Skip to content

🔍 Information Retrieval & RAG

💬 EMNLP2026 · 72 paper notes

📌 Same area in other venues: 🧠 NeurIPS2026 (18) · 🎞️ ECCV2026 (8) · 🔬 ICLR2026 (81) · 💬 ACL2026 (73) · 🧪 ICML2026 (26) · 🤖 AAAI2026 (21)

🔥 Top topics: RAG ×16 · Reasoning ×10 · LLM ×6 · Multimodal/VLM ×5 · Adversarial Robustness ×3

A Wolf in Sheep’s Clothing: Targeted Routing Hijacking in Federated RAG

This paper uncovers Routing Hijacking, a pre-retrieval vulnerability in Federated RAG where malicious clients forge semantic profiles to seize target queries without relevant underlying data, and proposes Trust-Aware Secure Routing (TASR) to reweight clients based on returned-evidence behavioral feedback.

HARVE: Hacking-Aware Reward-Head Vector Editing for Robust Reward Models

This paper introduces HARVE, a backpropagation-free vector editing method for scalar reward models that identifies a multi-directional hacking subspace from fooled contrastive pairs and projects out the sensitive reward-head component without fine-tuning, dramatically improving robustness in high-stakes professional and general domains.

Training-free LLM Verification via Recycling Few-shot Examples

ReFeri introduces a training-free framework that recycles few-shot examples for both generation and post-hoc verification, combining forward likelihood confidence with a leave-one-out backward reconstruction penalty to suppress demonstration-specific artifacts and achieve an 8.2% relative gain across seven diverse reasoning benchmarks.

AfrIFact: Cultural Information Retrieval, Evidence Extraction And Fact Checking for African Languages

AfrIFact introduces the first large-scale automated fact-checking benchmark for 10 native African languages and English (18,256 claims across healthcare and cultural news domains), demonstrating severe cross-lingual retrieval degradation in embedding models and revealing that LLMs require few-shot demonstration or parameter-efficient fine-tuning (QLoRA) to effectively leverage supporting evidence for veracity classification.

AnInter: Anti-Interference Training for Long Context Retrieval

Targeting the fundamental failure mode where local evidence in long documents is overwhelmed by distracting noise, AnInter introduces an efficient short-text anti-interference training framework that constructs noisy multi-positives via circular sliding concatenation and trains models under Multi-Positive InfoNCE, delivering a +16.4 point average boost over backbones on the challenging LongEmbed benchmark.

Argument Collapse: LLMs Flatten Long-Form Public Debate

Evaluating 23,381 essays generated by five frontier LLMs against 1,487 human essays across 256 public debates, this paper uncovers systemic "argument collapse" at the levels of main claims, supporting reasons, and rhetorical arcs, while revealing that standard preference evaluators actively penalize novel viewpoints and reinforce this homogenization.

Argus-Retriever: Vision-LLM Late-Interaction Retrieval with Region-Aware Query-Conditioned MoE for Visual Document Retrieval

Addressing the query-agnostic limitation of standard vision-language late-interaction retrievers, Argus incorporates a region-aware Mixture-of-Experts module conditioned on query context and spatial coordinates, producing adaptive document representations while maintaining ColPali-compatible multi-vector indexing and achieving 86.0 NDCG@5 on ViDoRe V1+V2.

Authority Bias in Conversational Search Engines for Academic Paper Recommendation

This paper introduces the first content-controlled counterfactual causal audit framework for conversational academic paper recommendation in LLMs, demonstrating that holding title and abstract constant while manipulating authority metadata causes 39.2% of recommendations to flip, with venue prestige acting as the dominant driver, frontier models exhibiting a prompt-level backfire, and an 11.1pp say-do gap obscuring behavioral bias.

Axis-LoRA: Probing and Adapting Embedding Models for Personal-Memory Retrieval

Addressing the severe under-exposure of temporal, spatial, entity, and event attributes in general-purpose dense text embeddings, Axis-LoRA introduces four axis-conditional LoRA branches inside the late layers of a frozen encoder with query-conditioned sparsemax gating, achieving consistent retrieval gains on long-term personal memory benchmarks with minimal parameter overhead.

