grained late interaction to encode holistic information and capture nuanced alignments, significantly improving retrieval performance. However, retrievers are still trained with coarse global relevance labels, without revealing which regions support the match. As a result, retrievers tend to rely on surface-level cues and struggle to capture implicit semantic connections, hindering their ability to handle non-extractive queries. To improve fine-grained relevance modeling, we propose a Attention-Grounded REtriever Enhancement (AGREE) framework. AGREE leverages cross-modal attention from multimodal large language models (MLLMs) as proxy supervision to guide the retriever in identifying relevant document regions. Specifically, AGREE extracts attention maps from the MLLM that highlight which document regions are attended to based on the query. These attention scores serve as local, region-level relevance signals. During training, AGREE combines local signals with the global document-level relevance label to jointly optimize the retriever. This dual-level supervision enables the model to learn not only whether documents match, but also which content drives relevance. Experiments on the challenging visual document retrieval benchmark, ViDoRe V2, show that AGREE significantly outperforms the global-supervision-only baseline by 12.82% and 5.03% in terms of average nDCG@1 and nDCG@5. Quantitative and qualitative analyses further demonstrate that AGREE promotes deeper alignment between query terms and document regions, moving beyond surface-level matching toward more accurate and interpretable retrieval. Our code is available at: https://github.com/VickiCui/AGREE.

Real-world document question-answering that relies on screenshots, such as bills and forms, requires evidence that is often spatially localised and visually cluttered. However, most Screenshot Language Models (SLMs) encode the entire page holistically and rely on implicit attention to ''find'' relevant content, which limits both accuracy and efficiency. We present RegionSLM, a region-aware SLM designed to explicitly connect the question to its supporting regions. RegionSLM has two key components: (1) a patch-relevance router that learns a query–region relevance distribution, enabling the model to produce a box-free relevance prior at inference; and (2) Relevance-Guided Region Pooling (RGRP), a query-conditional attention–pooling module that aggregates dense features into a small set of region tokens, which preserves grounding signals while reducing computational overhead. To support training and evaluation, we further curate ReDoc, a region-supervised corpus with 105k documents and 350k question-answer pairs, obtained via a question-guided two-step filtering procedure. Extensive experiments on 12 datasets demonstrate that explicitly learning query–region relevance and pooling it into compact region tokens is an effective and practical recipe for document retrieval and understanding.

Multimodal Document Question Answering (MDQA) systems commonly follow a retrieve-then-answer paradigm; however, the retrieval stage often introduces substantial noise, making an effective reranking component indispensable.Existing reranker training frameworks in MDQA rely predominantly on proxy supervision derived from human annotations or large language model (LLM) outputs, which are frequently noisy and, more critically, misaligned with downstream answer quality.To overcome this limitation, we propose \textbf{AD-Reranker}, a novel framework that shifts reranker training from proxy imitation to \textbf{answer-driven} utility optimization.Specifically, we reformulate the reranker as an environment-grounded agent that interacts with a downstream reader, modeled as a deterministic environment.We further design a composite reward function that integrates answer correctness, thereby explicitly incentivizing ranking strategies aligned with downstream task performance.To optimize the agent, we adopt Group Relative Policy Optimization (GRPO), enabling stable and effective group-wise policy learning.Empirical results demonstrate that AD-Reranker achieves superior reranking quality and an optimal accuracy-efficiency trade-off.When integrated into standard MDQA pipelines, AD-Reranker improves accuracy by 1.9\%–5.0\% while reducing the reader’s context token consumption by 15\%–52\%, providing strong evidence for the effectiveness of answer-driven reranker training.

Query auto-completion (QAC) has been widely studied in the context of web search, yet remains underexplored for in-document search, which we term DocQAC. DocQAC aims to enhance search productivity within long documents by helping users craft faster, more precise queries, even for complex or hard-to-spell terms. Unlike traditional WebQAC systems, DocQAC can leverage rich document context, having access not only to the partially typed user query and global historical queries, but also the content of the current document itself, and crucially, the document-specific history of user query interactions. To address this setting, we propose a novel adaptive trie-guided decoding framework that uses user query prefixes to softly steer language models toward high-quality completions. Our approach introduces an adaptive penalty mechanism with tunable hyperparameters, enabling a principled trade-off between model confidence and trie-based guidance. To efficiently incorporate document context, we explore retrieval-augmented generation (RAG) and lightweight contextual document signals such as titles, keyphrases, and summaries. When applied to encoder–decoder models like T5 and BART, our trie-guided framework outperforms strong baselines and even surpasses much larger instruction-tuned models such as LLaMA-3 and Phi-3 in seen-query settings. This demonstrates its practicality for real-world DocQAC system deployments, where efficiency and scalability are critical. We evaluate our method on a newly introduced DocQAC benchmark derived from ORCAS, enriched with query–document pairs. We make both the DocQAC dataset https://bit.ly/3IGEkbH and code https://github.com/rahcode7/DocQAC publicly available.

In this study, we applied the "personalized diversity nudge framework" with the goal of expanding readers' reading coverage in terms of news locality (i.e., domestic and world news). We designed a novel topic-locality dual calibration algorithmic nudge and a large language model-based news personalization presentation nudge, then launched a 5-week real-user study with 120 U.S. news readers on the news recommendation experiment platform POPROX. With interaction logs and survey responses, we found that algorithmic nudges can successfully increase exposure and consumption diversity, while the impact of LLM-based presentation nudges varied. Participant-level topic interest is a strong predictor of individual clicks, while highlighting the relevance of news articles to prior read articles outperforms generic topic-based and no personalization. We also demonstrate that longitudinal exposure to calibrated news may shift readers' reading habits to value a balanced news digest from both domestic and world articles. Our results provide direction for future work on nudging for diverse consumption in news recommendation systems.

