Skip to content

💬 EMNLP2026 Accepted Papers

1620 EMNLP2026 paper notes covering LLM Reasoning (601), LLM Agent (143), Multimodal VLM (131), Safety (129), Model Compression (85), Information Retrieval & RAG (72), LLM (Other) (49), NLP Understanding (49) and other 42 areas. Each note has TL;DR, motivation, method, experiments, highlights, and limitations — 5-minute reads of core ideas.


💡 LLM Reasoning (601)

A Manifold-Aware Topic Modeling Approach via Rank-Based Prototypes

MARETopic introduces a training-free, manifold-aware topic modeling framework that casts topic discovery as greedy prototype document selection over ordinal ranked lists, eliminating high-dimensional embedding distortions while achieving superior clustering purity and topic coherence without a single gradient step.

A Mechanistic Study of AI-Text Detection Neurons in Frozen BERT: Sparse Probing and Activation Patching on RAID

Applying an L1-then-L2 sparse probing protocol across all 9,216 CLS hidden-state dimensions in frozen BERT-base-uncased identifies a stable subset of fewer than 1% of neurons that causally drives AI-text detection (10–16× flip rate over random via activation patching), exhibits high distributed redundancy, and displays a distinct layer-12 concentration unique to instruction-tuned generators.

A Multimodal Automatic Redteaming Evaluation based on Atomic Jailbreak Strategy Decoupling and Combination

This work proposes HACA, an automated multimodal red-teaming framework that decomposes the jailbreak space into structural, semantic, and syntactic layers across text and image modalities into 30 atomic units, orchestrating them via a cross-modal joint planner and a unified generative executor to achieve an average 95.48% attack success rate across five mainstream MLLMs.

A retrieval conditioned rebinding circuit for dynamic entity tracking in large language models

Using causal mediation analysis, path patching, and subspace interventions, this work proves that LLMs do not eagerly maintain a globally updated situation model during dynamic state transitions; instead, they rely on a compact "retrieval-conditioned rebinding circuit" (comprising 3.9%–10.8% of attention heads) that defers binding resolution to query time via dereferencing pointers, with Gemma models exhibiting symmetric Q/K matching and Llama models being predominantly key-dominated.

ACE-Align: Attribute Causal Effect Alignment for Cultural Values under Varying Persona Granularities

ACE-Align addresses the instability and representational distortions of large language models across varying demographic persona granularities by leveraging controlled persona edits to match survey-grounded ordinal cumulative attribute causal effects (\(\Delta\text{CDF}\)) while anchoring absolute modes, substantially outperforming standard RL/SFT baselines and shrinking the Global North–South alignment gap across 14 countries.

ACTD: Anchor-Based Cross-Tokenizer Distillation with Residual Regularization

ACTD tackles vocabulary/sequence misalignment and long-tail projection noise in cross-tokenizer distillation via first-subtoken vocabulary mapping, nearest monotonic boundary alignment, a Top-k anchor loss, and an L1 residual regularization that suppresses non-anchor probability mass, yielding up to +4.5 AVG@8 gains on reasoning benchmarks and unlocking robust multi-teacher fusion.

Adapting from Downturns: Prediction of Long-Term Conversational-Skill Development in Mental-Health Crisis Counselors

This paper introduces the task of predicting early in a conversationalist's career whether they will autonomously improve over time, demonstrating on mental-health crisis counseling data that identifying conversational downturns, matching contextually analogous moments, and learning adaptation summaries via GRPO reinforcement learning successfully forecasts long-term counselor skill development months or years later.

Adaptive Boundaries: Context-Aware Detection for Synthetic Text

PC-CAT establishes a novel two-stage adaptive boundary detection framework that couples reference-free Simple Preference Optimization (SimPO) to enlarge the log-probability gap between human and machine text with a lightweight regressor trained via centering and ordinal losses to predict instance-specific decision thresholds, eliminating the catastrophic collapse of static thresholds under distribution shifts.