Benchmarking Aggregation Queries over Unstructured Text

Addressing the severe undercounting flaw of existing RAG and agent systems that prioritize partial evidence rather than exhaustive entity coverage on unstructured corpora, this paper formalizes corpus-bounded entity aggregation querying, constructs AGGBench spanning 498 single- and multi-constraint queries across finance and research domains with verifiable evidence, and evaluates 15 representative systems to uncover fundamental bottlenecks in evidence sparsity and completeness.

BenchPreS: A Benchmark for Context-Aware Personalized Preference Selectivity of Persistent-Memory LLMs

BenchPreS is the first benchmark assessing the context-aware preference selectivity of persistent-memory LLMs across formal communication scenarios, revealing that frontier models indiscriminately treat stored user preferences as globally enforceable commands—achieving misapplication rates of 40% to 86% across institutional communication tasks and demonstrating a fundamental lack of context-sensitive normative suppression.

Beyond Bag-of-Words: Generative Structural Regularization for Dense Table Retrieval

To alleviate the severe bag-of-words degradation where contrastive table retrievers rely on superficial lexical overlap rather than row-column bindings, this paper establishes a BoW-equivalent adversarial benchmark via Homologous Structural Swapping (HSS) and introduces GSR-Table, a dual-encoder framework that leverages orthogonal attention pooling and masked generative reconstruction to enforce structure-aware dense representations.

Beyond Output-Level Correction: Tracing Positional Bias to Early-Layer Commitment in Transformer Multiple-Choice Evaluation

Using permutation-invariant all-pairs causal activation patching across three model families and benchmarks, this work traces multiple-choice positional bias to the residual stream before normalized depth 0.60, demonstrating that the bias originates from early MLP content encoding followed by mid-layer attention-driven label commitment rather than shallow output-level distortion.

Beyond Retrieval: Structuring Evolving and Inconsistent External Knowledge with Proposition Relations for RAG

To tackle evolving external knowledge bases contaminated with outdated statements and factual conflicts, ProStruct-RAG shifts the burden of knowledge integration from inference-time reasoning to pre-retrieval structuring: it extracts timestamped propositions, mines directed updates, contradicts, and supports relations to build a global proposition graph, and retrieves relation-aware subgraphs via one-hop neighborhood expansion, substantially outperforming baselines that do not model inter-proposition relations.

Beyond the Vacuum: Combinatorial Strategy Selection for Competitor-Aware Generative Engine Optimization

Addressing the severe diminishing returns of isolated generative engine optimization under competitor adoption, this paper proposes a two-phase competitor-aware framework that pairs Bayesian Optimization of Combinatorial Structures (BOCS) with teacher-reasoning-grounded DPO fine-tuning, training a lightweight 2B selector model that achieves state-of-the-art citation visibility and shows the slowest performance degradation in saturated competitive environments.

CamoDocs: A Poisoning Attack Against Retrieval-Augmented Language Models Using Camouflaged Documents

CamoDocs proposes a query-free poisoning attack against retrieval-augmented language models that crafts stealthy adversarial documents by blending synthesized benign and adversarial chunks, dispersing document embeddings via gradient-guided token manipulation away from their centroid, and bounding fluency loss via perplexity re-ranking to evade seven state-of-the-art defenses.

Chiaroscuro for Emotions: A Contrastive Emotion Benchmark for Multimodal Language Models

This paper presents CHIARO, a 1,000-sentence benchmark grounded in appraisal theory for two-agent contrastive emotion inference from single shared events without explicit affect words, exposing a 25.7-point macro-F1 deficit in frontier LLMs relative to human agreement and establishing its utility as an orthogonal training signal that enhances emotion classifiers across six of ten external benchmarks.