Medication recommendations seek to deliver personalized drug combinations based on patients' clinical information, where ensuring transparent decision-making and personalized safety are of paramount importance. However, existing approaches predominantly operate as ''black boxes'', offering limited interpretability and hindering clinician trust. Furthermore, they typically rely on fixed, patient-agnostic safety constraints, neglecting the heterogeneity of risk tolerance across individuals, which poses significant risks to vulnerable populations. To address these gaps, we propose RES-MR, a novel Risk-aware Reasoning framework for Explainable and Safe Medication Recommendation using LLMs. RES-MR follows a two-stage training paradigm: (1) an explainable clinical reasoning distillation stage to elicit diagnostic capabilities by fine-tuning on synthesized reasoning trajectories, grounded in patient-specific knowledge from graph retrieval; and (2) a risk-aware policy optimization stage to dynamically balance therapeutic efficacy with personalized safety. This stage employs disentangled drug factors to construct patient-specific reference prototypes that capture desired therapeutic drug features and risk features to avoid. These prototypes enable risk boundary calibration and guide the model toward safe recommendations via reward shaping and clipping. Extensive experiments on benchmark datasets demonstrate that RES-MR significantly outperforms state-of-the-art baselines in terms of accuracy, safety, and explainability. Our code and data are available at: https://github.com/wangcong2001/RES-MR.

E-commerce search rankers optimized for engagement or conversion often homogenize first-page results, under-serving ambiguous or exploratory queries. Prior diversification methods typically assume per-query intent/aspect supervision or heavy re-ranking, which is costly and brittle under production constraints. We proposed a supervision-free, deployable framework for demand-calibrated facet diversification. From independent holdout traffic, we estimate a cohort-level target facet distribution using session-attributed downstream engagement, capturing how much diversity users actually demand. At serving time, we steer the first-page slate toward this target with a lightweight soft-quota score adjustment over a short candidate prefix, preserving the base ranker's within-facet ordering and adding negligible latency. Large-scale online experiments on high-volume traffic show consistent improvements in revenue and engagement while reducing abandonment. We further find that delivered first-page facet mixes move closer to the engagement-derived target, indicating effectiveness of the proposed mechanism.

Retrieval-augmented generation (RAG) has been used to improve the accuracy and transparency of outputs produced by large language models (LLMs) by integrating external knowledge; however, applying RAG to low-resource languages presents unique challenges, including poor embedding representations, low retrieval quality, and semantic gaps caused by the scarcity of digital documents. In this proposal, we address these challenges for a selected low-resource language, Amharic, by using translative and debate-based RAG techniques to improve retrieval and reasoning. The proposal outlines the key problems and research gaps in applying RAG to low-resource languages and introduces a method to enhance RAG performance for Amharic. Additionally, we introduce the first comprehensive Amharic Retrieval-Augmented Generation Benchmark (ARGB), designed to capture grammatical, cultural, and writing-system-specific constraints of the Amharic language. ARGB evaluates not only retrieval and generation quality, but also noise robustness, counterfactual robustness, negative rejection, and multi-source information integration, providing a holistic assessment of RAG capabilities. This work aims to improve the reliability, accessibility, and inclusiveness of AI systems for Amharic speakers while providing a scalable framework for other low-resource languages. Current progress on the code and benchmark can be found on this GitHub link: \hrefhttps://github.com/AbelAlemu155/AmharicRAG link.

The deployment of information retrieval (IR) systems in agriculture remains particularly challenging in low-resource environments, where limited access to expert knowledge, data resources, and digital infrastructure constrains effective information access. Recent advances in large language models (LLMs) have enabled conversational interfaces that generate direct answers to user queries, increasingly functioning as retrieval outputs in IR systems. However, little is known about how such generated answers behave in low-resource language settings. This study examines Swahili maize-related question answering as a low-resource agricultural IR task and investigates whether LLM-generated responses are detectably different from human-authored answers. Using a dataset of 127 questions with responses from human experts, ChatGPT, and Gemini, we analyze textual characteristics including answer length, lexical diversity, and TF-IDF cosine similarity. A multi-class TF-IDF and logistic regression classifier is then used to evaluate source detectability. Results show that LLM-generated answers are systematically longer, less lexically diverse, and exhibit moderate similarity to human responses. The classifier achieves 0.98 accuracy, indicating that generated responses exhibit consistent and learnable patterns. These findings highlight detectability as a useful diagnostic signal for evaluating generative IR systems, particularly in low-resource environments where ground-truth evaluation is limited.

Despite major advances in information retrieval (IR) and language technologies, speakers of extremely low-resource ethnic community languages (ECLs) remain largely excluded from digital information access. With Tanzania having over 120 languages, yet most IR systems and digital assistants primarily support English and, to a limited extent, Kiswahili. This disproportionately affects rural populations (over 65%), where ECLs are used alongside Kiswahili in daily life. This paper presents the ongoing Ethnic Voices initiative, which addresses this gap by building foundational text and speech resources explicitly relevant to IR research in extremely low-resource multilingual settings. The project targets six ECLs in Tanzania, with current data collection completed for Hehe and Luguru. Data collection from the two ECLs has produced over 800 Kiswahili-ECL parallel sentences per language and over 95 hours of speech. Preliminary survey results show strong demand for multilingual and voice-based access to information. It further identifies language barriers as a primary limitation of existing English-centric IR systems. This also paper discusses the IR implications of these findings and outlines directions toward inclusive retrieval and evaluation in extremely low-resource languages. Beyond resource creation, we position the collected dataset as IR-enabling infrastructure that supports the formulation of retrieval tasks, the development of baseline systems, and the design of evaluation protocols in benchmark-scarce, extremely low-resource multilingual settings.

Learned Sparse IR models, such as SPLADE, offer an excellent efficiency-effectiveness tradeoff. However, they rely on the underlying backbone vocabulary, which might hinder performance (polysemicity and synonymy) and pose a challenge for multi-lingual and multi-modal usages. To solve this limitation, we propose to replace the backbone vocabulary with a latent space of semantic concepts learned using Sparse Auto-Encoders (SAE). Throughout this paper, we study the compatibility of these 2 concepts, explore training approaches, and analyze the differences between our SAE-SPLADE model and traditional SPLADE models. Our experiments demonstrate that SAE-SPLADE achieves retrieval performance comparable to SPLADE on both in-domain and out-of-domain tasks while offering improved efficiency.