CoreMem: Riemannian Retrieval and Fisher-Guided Distillation for Long-Term Memory in Dialogue Agents

Addressing dual bottlenecks of consumer edge hardware (8 GB VRAM) and linear cloud API costs, CoreMem unifies long-term dialogue memory under information geometry via Woodbury-accelerated Fisher-Rao Riemannian retrieval and syntax-protected Fisher-guided token distillation, achieving decisive accuracy gains in open-domain and temporal reasoning.

ShapleyLaw: A Game-Theoretic Approach to Multilingual Scaling Laws

ShapleyLaw is the first game-theoretic multilingual scaling law that treats pretraining languages as players in a cooperative game, quantifies cross-lingual transfer via In-Run Data Shapley values, and enables accurate per-language loss prediction and provably optimal language-level mixture ratio optimization.

Browse all 601 LLM Reasoning papers →


🦾 LLM Agent (143)

A Comprehensive Survey on Linguistic Steganography in the LLM Era

This paper delivers the first comprehensive survey systematically situating linguistic steganography in the LLM era, categorizing 148 steganographic schemes, 60 steganalysis countermeasures, 23 evaluation metrics, and 9 open challenges while delineating five decisive paradigm shifts.

ABDATABENCH: Evaluating AI Agents for Evidence-Grounded Scientific Data Analysis

Introduces ABDATABENCH, the first benchmark assessing AI agents on constructing complete, evidence-grounded antibody records across scientific literature and patents (44 documents, 279 records, 22-field tiered schema with Hungarian alignment and omission penalties), alongside ABCURATOR, a record-centric multi-agent framework achieving 78.3 (proprietary) and 76.0 (open-source) through deterministic tool integration and iterative skill accumulation.

Agent-as-Peer-Debriefer: A Multi-Agent Framework with Perspective-Based Refinement for Qualitative Analysis

This paper proposes Agent-as-Peer-Debriefer, a multi-agent framework that translates the qualitative research credibility practice of peer debriefing into LLM-assisted qualitative coding. By pairing a Hierarchical Coding Agent generating self-reflection memos with three disinterested Peer-Debriefing Agents applying Theory-Driven, Data-Driven, and Applied perspectives across five structural operations, the framework substantially improves code alignment and recall against human expert codebooks.

Agent-G^2: Gaussian Guidance for Agentic Reinforcement Learning

To resolve the structural mismatch of scalar hint depths in long-horizon agent RL with sparse rewards, Agent-G^2 reveals that informative guidance depths form a Gaussian band around the optimum and introduces an online Gaussian schedule that dynamically updates depth distributions using regular policy rollouts with zero probe overhead.

AgentFold: Closed-Loop Agentic Search for Protein Folding Model Design

AgentFold establishes an autonomous closed-loop multi-agent framework that searches over executable code variants on a compact ESMFold substrate via an MCTS-style tree controller, boosting best lDDT by 7.5% over independent Codex proposals under matched evaluation budgets and uncovering three transferable empirical design patterns for geometric deep learning.

AgentJudgeBench: A Multi-Difficulty Benchmark for Evaluating LLM Judges on Agentic Tool-Calling

Addressing the uncalibrated reliance on LLM judges in agentic evaluation, AgentJudgeBench provides a comprehensive benchmark across 6 DAG topologies, 3 difficulty tiers, and 3,808 records, revealing that judge alignment degrades monotonically with task difficulty, plateaus at a 77–82% structural ceiling without ground truth, and suffers from counterproductive over-anchoring in frontier models when exposed to reference traces.

AgentProv: Auditing Agentic LLM API Providers via Tool-use Policy Probes

AgentProv is the first action-based black-box identity auditing framework for agentic LLM APIs, utilizing categorical tool-selection preferences under functionally equivalent probes and a block-decomposed MMD permutation test to catch backbone substitutions without text dependence and with strong robustness against hidden system prompt injections.