Concise and Sufficient Sub-Sentence Citations for Retrieval-Augmented Generation

Addressing the cognitive burden of verbose sentence-level citations and incomplete keyword-level attributions in RAG systems, this paper formulates concise and sufficient sub-sentence citation generation, constructs a double-annotated benchmark, and proposes an evaluator-filtered synthetic data augmentation framework with two-stage fine-tuning that boosts citation F1 to 80.46 while significantly cutting user verification time.

CORE-T: COherent REtrieval of Tables for Text-to-SQL

To tackle distractor proliferation and broken join paths in open-book multi-table text-to-SQL, CORE-T introduces offline table purpose enrichment with key-constraint compatibility caching, combined with single-shot LLM coherent subset selection and two-step additive adjustment, boosting multi-table execution accuracy while drastically pruning candidate tables and LLM token overhead.

CoRect: Context-Aware Logit Contrast for Hidden State Rectification to Resolve Knowledge Conflicts

To tackle knowledge conflicts where a language model prioritizes its parametric memory over contradictory retrieved evidence, CoRect introduces a training-free inference-time framework that contrastively selects evidence-supported token directions without gold labels and applies a minimal additive patch to suppressive FFN outputs in the residual stream, boosting factual alignment and preserving generation fluency.

Doc-REFRAG: Rethinking Multimodal Document Retrieval-Augmented Generation

To resolve visual token explosion and retrieval noise in multi-image multimodal document RAG, Doc-REFRAG introduces coarse-level chunk compression paired with a lightweight RL-based selector that selectively expands question-critical regions, supported by DocLongRAG with hard-negative mining, achieving state-of-the-art accuracy with markedly reduced TTFT and TTIT.

Dual-View Training for Instruction-Following Information Retrieval

Dual-View Training introduces a data synthesis strategy based on polarity reversal: given a query and a document pair where one is relevant and the other is an instruction-excluded negative, an LLM generates a complementary instruction that completely swaps their relevance labels, boosting FollowIR instruction sensitivity (p-MRR) by +57.5% across both 305M and 560M encoders without degrading general retrieval effectiveness.

DuoPRAG: Rank-Disentangled and Gated Knowledge-Reasoning Fusion

DuoPRAG disentangles retrieved factual knowledge from multi-step reasoning into two distinct LoRA adapters combined via static rank masks and dynamic token-level gates, replacing brittle lexical CoT prompts with continuous latent reasoning tokens to achieve state-of-the-art multi-hop QA performance in a query-only regime.

EARAL: Efficient Attribution of Retrieval-Augmented Answer Generation using Logits

EARAL prefixes retrieved documents with special identifier tokens and jointly fine-tunes LLMs to calibrate identifier logits alongside answer generation, enabling faithful, generation-time document attribution directly from pre-softmax token logits with zero extra inference passes and a 104x speedup over ablation-based baselines.

Efficient Iterative Retrieval with Heterogeneous Batching

Addressing the severe GPU underutilization and computational bubbles caused by alternating embedding (compute-bound) and generation (memory-bound) workloads in iterative RAG, ORTHRUS enables heterogeneous batching within a unified inference loop through chunked embedding, incremental pooling, and workload-aware intra-batch scheduling.

Efficient Temporal-aware Matryoshka Adaptation for Temporal Information Retrieval

This paper proposes Temporal-aware Matryoshka Representation Learning (TMRL), an efficient adaptation framework that explicitly allocates a compact prefix temporal subspace within Matryoshka embeddings, pairing it with self-distillation and temporal contrastive data augmentation to achieve high-accuracy, zero-overhead temporal retrieval and flexible funnel retrieval in temporal RAG.

Escaping the Ambiguity Trap: Content-Level Evidence Tracing for Open-Domain Table QA

To overcome the "ambiguity trap" where partitioned table segments share identical metadata and schemas, this paper proposes Probe-Grounded Reasoning (PGR), a coarse-to-fine framework that performs deterministic content-level evidence tracing via query-specific probes and generates verified answers through a decoupled Semantic-Plan-Execute (SPE) workflow aligned with GRPO.

eTracer: Polarity-Aware Evidence Grounding

This paper formalizes the task of Polarity-Aware Evidence Grounding and introduces ETRACER-BENCH, the first human-annotated benchmark spanning 7,905 sentence-level evidence pairs with three-way polarity labels, alongside ETRACER, an encoder-based post-hoc framework that jointly recovers fine-grained evidence localization and signed polarity in a single forward pass under full self-attention, matching a 235B LLM at 1/600 the parameter scale.