While advanced foundation models like ModernBERT significantly outperform older architectures in dense retrieval, they surprisingly lag behind the aging BERT-base baseline in learned sparse retrieval (LSR). We identify the root cause as the Vocabulary Gap : modern tokenizers utilize raw, case-sensitive vocabularies designed for lossless reconstruction, which map single semantic units to redundant surface forms, wasting model capacity on morphological noise and hindering lexical matching. We formalize this intuition through a theoretical framework, demonstrating that appropriate vocabulary coarse-graining can tighten the generalization bounds by reducing complexity of the hypothesis class, provided that semantic integrity is preserved. To resolve this, we propose Vocabulary Transfer (VT), a model-agnostic framework that migrates advanced encoders to sparse-friendly, normalized vocabularies with minimal computational cost. VT utilizes a novel Semantic Initialization via spatial topology to preserve geometric structure and an Activation Potential Calibration (APC) mechanism to align pre-trained manifolds with sparsity constraints, preventing the dead neuron and dense collapse observed in standard fine-tuning. Empirically, VT is universally effective: it enables ModernBERT to achieve state-of-the-art performance on the BEIR benchmark (52.4 nDCG, a +4.7 improvement), resuscitates failing models like RoBERTa-large, and generalizes seamlessly to inference-free architectures and specialized domains. These results confirm that the performance lag is not an architectural deficiency but a solvable vocabulary mismatch. We've released our code and models.. https://anonymous.4open.science/r/vocab-transfer/. All details included.

Recent work on neural sparse retrieval has demonstrated strong relevance by leveraging Large Language Models (LLMs) for semantic term expansion. However, learned models paired with previous sparsification techniques still yield overly long document and query vectors partly due to a large LLM vocabulary, imposing a serious challenge to retrieval time and space efficiency. This paper proposes a scheme for optimizing model sparsity through a synergy of adaptive strategies, including learnable soft top-??, per-term thresholding, and FLOPs regularization to increase the sparsity of query and document vectors. Experimental results with Lion-SP model on the MS MARCO and BEIR datasets demonstrate that the proposed scheme can outperform the baselines by significantly reducing the average query and document lengths. Our scheme can achieve much shorter retrieval latency and lower storage cost while maintaining highly competitive relevance.

Learned sparse retrieval (LSR) is a popular method for first-stage retrieval because it combines the semantic matching of language models with efficient CPU-friendly algorithms. Previous work aggregates blocks into ''superblocks'' to quickly skip the visitation of blocks during query processing by using an advanced pruning heuristic. This paper proposes a simple and effective superblock pruning scheme that reduces the overhead of superblock score computation while preserving competitive relevance. It combines this scheme with a compact index structure and a robust zero-shot configuration that is effective across LSR models and multiple datasets. This paper provides an analytical justification and evaluation on the MS MARCO and BEIR datasets, demonstrating that the proposed scheme can be a strong alternative for efficient sparse retrieval.

Score-at-a-Time (SaaT) retrieval, utilising impact-ordered indexes, remains a relatively overlooked search strategy with increasing importance. Within the context of learned sparse representations and approximate retrieval, SaaT has proven to be a competitive alternative to the popular Document-at-a-Time (DaaT) approach. Yet, there are only two notable implementations of SaaT search engines: JASSv2 and IOQP. We are interested in the differences between these systems and how those differences affect performance. We identify differences in postings lists due to indexing, ranking functions, and quantization. Thus, we introduce ciffTools for quantizing the ciff indexes used by both search engines; eliminating these differences. Then, in a reproducibility study we reproduce previous experiments investigating the efficiency gap between the systems. They also differ in their compression codecs, with IOQP using SIMD BP-128 and JASSv2 using Elias Gamma SIMD VB. We, unexpectedly, find in another reproducibility experiment that SIMD BP-128 and QMX outperform Elias Gamma SIMD VB in situ. Finally, we investigate the CPU effect, and find the engines are affected differently. Overall, JASSv2 has faster median latency on a server-grade CPU, while on a desktop-grade they are evenly matched. IOQP has faster tail latency regardless of CPU. Our work reduces the throughput difference between JASSv2 and IOQP and offers insights into which aspects affect the efficiency of SaaT.

Learned sparse retrieval models such as SPLADE combine the effectiveness of neural architectures with the efficiency of inverted indices. As these models assign weights to terms from a fixed vocabulary, interpretability is often touted as a major benefit of these models. However, the emergence of wacky weights, i.e., expansion terms that appear semantically unrelated to the input, limits interpretability. While prior research has anecdotally observed this phenomenon, there is a lack of systematic understanding regarding their origins, prevalence, and contribution to retrieval effectiveness. In this paper, we reproduce SPLADE-v2 to systematically investigate wacky weights across the SPLADE family of models. We present a comprehensive dissection of wacky weights, providing a formal definition of wackiness based on the lexical utility of expansion terms. Furthermore, we introduce a novel measure to compare the prevalence of these tokens across models with varying vocabularies and sparsity levels. Beyond reproducing the original SPLADE-v2, we train it with various loss functions, datasets, and backbone transformers to isolate the factors contributing to wackiness. Our results show that larger vocabularies are associated with a higher prevalence of wacky tokens, while stricter sparsity regularizers are associated with lower prevalence. Finally, we find that wacky weights are used primarily for in-domain effectiveness rather than out-of-domain generalization.