AgentSpec: Speculative Decoding for Batch Inference of LLM Agents

Addressing the severe throughput degradation of speculative decoding in large-batch LLM agent serving caused by high token rejection rates and dynamic budget under-utilization, AgentSpec introduces pushdown automaton-guided structure-isolated drafting and redundancy-aware token budget allocation, achieving up to 2.02× end-to-end goodput speedup in vLLM without compromising generation quality.

Aggregate While Searching: Online Partial-Trajectory Aggregation for Long-Horizon Reasoning

Addressing the actionability gap where parallel search rollouts fail to correct reasoning drift before termination, this paper introduces OPTA to periodically synthesize a typed atomic global shared state and selectively inject relevant peer corrections, raising Qwen3-30B accuracy on BrowseComp-Plus from 37.9% to 59.6% with just 2.3% coordination overhead.

AGORA: An Archive-Grounded Benchmark for Agentic Workplace Document Reasoning

AGORA establishes an archive-grounded, cross-domain benchmark for agentic workplace document reasoning containing 9,664 authentic documents (372M tokens) across eight professional domains and 362 deterministic numeric queries, demonstrating that frontier LLM agents struggle to navigate messy corporate archives with even the strongest model reaching only 59.39% accuracy.

Browse all 143 LLM Agent papers →


⚖️ Alignment & RLHF (16)

Aligning Cultural Behaviors in Arabic Language Models via Preferences

To address the lack of country-specific cultural behavioral alignment in Arabic and multilingual LLMs, this paper introduces ARABPREF—the first large-scale pan-Arab preference dataset and native-validated evaluation benchmark covering all 22 Arab League nations across 16 categories and 200 aspects—demonstrating significant gains in cultural appropriateness and robustness via preference optimization.

An Empirical Study on Preference Tuning Generalization and Diversity Under Domain Shift

This paper conducts the first systematic investigation of five preference alignment objectives and various domain adaptation strategies under domain shift across summarization, QA helpfulness, and safety, proving that adaptation strategy dominates alignment objective and revealing that pseudo-labeling achieves high target win rates at the expense of severe linguistic mode collapse.

ARCA: Attribution-Regularized Causal Alignment for Task-Aware Bias Mitigation

ARCA addresses the issue of social bias re-emergence during downstream fine-tuning by automatically extracting task-specific bias-propagating words via differential gradient attribution and jointly optimizing task classification with representation-level causal alignment on dynamically generated counterfactual pairs.

Do Implicit Personalization and Explicit Styles Conflict? PsPLUG: A Lightweight Plug-in for Balancing Personalization and Style in Customized LLMs

To tackle "personalization collapse"—where explicit style instructions override and suppress implicit user habits—PsPLUG models personalization as a distributional residual via a lightweight plug-in prefix and style-conditioned preference optimization, achieving a controllable trade-off between style adherence and personalization strength under a frozen backbone.

DSPA: Dynamic SAE Steering for Data-Efficient Preference Alignment

DSPA introduces a training-free inference-time preference alignment method that featurizes prompts via early-layer SAEs and routes interventions to late-layer SAE latents through a sparse conditional-difference map, editing only token-active features during decoding to achieve competitive alignment under extreme data scarcity without degrading general capability.

Expert Personas Improve LLM Alignment but Damage Accuracy: Bootstrapping Intent-Based Persona Routing with PRISM

This paper systematically demonstrates that expert persona prompts improve instruction-alignment tasks (safety and preference) while consistently degrading pretraining-dependent knowledge retrieval accuracy, and proposes PRISM—a fully bootstrapped, gated LoRA self-distillation framework that captures alignment benefits while preserving knowledge accuracy without external supervision.

From Rollouts to Recipes: Self-Contained Post-Training for LLMs

Addressing the limitation that standard verifier-based post-training applies a single uniform optimization recipe to all samples regardless of learning state, this paper proposes Self-Routing—a self-contained framework that leverages rollout correctness and predictive entropy to dynamically assign samples to GRPO, on-policy self-distillation, regularization, or skipping, achieving consistent gains across mathematical reasoning benchmarks without external teachers or extra sampling overhead.