EviRank: Structured Relevance Evidence for Multimodal Image Re-ranking

Addressing the opacity of dense embeddings and the omission/hallucination vulnerabilities of free-form chain-of-thought reasoning, EviRank parses multimodal queries into a structured evidence package of required, forbidden, and ignorable criteria across six semantic slots, achieving state-of-the-art training-free re-ranking and enabling effective distillation into an MLLM-free lightweight student.

Finding What Matters: Anchoring Context Knowledge with Evolving Indices for Iterative Retrieval

Addressing the challenge where iterative retrieval accumulates context noise and hinders multi-hop evidence synthesis, KAIR maintains an in-context evolving index of keywords and RDF relation triples that dynamically anchors salient facts, acts as a cognitive navigational guide for the LLM, and achieves substantial multi-hop QA gains while preserving raw documents.

From Gaps to Active Retrieval: Open-Domain Iterative Timeline Completion via Historical Trajectory Analogy and Dynamic Ontology

Addressing the failure of passive semantic retrieval to recover logically necessary milestones and fine-grained facts in open-domain timeline summarization, this paper presents Active Timeline Completion (ATC), an iterative framework that transforms causal and element gaps into targeted retrieval questions via parametric detection, historical trajectory analogies, and dynamic ontologies, verified by a logical utility filter.

From Sufficiency to Reflection: Reinforcement-Guided Thinking Quality in Retrieval-Augmented Reasoning for LLMs

Addressing the failure patterns of outcome-only reinforcement learning in retrieval-augmented reasoning—specifically premature search termination, flawed intermediate logic, and answer-reasoning mismatches—this paper proposes TIRESRAG-R1, a unified framework combining a think-retrieve-reflect paradigm with multidimensional process rewards (sufficiency, thinking quality, and reflection), achieving consistent state-of-the-art gains on multi-hop benchmarks while preventing late-stage training collapse.

Gated Residual Discrepancy for Efficient Bi-Encoder Factuality Verification

To overcome symmetric similarity saturation and contextual noise vulnerability in bi-encoders for factuality verification, GRD introduces a multi-slot dual-view residual discrepancy operator with adaptive gating amplification, matching cross-encoder accuracy (79.99% Macro-F1) while delivering a 23.55× inference speedup.

hievi rag hierarchical evidence driven reasoning for long document understanding

To resolve answer-void distractor interference and catastrophic omission in single-pass long document multimodal QA, HiEvi-RAG establishes a cooperative four-stage framework featuring hierarchical question decomposition, GRPO-trained EviAgent cross-page evidence verification with soft reranking, and explicit memory-guided sliding-window iterative reasoning, outperforming open-source baselines by an average of 8.05% in accuracy across four long-document benchmarks.

Hyperspherical Semantic Expansion via von Mises–Fisher Mixtures for Zero-Shot Dense Retrieval

Addressing the norm shrinkage defect induced by Euclidean Gaussian Mixture Models on the unit hypersphere, this paper proposes vMF-HSE, a training-free hyperspherical semantic expansion framework based on von Mises–Fisher mixtures that decouples semantic direction from variance, combined with an adaptive \(\kappa\)-Stop mechanism based on marginal information gain convergence to significantly boost zero-shot dense retrieval on BEIR.

Learning Diagnostic Clarifying Questions for Conversational Question Answering via Data-Efficient Dynamic Distillation

To tackle the lack of progressive disambiguation protocols and the prevalence of non-diagnostic questions in open-domain conversational QA, D4CQ introduces procedural scaffolding with contrastive intent-convergence alignment, diagnosticity-based dynamic distillation, and hindsight relabeling with compatibility gating, achieving superior ambiguity resolution and answer accuracy under limited interaction budgets.