Generative List Recommendation (GLR) models have shown superior performance in E-commerce by directly generating high-quality recommendation lists through sequential item selection. While Online Learning (OL) has proven valuable for streaming point-wise recommendation models in adapting to dynamic user preferences, its application to GLR remains largely unexplored, due in large part to the inefficiency and instability of conventional on-policy reinforcement learning algorithms used in existing GLR approaches. Existing approaches typically rely on surrogate losses, which provide indirect and biased gradient estimates, making them ill-suited for the rapid, subtle distribution shifts common in real-world E-commerce environments. In this paper, we propose Environment Policy Optimization (EPO), a novel GLR model that fundamentally reshapes policy learning by exploiting the differentiability of the environment within the Generator-Evaluator framework. EPO recognizes that the evaluator is a neural network, capable of providing gradient signals. By directly utilizing these gradients, EPO enables end-to-end optimization of the total list-wise reward—the true objective. To ensure differentiable list generation, EPO introduces a new indexing and exploration strategy based on NeuralSort and Gumbel noise, which relaxes discrete item selection into a continuous, gradient-friendly operation. EPO not only demonstrates strong performance in offline evaluations but also unlocks the potential of online learning for GLR at an industrial scale, yielding a 1.18% relative improvement in user clicks in online A/B tests. The results underscore the critical role of EPO in providing the sensitivity, stability, and gradient fidelity necessary for effective real-time adaptation in streaming and dynamic recommendation environments. Moreover, EPO reached the baseline performance with a training time deduction of 76% under identical hardware conditions.

Generative recommendation (GR) has demonstrated substantial potential for sequential recommendation in an end-to-end generation paradigm. However, existing models suffer from severe cold-start collapse, i.e., recommendation accuracy on cold-start items drops to near zero. Current solutions rely on retraining with cold-start interactions, which faces sparse feedback, high computational cost, and delayed updates, diminishing practical utility in rapidly evolving catalogs in recommendation. Inspired by NLP model editing (enabling training-free knowledge injection into large language models), we explore applying this paradigm to generative recommendation, but face two fundamental challenges: (1)~sequential data in GR lacks explicit subject–object binding (a core NLP sentence structure), hindering targeted model edits; (2)~sequential data in GR has no fixed token co-occurrence patterns (unlike NLP's stable phrases), making multi-token injection unreliable. To address these, we propose GenRecEdit, the first model editing framework tailored for generative recommendation. Specifically, we: (1)~mitigate the absence of sentence structure by explicitly modeling the intrinsic relationship between the entire sequence context and the next-token (e.g., semantic IDs, SIDs) generation; (2)~adopt iterative token-level edits to effectively inject token bundles (e.g., items); and (3)~introduce a One-One trigger mechanism to avoid interactions among multiple token-level edits during inference. Extensive experiments across multiple datasets demonstrate that øurname substantially improves recommendation performance on cold-start items while preserving the model's original recommendation quality. Moreover, øurname achieves these gains with only approximately 9.5% of the training time required by retraining, significantly reducing computational cost and enabling efficient, frequent model updates.

Recommender systems aim to leverage user interaction signals to recommend items that users are likely to be interested in. Motivated by the success of Large Language Model (LLMs), generative recommendation (GR) has recently gained increasing attention, typically following a two-stage paradigm: supervised fine-tuning followed by preference alignment. However, aligning generative recommenders with users' personalized preferences remains challenging, as user feedback is inherently heterogeneous and uncertain. Different types of user interaction signals reflect varying levels of intent and should therefore be modeled differently. In this work, we propose Adaptive Mix Preference Optimization (AMPO), an adaptive alignment framework that mixes likelihood and preference objectives with self-calibrated, sample-wise confidence adjustment. AMPO introduces an adaptive target margin that leverages the model's own probability ratio to modulate optimization strength: confident pairs receive full margins that reinforce correct rankings, while uncertain pairs receive reduced margins that prevent overfitting to ambiguous signals. Additionally, AMPO incorporates negative log-likelihood regularization on preferred items to counteract likelihood displacement, a phenomenon where contrastive objectives cause preferred and non-preferred probabilities to collapse simultaneously. Such a design eliminates the need for a reference model, yielding up to 2.5x speedup and 35% memory reduction. Extensive experiments on public benchmarks and a large-scale industrial dataset demonstrate consistent improvements in ranking metrics. Online A/B tests on a major e-commerce platform further confirm statistically significant gains in click-through and conversion rates. The code is available at https://github.com/jumbo-q/ampo.

Language Models (LMs) have been widely used in recommender systems to incorporate textual information of items into item IDs, leveraging their advanced language understanding and generation capabilities. Recently, generative recommender systems have utilized the reasoning abilities of LMs to directly generate index tokens for potential items of interest based on the user's interaction history. To inject diverse item knowledge into LMs, prompt templates with detailed task descriptions and various indexing techniques derived from diverse item information have been explored. This paper focuses on the inconsistency in outputs generated by variations in input prompt templates and item index types, even with the same user's interaction history. Our in-depth quantitative analysis reveals that preference knowledge learned from diverse prompt templates and heterogeneous indices differs significantly, indicating a high potential for complementarity. To fully exploit this complementarity and provide consistent performance under varying prompts and item indices, we propose MVIGER, a unified variational framework that models selection among these information sources as a categorical latent variable with a learnable prior. During inference, this prior enables the model to adaptively select the most relevant source or aggregate predictions across multiple sources, thereby ensuring high-quality recommendation across diverse template-index combinations. We validate the effectiveness of MVIGER on three real-world datasets, demonstrating its superior performance over existing generative recommender baselines through the effective integration of complementary knowledge.

Recent years have seen a rapid surge in research leveraging Large Language Models (LLMs) for recommendation. These methods typically employ supervised fine-tuning (SFT) to adapt LLMs to recommendation scenarios, and utilize beam search during inference to efficiently retrieve B top-ranked recommended items. However, we identify a critical training-inference inconsistency: while SFT optimizes the overall probability of positive items, it does not guarantee that such items will be retrieved by beam search even if they possess high overall probabilities. Due to the greedy pruning mechanism, beam search can prematurely discard a positive item once its prefix probability is insufficient. To address this inconsistency, we propose BEAR (Beam-SEarch-Aware Regularization), a novel fine-tuning objective that explicitly accounts for beam search behavior during training. Rather than directly simulating beam search for each instance during training, which is computationally prohibitive, BEAR enforces a relaxed necessary condition: each token in a positive item must rank within the top-B candidate tokens at each decoding step. This objective effectively mitigates the risk of incorrect pruning while incurring negligible computational overhead compared to standard SFT. Extensive experiments across four real-world datasets demonstrate that BEAR significantly outperforms strong baselines. Code is available at https://github.com/Tiny-Snow/BEAR-SIGIR-2026.