GAC: Noise-Aware Adaptive Mixing for Hybrid SFT-RL Post-Training

Addressing the vulnerability of fixed mixing schedules in hybrid SFT-RL post-training to shifting gradient noise and objective conflict, GAC derives a closed-form mixing weight by minimizing the mean squared error (MSE) of the mixed stochastic gradient and pairs it with online lightweight proxies and a three-tier stabilization guardrail stack, achieving notable accuracy gains on math and reasoning benchmarks with less than 1% overhead.

LQ-Rewriter: Generative Reward Model-Guided Lazy Query Rewriting for Role-Playing Dialogue

To tackle conversational stagnation caused by ambiguous and low-effort user inputs in role-playing interactions, this paper introduces LQ-Rewriter, a lightweight query rewriter guided by an online regeneration-derived Generative Reward Model (GRM), which leverages narrative-style rewriting strategies and two-stage alignment (RFT + CPO) to substantially enrich downstream role-play responses and boost online conversation turns by +11.1%.

Mechanism Alignment for Socially Embedded Decision Making in LLMs

Addressing the failure of large language models to implement true peer-weighted utility aggregation in multi-agent social simulations, this paper introduces a network-utility benchmark and a theory-supervised mechanism alignment framework (SEDM-alignment) that significantly improves micro-level decision rule recovery and restores faithful macro emergent dynamics in financial and labor markets.

Browse all 16 Alignment & RLHF papers →


🔒 Safety (129)

A Novel Semantic Manifold Alignment Attack against Embedding-to-Embedding Obfuscation in Privacy-Preserving LLMs

By conceptualizing the obfuscated vector stream as an unknown foreign tokenizer-language, Proxy Manifold Alignment (PMA) leverages Gromov-Wasserstein optimal transport and mutual nearest neighbor refinement over Word2Vec-induced proxy semantic manifolds to achieve over 80–99 ROUGE-L plaintext reconstruction without model weights, parallel data, or original embedding values.

Addressing Corpus Knowledge Poisoning Attacks on RAG Using Sparse Attention

Reveals that standard causal cross-document attention in decoder-only LLMs is the primary vulnerability channel for corpus knowledge poisoning in RAG, and proposes Sparse Document Attention RAG (SDAG)—an inference-only block-sparse attention mask that eliminates cross-document contamination, dramatically slashing attack success rates and establishing new SOTA benchmarks.

Affective Context Amplifies Sycophancy in LLM Responses

Grounding evaluation in ingratiation theory, this work shows that LLMs systematically soften or withhold negative evaluations when directly addressing users, an effect significantly amplified by negative affective context (especially loneliness and distress) through evasive sycophancy.

AI Psychosis: Does Conversational AI Amplify Delusions?

By constructing simulated users (SimUsers) matched via Stratified Propensity Score Matching from longitudinal Reddit posting histories and developing a supervised DelusionScore metric, this paper demonstrates that multi-turn LLM interactions progressively amplify delusion-related language by an average of 233% in vulnerable profiles (strongest in reality skepticism and compulsive reasoning), whereas a lightweight state-aware prompt conditioning intervention successfully reverses this amplification feedback loop.

Assessing Adversarial Robustness of Latent Reasoning Models

This work provides the first systematic adversarial robustness assessment of latent reasoning models across textual and vision-language domains, finding that implicit reasoners degrade substantially more than explicit CoT baselines, plagued by brittle latent representations in text and complete functional redundancy of latent tokens in multimodal models.

Benchmarking Bengali Dialectal Bias: A Multi-Stage Framework Integrating RAG-Based Translation and Human-Augmented RLAIF

BRAC University researchers propose a two-phase framework to evaluate dialectal bias in low-resource languages by coupling RAG-based dialect translation, CoT-guided LLM judges, and human fallback verification, creating a 4,000-question gold-standard benchmark across 9 Bengali dialects and demonstrating across 68,395 evaluations over 19 LLMs that linguistic divergence directly causes severe comprehension degradation.