MEMORA: A Memory-Enhanced Multimodal Committee Reranking Agent for Reasoning-Intensive Retrieval

Addressing the paradox where visual inputs act as noise and monolithic rerankers fail under complex inference, MEMORA establishes an agentic retrieval framework combining factor-committee reranking, cross-query episodic memory, and a CoT-fine-tuned dense retriever to achieve state-of-the-art results on MM-BRIGHT.

MetaRAG: Belief-Action Aligned Policy Optimization for Agentic RAG

MetaRAG frames the SEARCH/ANSWER decision boundary in agentic RAG as belief-action alignment, employing verify-first action generation alongside lightweight internal belief probing with a correctness-gated consistency reward to significantly improve accuracy-efficiency trade-offs without requiring external LLM judges.

MIMO: Multilingual Information Retrieval via Monolingual Objectives

MIMO is a two-stage training framework for Multilingual Information Retrieval (MLIR) that leverages a high-performing teacher model's English semantic space as an absolute anchor, warming up cross-lingual alignment via distillation and jointly optimizing distillation with contrastive learning to break the alignment-uniformity trade-off and substantially outperform prior baselines.

Multi-Source Retrieval and Reasoning for Legal Sentencing Prediction

Addressing the dual bottlenecks of fragmented objective regulatory criteria and vulnerable discretionary subjective reasoning in legal sentencing prediction, this paper proposes MSR2—a reinforcement learning framework that seamlessly interleaves multi-source legal retrieval into the LLM action space and guides intermediate sentencing factor extraction via a rubric-based process-level reward under GRPO, setting new state-of-the-art accuracy and interpretability benchmarks on CAIL2018 and CJO22.

Muon^2: Boosting Muon via Adaptive Second-Moment Preconditioning

Muon^2 introduces Adam-style adaptive second-moment element-wise preconditioning prior to the Newton-Schulz polar orthogonalization step in Muon (together with a memory-efficient factorized variant Muon^2-F), eliminating singular value dead zones to reduce Newton-Schulz iterations by 40% while speeding up pre-training and boosting downstream task performance.

OBLIQ-IR: Training a Dense Retriever for Oblique Queries

Addressing the severe first-stage retrieval failure on "oblique queries" whose relevance hinges on latent document attributes with minimal surface overlap, OBLIQ-IR fine-tunes a 3B single-vector bi-encoder using mechanism-lensed synthetic queries alongside score-free \(k\text{NN}\)-graph distillation from a frozen authorship specialist, consistently outperforming frontier models like Gemini-2-Embedding and the GPT-5.2 Multi-Hop Agent across OBLIQ-Bench tasks.

OpenResearcher: A Fully Open Pipeline for Long-Horizon Deep Research Trajectory Synthesis

Addressing the excessive API costs, environment instability, and lack of reproducibility in training deep research agents via live commercial search APIs, OpenResearcher decouples one-time online corpus bootstrapping from multi-turn offline trajectory synthesis, using three explicit browser primitives (search, open, and find) over a 15M-document corpus to synthesize 97K+ long-horizon trajectories and boosting a 30B student model to 54.8% on BrowseComp-Plus.

oscar optimization steered agentic planning for composed image retrieval

OSCAR reframes agentic composed image retrieval (CIR) from heuristic trial-and-error into a two-stage mixed-integer programming (MIP) trajectory optimization problem, building a golden trajectory library of positive unions and conservative negative intersections offline to steer a vision-language model (VLM) planner via in-context demonstrations during single-pass online inference, outperforming prior SOTAs using only 10% of training data.

PromptEmbedder: Efficient and Transferable Text Embedding via Dual-LLM Soft Prompting

PromptEmbedder decouples embedding task representations from specific backbone parameters using a dual-LLM architecture, where a compact Prompting LLM differentiably generates instruction-aware soft prompts for a frozen Embedding LLM via continuous vocabulary relaxation, approaching LoRA fine-tuning quality while enabling seamless cross-architecture transfer via a single-layer linear adapter.