Community search is a fundamental graph-based retrieval problem that aims to identify a query-dependent subgraph whose nodes exhibit strong internal connectivity. While recent learning-based methods improve retrieval effectiveness via graph representation learning, they follow a ''one-use-one-train'' paradigm that requires retraining or fine-tuning for each target graph, leading to high data dependency, high training costs, and limited generalization. To handle this, we propose OFA-CS, a ''one-for-all'' community search framework trained once on source datasets and directly deployed to arbitrary unseen graphs without retraining or fine-tuning, while preserving strong performance. Specifically, we introduce a Spectral-Aware Feature Alignment module to unify feature dimensionality and align cross-domain semantics in a community-aware manner. We further develop a Graph Diffusion Tokenized Transformer that constructs hybrid token sequences from local and global structural contexts for Transformer encoding, and applies diffusion-based refinement to mitigate distribution shifts on unseen graphs. With the unified representations, communities are efficiently retrieved via a modularity-driven search procedure. Extensive experiments on diverse real-world graphs demonstrate that OFA-CS achieves strong cross-domain generalization and competitive retrieval effectiveness against state-of-the-art methods, without requiring target-domain supervision.

Several recent works utilize deep learning (DL) techniques for subgraph retrieval via matching, yet most only return approximate isomorphism relations between queries and data graphs--failing to retrieve all exact matching locations, a critical demand for structured graph retrieval in information retrieval. Unlike these DL-based approximate methods, we propose a learning-based framework for subgraph retrieval, called the graph neural network (GNN)-based anchor embedding framework (GNN-AE), which can efficiently retrieve all exact matching locations. In contrast to most traditional exact subgraph matching methods, which create auxiliary structures online for each query, our method has two core optimizations: (1) We construct offline, one-time-only efficient embedding indices for small feature subgraphs (namely, anchored subgraphs and anchored paths) in the data graph and obtain candidates for the query on these indexed feature subgraphs, trading space for time to reduce online query latency; (2) We leverage GNNs to perform graph isomorphism tests on indexed feature subgraphs and generate low-conflict embeddings for these feature subgraphs, yielding a high-quality, compact set of candidates that further enhances query efficiency. Beyond these core optimizations, we develop a parallel matching growth algorithm and design a cost-based DFS query strategy to retrieve all matching locations. Extensive experiments on both real and synthetic datasets validate the efficiency and effectiveness of our GNN-AE for exact subgraph retrieval.

Discovering cohesive groups is a fundamental primitive in graph-based recommender systems, underpinning tasks such as social recommendation, bundle discovery, and community-aware modeling. In interaction graphs, cohesion is often modeled as the γ-quasi-clique, an induced subgraph whose internal edge density meets a user-defined threshold γ. This formulation provides explicit control over within-group connectivity while accommodating the sparsity inherent in real-world data. However, ensuring explicit density constraints while maintaining robustness remains challenging for existing heuristic approaches. This paper presents EDQC, an effective framework for cohesive group discovery under explicit density constraints. EDQC leverages a lightweight energy diffusion process to rank vertices for localizing promising candidate regions. Guided by this ranking, the framework extracts and refines a candidate subgraph to ensure the output strictly satisfies the target density requirement. Extensive experiments on 75 real-world graphs across varying density thresholds demonstrate that EDQC identifies the largest mean γ-quasi-cliques in the vast majority of cases, achieving lower variance than the state-of-the-art methods while maintaining competitive runtime, making it a robust and practical solution for cohesive group discovery in graph-based recommender systems.

Dynamic signed networks (DSNs) are common in online platforms, where time-stamped positive and negative relations evolve over time. A core task in DSNs is dynamic edge prediction, which forecasts future relations by jointly modeling edge existence and polarity (positive, negative, or non-existent). However, existing dynamic signed network embedding (DSNE) methods often entangle positive and negative signals within a shared temporal state and rely on node-specific temporal trajectories, which can obscure polarity-asymmetric dynamics and harm inductive generalization, especially under cold-start evaluation. We study an inductive setting where each test edge contains at least one endpoint node held out from training, while its interactions prior to the prediction time are available as historical evidence. The model must therefore infer representations for unseen nodes solely from such limited history. We propose IDP-DSN, an Inductive Dual-Polarity framework for Dynamic Signed Networks. IDP-DSN maintains sign-selective memories to model positive and negative temporal dynamics separately, performs history-only neighborhood inference for unseen nodes (instead of learned node-wise trajectories), and enforces polarity-wise static--dynamic disentanglement via an orthogonality regularizer. Experiments on BitcoinAlpha, BitcoinOTC, Wiki-RfA, and Epinions demonstrate consistent improvements over the strongest baselines, achieving relative Macro-F1 gains of 16.8/23.4%, 16.9/24%, 30.1/25.5%, and 18.7/28.9% in the transductive/inductive settings, respectively. These results highlight the effectiveness of IDP-DSN on DSNs, particularly under inductive cold-start evaluation for dynamic signed edge prediction.

Knowledge Graphs (KGs) have proven highly effective for recommendation systems by capturing latent item relationships, while recent integration of Large Language Models (LLMs) has further enhanced semantic understanding and addressed knowledge sparsity issues. Nevertheless, current KG&LLM-based methods still face three main limitations: 1) inadequate modeling of implicit semantics relationships beyond explicit KG links; 2) suboptimal single-channel fusion of ID and LLM embeddings, which often leads to signal interference and blurred representations; and 3) insufficient consideration of user-item interaction frequency variations in recommendation strategies. To address these challenges, we propose the Dual-Channel Graph Learning (DCGL) framework, featuring three key innovations: 1) a dual-channel architecture that structurally decoupling rich semantic information from user behavioral patterns, preventing early interference; 2) a multi-level contrastive learning mechanism that enhances robustness against KG noise through intra-view contrast and bridges semantic gaps between channels via inter-view alignment; and 3) a dynamic fusion mechanism that adaptively balances semantic generalization and behavioral specificity based on interaction frequency, resolving the cascading limitation. Extensive experiments on four real-world datasets show that DCGL consistently outperforms state-of-the-art methods, yielding substantial improvements in sparse scenarios while maintaining precision for active users. Our code is available at https://github.com/XinchiZou/DCGL.