Beyond Leaderboards: A Frozen-Rubric Contract for Position-Invariant Winning-Argument Detection

Addressing the pervasive speaker-order bias and runtime rubric leakage in LLM-as-a-judge systems, this paper introduces TAPIR, an evaluation protocol that internalizes survey-grounded persuasion rubrics via mode-split privileged self-distillation, algebraically guarantees exact position invariance (\(WPI=1\)) through an anti-symmetric scalar head and tied-order wrapper, and outperforms the same-backbone 24-call MS-PS-MLP baseline (\(65.80\%\) vs. \(58.12\%\)) with a \(39.41\times\) inference speedup on Change My View.

Beyond Safe Answers: Segment-Aware Listwise Alignment for Reasoning Safety in Large Reasoning Models

To tackle the dual-surface safety discrepancy in Large Reasoning Models (LRMs) where hazardous rationales hide behind safe refusals, SaLT-DPO decomposes responses into reasoning and answer segments, fits length-normalized listwise soft target distributions, enforces weakest-link joint safety coherence, and anchors benign utility to suppress over-refusal and reasoning degradation.

BiG-SURE: Bipartite Graph for Semantic Uncertainty and Reliability Estimation of LLMs

BiG-SURE introduces an unsupervised, black-box uncertainty estimator that uses stable low-temperature generations as anchors and perturbed high-temperature responses as probes in a bipartite graph, quantifying semantic confidence via normalized squared spectral energy.

Black-Box Membership Inference via Word-Level Probability Estimation

Addressing the fundamental challenge of auditing proprietary LLMs that expose only generated textual continuations without token-level logits, WPMIA estimates word-level generation probabilities via Monte Carlo sampling with local semantic kernel smoothing, aggregates them additively into sequence-level log-likelihoods, and contrasts member and non-member prefix interventions, achieving an average TPR@5%FPR of 42.0% across GPT-5-Chat, Gemini-2.5-Flash, and Claude-4.5-Haiku.

Browse all 129 Safety papers →


👻 Hallucination Detection (6)

HaluMem: Evaluating Hallucinations in Memory Systems of Agents

HaluMem establishes the first operation-level hallucination evaluation benchmark for LLM agent memory systems, decoupling traditional black-box QA into memory extraction, updating, and question answering stages across dialogues reaching one million tokens, revealing how upstream extraction omissions cascade into downstream updating and QA failures.

Library Hallucinations in LLM-Generated Code: A Risk Analysis Grounded in Developer Queries

This paper presents the first systematic study of how realistic developer prompt variations (natural language descriptors and controlled user typos/fabrications) trigger library name and member hallucinations in LLM-generated code, revealing alarming vulnerabilities to temporal constraints and typos, and introducing LIBHALLUBENCH for reproducible evaluation.

PVFH-Bench: Benchmarking Premise-Induced Visual Fixation Hallucinations in Vision-Language Models

This paper identifies a critical compounded failure mode termed Premise-Induced Visual Fixation Hallucination (PVFH) along with a diagnostic benchmark PVFH-Bench, revealing that contemporary LVLMs anchor onto visually confusable regions under misleading textual premises and that explicit chain-of-thought reasoning exacerbates rather than mitigates this hallucination trap.

Teaching LLMs to Abstain via Fine-Grained Semantic Confidence Reward

Addressing the ambiguous knowledge boundary caused by coarse-grained global signals in LLM abstention alignment, this paper proposes FISCORE, an RL framework that derives sample-specific semantic confidence rewards from clustering cluster sizes via GRPO, and introduces the monotonically sound F1rel reliability metric.

The Trilemma of Truth in Large Language Models

Addressing three flawed assumptions in conventional binary veracity probing, this paper proposes sAwMIL, a multiclass probing framework combining sparse multiple-instance learning with conformal prediction that reliably isolates true, false, and neither signals with causal steering guarantees across 16 LLMs.