RAGSentinel: Certifiable Geometric Consensus for Robust Retrieval-Augmented Generation

RAGSentinel proposes a training-free, black-box-compatible post-retrieval defense against RAG knowledge poisoning by measuring document-induced hidden-state shifts with an independent surrogate encoder, removing shared query-topic directions, and filtering malicious documents as geometric outliers relative to a robust geometric median consensus with certifiable guarantees.

ReBOL: Retrieval via Bayesian Optimization with Batched LLM Relevance Observations and Query Reformulation

ReBOL reformulates the standard reformulate-retrieve-rerank pipeline into a Bayesian optimization framework, using LLM query reformulations to seed a multimodal Gaussian process posterior and actively acquiring diversified document batches for LLM relevance scoring to overcome vector retrieval recall bottlenecks.

RedNote-Vibe: A Lifestyle Social Media Dataset for Detecting AI-Generated Text and Analyzing Its Impact

To address the challenges of detecting high-fidelity AI-generated text and auditing long-term engagement dynamics on lifestyle social platforms, this paper constructs RedNote-Vibe—the first five-year (2020.10–2025.07), 148k-note benchmark from RedNote (Xiaohongshu)—and develops the PsychoLinguistic AIGT Detection Framework (PLAD), an interpretable system based on 31 cognitive-linguistic features that reveals how AI and human authorship diverge in driving audience engagement.

ReGround: Grounding Reviewer Comments in Multimodal Evidence

Addressing the lack of fine-grained multimodal evidence alignment benchmarks in peer review, ReGround leverages explicit author rebuttal references to construct a large-scale retrieval benchmark across 3,656 papers and 16,274 comment–evidence pairs, exposing evidence-type routing bottlenecks and substantial text-visual complementarity.

REPAIR: Resolving Long-Tail Confusion in Scientific Retrievers via Fact-Verified Iterative Refinement

Addressing structural hallucinations and catastrophic semantic collapse caused by heavy-tailed entities and extreme fact-sensitivity in scientific corpora, REPAIR introduces a three-stage self-evolving framework—diagnosing low-margin confused queries, expanding fact-verified evidence via external scientific APIs, and differentiating model-aware hard negatives—outperforming 19 competitive baselines across nine materials science and biomedical benchmarks using only 4M training pairs.

Retrieval-Augmented Generation with Retrieval-optimized TRA Structure

Addressing the contextual loss and structural rigidity of triplet graphs in Graph-based RAG, this paper proposes an entity-centric hierarchical Title-Relations-Attributes (TRA) database structure and atomic Fact-level indexing, coupled with dependency-aware query decomposition and reciprocal rank fusion, achieving comprehensive state-of-the-art performance across simple and multi-hop QA benchmarks.

Retrieving Facts or Reinforcing Beliefs? Towards Understanding Confirmation Bias in RAG Systems for Scientific Claim Verification

Presents CoBiT, the first systematic benchmark comprising 707 minimally perturbed claim–counterclaim pairs to evaluate confirmation bias in scientific RAG systems, introducing the RBOH/RBM ranking bias metrics and generative consistency measures to uncover multi-stage bias amplification across retrieval, re-ranking, and generation.

Retrieving Facts or Reinforcing Beliefs? Towards Understanding Confirmation Bias in RAG Systems for Scientific Claim Verification

Addressing the insidious tendency of retrieval-augmented generation (RAG) systems to pander to user query stances by preferentially retrieving and generating confirmatory evidence, this paper introduces CoBiT—a benchmark of 707 minimally perturbed claim-counterclaim pairs—and proposes stage-specific quantitative metrics (RBM, CBR) that uncover how confirmation bias emerges and amplifies across re-ranking and generation.

SAILRec: Steering LLM Attention to Dual-Side Semantically Aligned Collaborative Embeddings for Recommendation

SAILRec overcomes semantic incompatibility and layer-wise attention interference in collaborative LLM recommenders by integrating codebook-guided dual-side semantic alignment with shallow-suppressed, deep-enhanced hierarchical attention steering.