Multimodal Retrieval-Augmented Generation (MRAG) has shown promise in mitigating hallucinations in Multimodal Large Language Models (MLLMs) by incorporating external knowledge. However, existing methods typically adhere to rigid retrieval paradigms by mimicking fixed retrieval trajectories and thus fail to fully exploit the knowledge of different retrieval experts through dynamic interaction based on the model's knowledge needs or evolving reasoning states. To overcome this limitation, we introduce Mixture-of-Retrieval Experts (MoRE), a novel framework that enables MLLMs to collaboratively interact with diverse retrieval experts for more effective knowledge exploitation. Specifically, MoRE learns to dynamically determine which expert to engage with, conditioned on the evolving reasoning state. To effectively train this capability, we propose Stepwise Group Relative Policy Optimization (Step-GRPO), which goes beyond sparse outcome-based supervision by encouraging MLLMs to interact with multiple retrieval experts and synthesize fine-grained rewards, thereby teaching the MLLM to fully coordinate all experts when answering a given query. Experimental results on diverse open-domain QA benchmarks demonstrate the effectiveness of MoRE, achieving average performance gains of over 7% compared to competitive baselines. Notably, MoRE exhibits strong adaptability by dynamically coordinating heterogeneous experts to precisely locate relevant information, validating its capability for robust, reasoning-driven expert collaboration. All codes and data are released on https://github.com/OpenBMB/MoRE.

Retrieval-Augmented Generation (RAG) has emerged as a powerful paradigm for enhancing Large Language Models (LLMs) by retrieving relevant documents from an external corpus. However, existing RAG systems primarily focus on unimodal text documents, and often fall short in real-world scenarios where both queries and documents may contain mixed modalities (such as text and images). In this paper, we address the challenge of Universal Retrieval-Augmented Generation (URAG), which involves retrieving and reasoning over mixed-modal information to improve vision-language generation. To this end, we propose Nyx, a unified mixed-modal-to-mixed-modal retriever tailored for URAG scenarios. To mitigate the scarcity of realistic mixed-modal data, we introduce a four-stage automated pipeline for data generation and filtering, leveraging web documents to construct NyxQA, a dataset comprising diverse mixed-modal question-answer pairs that better reflect real-world information needs. Building on this high-quality dataset, we adopt a two-stage training framework for Nyx: we first perform pre-training on NyxQA along with a variety of open-source retrieval datasets, followed by supervised fine-tuning using feedback from downstream Vision-Language Models (VLMs) to align retrieval outputs with generative preferences. Experimental results demonstrate that Nyx not only performs competitively on standard text-only RAG benchmarks, but also excels in the more general and realistic URAG setting, significantly improving generation quality in vision-language tasks.

Retrieval-Augmented Generation (RAG) has emerged as an effective paradigm for expanding the knowledge capacity of Multimodal Large Language Models (MLLMs) by incorporating external knowledge sources into the generation process, and has been widely adopted for knowledge-based Visual Question Answering (VQA). Despite impressive advancements, vanilla RAG-based VQA methods that rely on unstructured documents and overlook the structural relations among knowledge elements frequently introduce irrelevant or misleading content, degrading answer accuracy and reliability. To overcome these challenges, a promising solution is to integrate multimodal knowledge graphs (KGs) into RAG-based VQA frameworks, thereby enhancing generation through structured multimodal knowledge. To this end, this paper proposes mKG-RAG, a novel retrieval-augmented generation framework built upon multimodal KGs for knowledge-intensive VQA tasks. Specifically, mKG-RAG leverages MLLM-driven graph extraction and vision-text matching to distill semantically consistent, modality-complementary entities and relations from multimodal documents, constructing high-quality multimodal KGs as structured knowledge representations. Furthermore, a dual-stage retrieval strategy equipped with a query-aware multimodal retriever is introduced to improve retrieval efficiency while progressively refining precision. Comprehensive experiments demonstrate that our approach significantly outperforms existing approaches and sets new state-of-the-art results for knowledge-based VQA. The code is available at https://github.com/xandery-geek/mKG-RAG.

Iterative Retrieval-Augmented Generation (iRAG) has emerged as a powerful paradigm for answering complex multi-hop questions by progressively retrieving and reasoning over external documents. However, current systems predominantly operate on parsed text, which creates two critical bottlenecks: (1) Coarse-grained attribution, where users are burdened with manually locating evidence within lengthy documents based on vague text-level citations; and (2) Visual semantic loss, where the conversion of visually rich documents (e.g., slides, PDFs with charts) into text discards spatial logic and layout cues essential for reasoning. To bridge this gap, we present Chain of Evidence (CoE), a retriever-agnostic visual attribution framework that leverages Vision-Language Models to reason directly over screenshots of retrieved document candidates. CoE eliminates format-specific parsing and outputs precise bounding boxes, visualizing the complete reasoning chain within the retrieved candidate set. We evaluate CoE on two distinct benchmarks: Wiki-CoE, a large-scale dataset of structured web pages derived from 2WikiMultiHopQA, and SlideVQA, a challenging dataset of presentation slides featuring complex diagrams and free-form layouts. Experiments demonstrate that fine-tuned Qwen3-VL-8B-Instruct achieves robust performance, significantly outperforming text-based baselines in scenarios requiring visual layout understanding, while establishing a retriever-agnostic solution for pixel-level interpretable iRAG. Our code is available at https://github.com/PeiYangLiu/CoE.git.