To Answer or to Abstain: Mitigating Search-Agent Hallucinations via Abstention-Aware Reinforcement Learning

Addressing the over-answering hallucination issue where search-augmented LLM agents fabricate answers upon retrieval failure, this paper proposes Abstention-aWAre Reinforcement Learning (AWA-RL), dynamically shaping refusal rewards via query-specific prior capability estimation, an optimistic courage factor, and on-policy excess penalty to achieve a superior capability-reliability trade-off.


📊 LLM Evaluation (29)

Auditing MCQA Benchmarks through Probability Landscapes

Proposes a two-level MCQA benchmark auditing framework using model output probability distributions, characterizing macro benchmark structure via top prediction probability and normalized residual entropy summarized by Mean Pairwise Distance, and filtering distractor competition via noise injection to effectively prioritize flawed items for targeted human review.

Benchmarking LLM-as-a-Judge for Long-Form Output Evaluation

Tsinghua University and collaborating institutions introduce LongJudgeBench, the first meta-evaluation benchmark tailored for LLM-as-a-judge on long-form outputs averaging 9,249.7 tokens across five realistic scenarios, exposing a severe reliability gap, position biases, and superficial coverage illusions in current judges.

Check The Scoreboard: An Analysis of Scoring Schemes on Multiple Choice Evaluation

By systematically adapting six educational measurement scoring schemes to multiple-choice QA, this paper penalizes guessing, differentiates model capabilities beyond standard accuracy, and demonstrates that Answer Until Correct scoring strongly predicts real-world human preferences in LLM Arena.

CyclicJudge: Mitigating Judge Bias Efficiently in LLM-based Evaluation

Addressing systemic scoring biases in LLM-as-a-judge evaluation that cannot be averaged away by increasing test scenarios, this work establishes a generalizability theory variance decomposition and demonstrates that a round-robin judge assignment strategy (CyclicJudge) cancels panel-relative judge offsets and maximizes generation diversity at zero additional cost, cutting top-1 ranking flips by 61%.

docfollow a hierarchical fine grained constraints following benchmark in long co

Addressing the critical gap in evaluating Long-In-Long-Out text generation and the unreliability of LLM-as-a-judge scoring, DocFollow introduces a benchmark of 2,436 expert-validated instances across 5 domains and 20 subdomains with hierarchical constraints and a deterministic rule-based evaluation protocol.

Does task decomposition improve automatic NLG evaluation?

A systematic benchmark comparing decomposition-based and direct prediction LLM-as-a-Judge frameworks reveals that task decomposition yields no intrinsic performance gains, that historical advantages were artifacts of human label supervision and continuous float tie-breaking, and that direct prediction reaches human-level consensus when fairly calibrated.

Evaluating and Improving LLM Self-Modeling

This paper operationalizes LLM self-modeling as an empirical behavioral prediction capability over verifiable input-output distributions, introduces an unbiased skill benchmark spanning 10 task formats across four domains, and demonstrates that multi-task RL on closed-loop synthetic perturbation data reliably enhances self-modeling skill without requiring privileged internal introspection.

Evaluating LLM-generated Explanatory Dialogue Turns through Dialogue Completion

This paper uses dialogue completion as a dual lens across human-spoken, human-written, and synthetic modalities to investigate explanatory dialogue quality and metric behavior, uncovering a systematic written-text bias in the widely used reference-free FED metric that penalizes spoken interactivity, while showing that metric-driven zero-shot self-refinement reliably closes the model interactivity gap.

Exposía: Teaching and Assessment of Academic Writing Skills for Research Project Proposals and Peer Feedback

This paper introduces Exposía, the first publicly available multi-stage dataset linking student research project proposal writing and peer review feedback in higher education, paired with fine-grained pedagogically grounded rubrics across 36 proposal criteria and 24 review criteria, and benchmarks leading open-weight and closed-source LLMs on automated multi-aspect scoring.

From Rubrics to Reliable Scores: Evidence-Grounded Text Evaluation with LLM Judges

RULERS introduces a criteria-transfer framework for frozen LLM judges that locks human rubrics into structured task specifications, extracts mechanically verifiable textual evidence via typed operators, and calibrates model signals through monotone quantile distribution mapping, substantially outperforming direct scoring baselines in Quadratic Weighted Kappa (QWK) across four diverse benchmarks.