Sampling Less, Knowing More: Contradiction-aware Adaptive Spectral and Entropic Uncertainty Quantification for LLMs

To tackle computational redundancy from fixed sampling budgets and semantic collapse under consistent errors in multi-sample uncertainty quantification, this paper presents CASE, which builds a geometrically symmetrized, sparsified contradiction graph and evaluates scale-invariant normalized spectral norms for adaptive early stopping, paired with low-temperature token predictive entropy to cut sampling by over 30% while achieving superior AUROC and AUARC across factual QA benchmarks.

SCALE: Structured Case-law-guided Anticipatory Legal Reasoning

Addressing standard legal RAG's inability to synthesize dense, heterogeneous case law and its reliance on surface factual similarity, SCALE transforms ECHR Case Law Guides into an explicit Issue–Factor Hierarchy, anticipates the legal issues and factual factors implicated by a case, and retrieves specific authority passages to generate faithful, precedent-grounded judicial reasoning.

SemBridge: Language Transfer in Sparse Encoders via Multilingual Semantic Bridge

SemBridge proposes a novel embedding initialization framework for transferring English-centric sparse retrieval encoders to diverse target languages by leveraging multilingual dense models as semantic bridges and Entmax-based sparse truncation, reconstructing new target tokens as precise linear combinations of core source synonyms to achieve superior zero-shot retrieval and faster fine-tuning convergence.

SemFlowRAG: Directed Semantic Flow from Abstraction to Evidence for Complex Reasoning

Addressing the "probability black hole" and semantic drift caused by high-degree abstract concept nodes in undirected Graph RAG topologies, SemFlowRAG constructs a corpus-adaptive semantic gradient graph via passage embedding covariance and guides retrieval from abstract concepts down to concrete evidence via a smooth-penalized directed Personalized PageRank algorithm.

Small Models, Fewer Heads, More Forgetting: Sparse Attention Fine-Tuning for SLM Continual Learning

Addressing the structural bottleneck where Small Language Models (SLMs) suffer severe catastrophic forgetting during continual learning due to limited attention heads, this paper introduces LightLoRA—a parameter-free sparse attention fine-tuning framework that quantifies parameter updates via SVD and penalizes historical gradient directional alignment to update only a sparse subset of high-contribution, low-conflict attention heads.

SmallReason-ColBERT: An Ultra-Small Late-Interaction Retriever for Reasoning-Intensive Retrieval

Addressing the severe performance degradation of edge-scale late-interaction retrievers on reasoning-intensive retrieval tasks, this paper introduces SmallReason-ColBERT—a 32M retriever that bridges the gap to 150M reasoning-tuned models on the BRIGHT benchmark via projection widening, two-stage reasoning fine-tuning, and a minimal 129-parameter query-token importance head trained under an asymmetric un-normalised MaxSim objective.

SpeechEditBench: A Bilingual Multi-Attribute Benchmark for Instruction-Guided Speech Editing

SpeechEditBench establishes the first bilingual benchmark covering seven atomic tasks and multi-intent compositional edits for instruction-guided speech editing, introducing an anchor-based dual-constraint evaluation framework that decouples target attribute manipulation from non-target content preservation.

STEB: Style Text Embedding Benchmark

Introduces the Style Text Embedding Benchmark (STEB) across 96 datasets in 7 languages and 5 task types, benchmarking 40 representation models to reveal that general-purpose semantic embeddings consistently fail on stylistic tasks and no single style model dominates universally.