Multimodal Retrieval-Augmented Generation (MRAG) is widely adopted for Multimodal Large Language Models (MLLMs) with external evidence to reduce hallucinations. Despite its success, most existing MRAG frameworks treat retrieved evidence as indivisible documents, implicitly assuming that all content within a document is equally informative. In practice, however, sometimes only a small fraction of a document is relevant to a given query, while the remaining content introduces substantial noise that may lead to performance degradation. We address this fundamental limitation by reframing MRAG as a fine-grained evidence selection problem. We propose Fragment-level Evidence Selection for RAG (FES-RAG), a framework that selects atomic multimodal fragments rather than entire documents as grounding evidence. FES-RAG decomposes retrieved multimodal documents into sentence-level textual fragments and region-level visual fragments, enabling precise identification of evidence that directly supports generation. To guide fragment selection, we introduce Fragment Information Gain (FIG), a principled metric that measures the marginal contribution of each fragment to the MLLM's generation confidence. Based on FIG, we distill fragment-level utility judgments from a high-capacity MLLM into a lightweight selector, achieving accurate evidence selection with low inference overhead. Experiments on the M2RAG benchmark show that FES-RAG consistently outperforms state-of-the-art document-level MRAG methods, achieving up to 27% relative improvement in CIDEr. By selecting fewer yet more informative fragments, our approach substantially reduces context length while improving factual accuracy and generation coherence.

Previous query performance prediction (QPP) methods typically return a single scalar value and do not require the predicted values to approximate a specific IR evaluation measure, leading to drawbacks: (i) a single scalar is insufficient to accurately represent different IR measures, especially when metrics do not correlate, and (ii) a single scalar limits the interpretability of QPP methods because solely using a scalar is insufficient to explain QPP results. To address these issues, we propose a QPP framework using automatically generated relevance judgments (QPP-GenRE), which decomposes QPP into independent subtasks of predicting the relevance of each item in a ranked list to a given query. This allows us to predict any IR measure using the generated relevance judgments as pseudo-labels. This also allows us to interpret predicted IR measures, and identify errors in generated relevance judgments to improve QPP quality. We predict an item’s relevance by using open-source LLMs. We face two main challenges: (i) excessive computational costs of judging an entire corpus for predicting a metric considering recall, and (ii) limited performance in prompting open-source LLMs. To solve both, we devise an approximation strategy and propose to fine-tune open-source LLMs using human-labeled relevance judgments. Experiments on multiple ad-hoc and conversational search datasets show that QPP-GenRE achieves state-of-the-art QPP quality.

Large Language Models (LLMs) have made query reformulation ubiquitous in modern retrieval and Retrieval-Augmented Generation (RAG) pipelines, enabling the generation of multiple semantically equivalent query variants. However, executing the full pipeline for every reformulation is computationally expensive, motivating selective execution: can we identify the best query variant before incurring downstream retrieval and generation costs?We investigate Query Performance Prediction (QPP) as a mechanism for variant selection across ad-hoc retrieval, and end-to-end RAG. Unlike traditional QPP, which estimates query difficulty across topics, we study intra-topic discrimination—selecting the optimal reformulation among competing variants of the same information need.Through large-scale experiments on TREC-RAG using both sparse and dense retrievers, we evaluate pre- and post-retrieval predictors under correlation- and decision-based metrics. Our results reveal a systematic divergence between retrieval and generation objectives: variants that maximize ranking metrics such as nDCG often fail to produce the best generated answers, exposing a “utility gap” between retrieval relevance and generation fidelity.Nevertheless, QPP can reliably identify variants that improve end-to-end quality over the original query. Notably, lightweight pre-retrieval predictors frequently match or outperform more expensive post-retrieval methods, offering a latency-efficient approach to robust RAG.

Modern information retrieval systems typically operationalise relevance as a query-conditional score computed at inference time. This design choice has become dominant such that alternative decompositions of relevance are rarely discussed, despite the long history of document and query priors in probabilistic retrieval and large-scale search. As neural ranking models grow more computationally expensive and retrieval pipelines expand to include multi-stage ranking, recommendation, and retrieval-augmented generation, this monolithic view of query-time scoring becomes increasingly limiting. We argue that modern information access systems are more naturally understood as performing approximate posterior inference, in which relevance is refined through a staged combination of query-dependent likelihoods and query-independent priors. We extend classical probabilistic retrieval formalisms to contemporary learned systems and show how explicit likelihood–prior decomposition exposes new opportunities to shift computation offline while disentangling document-level and interaction-level beliefs. We present empirical evidence that incorporating query-independent document utility can complement existing rankers and improve effectiveness with minimal query-time computation (solely score fusion). Concretely, a learned prior improves first-stage retrieval through rank fusion (up to ? nDCG@10 ? 0.046 on TREC DL-2019 and ? 0.029 on TREC DL-2020) and also improves downstream re-ranking, with the largest gains observed for the LLM re-ranker RankZephyr (up to ? nDCG@10 ? 0.054 on TREC DL-2020). Finally, we discuss how this decomposition connects to broader information access and outline research directions for designing retrieval systems that explicitly allocate modelling capacity between offline priors and online interaction.

Large Language Model (LLM)-based agent simulation has emerged as a promising approach to meet the increasing demand for real-time and rigorous evaluation in modern recommender systems. A typical LLM-driven simulation framework comprises three essential components: the profile module, memory module, and action module. However, existing studies have primarily concentrated on enhancing the memory and action modules, with limited attention to profile generation, which plays a pivotal role in ensuring realistic agent behaviours and aligning simulated interactions with real user dynamics. Moreover, the scarcity of datasets specifically designed for recommendation simulations has led to heavy reliance on manually crafted profiles, significantly limiting the scalability and generalisability of simulation frameworks across different datasets. To address these challenges, this work proposes an Automated Profile Generation Framework for Recommendation Simulation, APG4RecSim, that constructs realistic, coherent, and robust user profiles with minimal supervision. Extensive experiments on three benchmark datasets demonstrate that APG4RecSim achieves the best overall performance on discrimination, ranking, and rating tasks, improving ranking quality by up to 7% in nDCG@10 and reducing rating distribution divergence by 8% in JSD compared to existing profile-generation baselines. Beyond overall performance gains, our results show that APG4RecSim produces profiles that are resilient to popularity- and position-induced biases and maintain stable performance across datasets and different LLMs.

User interactions with rankings are affected by both items' relevances and display positions. Accordingly, click probabilities are often modeled as a product of relevance and position factors; and for improving recommendation and search, one needs to disentangle relevance from position bias. However, existing click models only provide frequentist point-estimates that do not capture any measure of epistemic uncertainty. Consequently, there is no indication of how much confidence one should have in their predictions. In this work, we introduce the first evidential deep-learning approach to form an epistemic alternative to the important position-based click model. Our learned model takes as input item and position features and outputs a beta-distribution for every relevance and position-bias variable of the position-based model. These distributions capture epistemic uncertainty about click probabilities and the underlying effects of attraction and position bias. The main challenge of our approach is its optimization for which we propose approximation and conditioning techniques to provide numerical stability and variance reduction. Our experiments indicate that our approach captures epistemic uncertainty in predictions on previously-unseen data, whereas standard policy gradients fail to learn meaningful distributions. We believe our contribution of the first contextual epistemic click model constitutes an important step in incorporating Bayesian uncertainty into click modeling.

Cognitive diagnosis aims to infer students' mastery levels over knowledge components from their learning interactions, supporting personalized education applications. However, existing models encode responses as binary correctness labels, discarding information about which specific option a student selected. This input-level information loss limits their ability to distinguish qualitatively different error types. As a result, providing interpretable diagnostic outputs becomes challenging. To address these limitations, we propose SACD, a semantic-augmented cognitive diagnosis framework that integrates LLM-based semantic analysis with student behavioral modeling. SACD comprises the following key components. First, an LLM-based exercise diagnostic generator analyzes exercise content and produces structured semantic annotations for each answer choice, capturing the specific misconceptions each option represents. Second, a kernel-based alignment mechanism projects semantic embeddings and behavioral representations into a unified kernel space, enabling effective fusion of heterogeneous information. Third, an interpretable diagnosis layer predicts student performance and generates fine-grained mastery estimates, which LLMs further process to produce actionable learning plans. Extensive experiments on three real-world datasets demonstrate that SACD achieves superior prediction accuracy while enabling interpretable, actionable diagnostics.

Many Namibian schools face recurring connectivity and infrastructure constraints, yet ICT-in-education initiatives often assume always-on internet and cloud-based platforms for STEM teaching and learning. This paper asks what changes when the school, rather than the cloud, is treated as the primary locus of information retrieval. We describe the design of an offline-first IR prototype that hosts a curriculum-aligned STEM repository locally and supports keyword search, BM25 ranking, and curriculum-aware facets. Framing the system explicitly as IR raises questions that current offline educational platforms tend to leave under-specified: how should ranking and faceted search be designed when bandwidth, devices, storage, and power are scarce; how might teachers' and learners' retrieval behaviour change when queries remain inside the school network; and how can such systems be evaluated when large-scale logging and A/B testing are unlikely to be feasible in many classrooms? We outline the system architecture and a mixed-methods evaluation plan combining local interaction logs, teacher-rated relevance judgements, and qualitative classroom evidence.

Web-augmented large language models (LLMs) promise end-to-end retrieval agents for time-sensitive information access, yet their reliability in low-resource environments remains unclear. We introduce MarathiWeb, a benchmark for evaluating LLM-native web retrieval in Marathi that separates translated global queries from natively authored, local information needs. We evaluate two frontier LLMs with built-in web search against a controlled Google-based retrieval baseline, and analyse retrieval depth, failure modes, and cost-accuracy trade-offs. Despite high tool invocation rates, accuracy drops sharply in Marathi, most severely for local queries, with the dominant failures arising from evidence integration rather than missing sources. We further find that shallow retrieval (one web call) is consistently the most accurate and cost-efficient strategy, while deeper multi-step search degrades performance. MarathiWeb exposes persistent gaps in local-language information access and provides an open benchmark for low-resource retrieval research.

Machine translation for low-resource languages suffers from domain-imbalanced corpora, causing quality degradation on technical text. However, in-context learning opens the possibility to rely on limited in-domain corpora to inform translation. We present lessons learned from Tulun, a retrieval-augmented system combining neural MT with LLM post-editing, guided by user-configurable translation memories and glossaries. Deployed for medical translation in Timor-Leste (Tetun) and disaster relief translation in Vanuatu (Bislama), the system achieves accuracy improvements over baseline MT by 16.90-22.41 ChrF++ points, while offering rapid adaptability and transparency to end-users. Key recommendations include: domain granularity matters more than broad categories; translation target audience should inform retrieval; and RAG-augmented MT is most effective for languages that lack domain corpora but remain within LLM pretraining distributions.

Large Vision--Language Models (LVLMs) frequently fail on knowledge-intensive educational video QA despite the presence of requisite visual evidence. Through region-level analysis, we identify a systematic evidence-suppression pattern: task-critical tokens (e.g., diagrams) exhibit lower representational energy than distractors at the encoder output, rendering them prone to persistent attention neglect during decoding. While a controlled study shows that performance is sensitive to the strength and amount of injected candidate evidence signals, we find that rigid heuristics are insufficient due to sensitivity to token quality. To address this, we propose Bi-level Active Evidence Injection (BAEI), a decoding-time intervention that keeps the LVLM backbone frozen. BAEI employs a lightweight Injection Policy Network (IPN), optimized via GRPO, to dynamically select candidate evidence tokens and predict their token-wise injection strengths. The framework operates on two levels: increasing the contribution of candidate evidence-related signals in shallow layers and performing adaptive correction in deep layers triggered by predictive entropy. Experiments on educational benchmarks demonstrate consistent gains, validating decoding-time evidence intervention as an effective solution for factual alignment. The code is available at https://github.com/diojojolc-cell/BAEI.