Structural Anchor Pruning: Training-Free Multi-Vector Compression for Visual Document Retrieval

Challenging the prevailing consensus that high-ratio visual token compression requires query-dependent fine-tuning, this paper uncovers an Alignment-Aggregation Divergence in deep vision-language retrievers and introduces Structural Anchor Pruning (SAP), a self-calibrating, training-free, query-agnostic index-time framework that retains 93%–96% of NDCG@5 on ViDoRe while pruning 90% of visual tokens.

tempo realistic multi domain benchmark for temporal reasoning intensive retrieva

Addressing the limitations where existing temporal QA benchmarks focus on simple fact-seeking and reasoning-intensive retrieval lacks temporal grounding, TEMPO introduces the first benchmark for temporal reasoning-intensive retrieval across 13 domains, featuring 1,730 complex queries, 3,976 step-wise retrieval planning steps, and specialized temporal metrics like Temporal Coverage, revealing major bottlenecks in retrieving temporally complete evidence.

The Magnitude Mirage: Rethinking Confidence for Reasoning-Intensive Retrieval

The paper reveals that the standard industry practice of thresholding raw similarity score magnitude for retrieval abstention in RAG systems systematically collapses on reasoning-intensive queries ("Magnitude Mirage"), and demonstrates that zero-cost score-distribution metrics (specifically Score Gap and LSMV) restore confidence calibration without extra inference or retraining, improving downstream answer accuracy by an average of 8.3 percentage points at 10% coverage on BRIGHT.

The RAT: A Unified Bayesian Model for RAG Evaluation

The paper introduces RAT, a unified Bayesian evaluation framework for Retrieval-Augmented Generation (RAG) that factorizes retrieval success, abstention behavior, and answer correctness along the pipeline's information flow, disentangles task success from generator policy adherence, reveals through information theory why retrieval annotations are far more informative than task correctness annotations, and seamlessly integrates calibrated noisy LLM judges.

ToolSpec: Accelerating Tool Calling via Schema-Aware and Retrieval-Augmented Speculative Decoding

Addressing the severe inference latency bottleneck of autoregressive generation in complex multi-turn LLM tool calling, ToolSpec introduces a training-free speculative decoding framework that couples finite-state machine-guided schema drafting with semantic retrieval of historical execution trajectories, achieving 3.5×–4.2× end-to-end lossless speedups across diverse benchmarks.

Towards AI-Assisted Research Writing: Benchmarking Citation Verification and Claim Grounding

This paper introduces the first large-scale benchmark and evaluation framework for Scientific Introduction Generation (SciIG), comprising 3,900 top-tier papers and a 30-metric multi-channel suite that systematically diagnoses modern LLMs across citation correctness, factuality, and rhetorical argumentation.

TRIAD: A Three-Axis Framework for Knowledge Conflict Resolution in RAG

Addressing the vulnerability of retrieval-augmented generation under conflicting and misleading context, TRIAD uncovers three coupled failure axes—question complexity, long misleading evidence, and off-target parametric recall—and resolves them via a training-free framework combining subquestion decomposition, SNR-based evidence calibration, and indirect clue-guided parametric knowledge elicitation.

UltraX: Refining Pre-Training Data at Scale with Adaptive Programmatic Editing

UltraX proposes an adaptive programmatic editing framework for large-scale pre-training data refinement that completes the editing function space by integrating insertion alongside deletion and modification; guided by dataset-adaptive expert prompting and conflict-free post-processing, a 0.6B lightweight model achieves fine-grained instance-level data cleaning at minimal inference cost, consistently surpassing baselines in from-scratch 20B/100B pre-training across five diverse corpora.

Understanding Reliability in LLM-based Human Behavior Simulation

ReliMap decomposes LLM-based human behavior simulation into a three-layer generative framework to audit individual and population reliability across model capacity, profile completeness, and population coverage, revealing that individual accuracy gains decouple from macro-distribution alignment and increasing sample coverage reduces variance but never systematic bias.

VerTox: Verifiable Reward-Guided Corpus Poisoning Against Neural Ranking Models

VerTox is the first framework to formulate corpus poisoning against neural ranking models as a verifiable reward-guided reinforcement learning (RLVR) problem, using GRPO to fine-tune compact LLM generators against normalized rank-distortion rewards, bidirectional NLI factual-corruption rewards, and query-repetition penalties, achieving near-perfect attack success rates across dense, learned sparse, and cross-encoder rankers while severely degrading downstream RAG reliability.