Browse all 29 LLM Evaluation papers →


⚡ LLM Efficiency (5)

Compressing Sequences in the Latent Embedding Space: K-Token Merging for Large Language Models

This paper proposes K-Token Merging, a latent-space soft prompt compression method that fuses contiguous blocks of \(K\) token embeddings into a single compressed embedding via an average-initialized residual MLP with LoRA fine-tuning, achieving up to 75% prompt length reduction and a 27.8% throughput speedup with negligible accuracy degradation.

Efficient Iterative Retrieval with Heterogeneous Batching

Addressing the complementary nature of compute-bound embedding and memory-bound generation in iterative RAG, ORTHRUS introduces a unified inference loop with heterogeneous batching, chunked embedding with incremental pooling, and workload-aware intra-batch scheduling to eliminate static GPU partitioning bubbles, achieving 1.28×–4.52× higher throughput and cutting end-to-end p99 latency by up to 55.8%.

SpecSteer: Synergizing Local Context and Global Reasoning for Efficient Personalized Generation

Addressing the core tension between the reasoning capacity deficit of on-device models and the privacy risks of cloud-centralized LLMs in personalized generation, SpecSteer introduces an asymmetric collaborative speculative decoding framework where an on-device private specialist drafts candidate sequences, a cloud generalist conducts non-private ratio-based verification, and a steering recovery mechanism corrects rejected tokens, closing the reasoning gap while delivering a 2.36× inference speedup.

Spexis: Speculative Lookahead Scheduling for LLM Inference

Spexis introduces speculative parallelism (SP) as an orthogonal scheduling dimension for multi-GPU LLM inference that overlaps self-speculative drafting with downstream pipeline verification while sharing the same KV cache, augmented with confidence-based selective speculation and lookahead remaining-length prediction to achieve up to 34% serving throughput speedup over optimized TP+PP baselines.

Survival-Guided Length Control for Efficient Reasoning

This paper proposes a plug-in, training-free survival-guided length predictor for masked diffusion language models (DLMs) that recasts sequence termination as a discrete-time survival process over the end-of-sequence token, deriving a closed-form expected length from single-pass EOS probabilities and delivering up to 7× decoding speedups across reasoning and code benchmarks with zero accuracy loss.


📚 Pretraining (7)

Beyond Junk Tokens: Subsumed Tokens in Japanese Subword Tokenization

Addressing the suppression of valid words caused by longer dominant tokens in frequency-driven subword tokenization for languages without explicit boundaries, this paper formalizes the notion of subsumed tokens and develops a tokenizer-agnostic post-processing framework that searches optimal binary segmentation trees via dynamic programming to restore valid word statistics without altering tokenizer vocabularies.

DebugLM: Learning Traceable Training Data Provenance for LLMs

DebugLM equips LLMs with internal parametric data provenance by binding discrete dataset tags via dual-mode training, enabling accurate multi-stage behavior tracing (>95% TSR) in a single forward pass and targeted test-time remediation without retraining.

How Should LLMs Consume High-Quality Data? Optimal Data Scheduling via Quality-Aware Functional Scaling Laws

This paper extends functional scaling laws to time-varying data quality, derives asymptotically optimal joint data-quality and batch-size schedules via variational analysis, and proposes the Drop-Stable-Rampup recipe for LLM midtraining that achieves substantial gains across dense and MoE architectures.

MDM-Prime-v2: Binary Encoding and Index Shuffling Enable Scaling of Diffusion Language Models

MDM-Prime-v2 reveals that Byte-Pair Encoding induces severe low-entropy sub-tokens in masked diffusion models, degrading cross-entropy loss; by introducing offline index shuffling to maximize sub-token entropy and selecting the theoretical limit of binary encoding (\(\ell=\lceil\log_2 V\rceil\)), it scales diffusion language models to 1.1B parameters, achieving superior zero-shot accuracy across eight commonsense reasoning benchmarks over both autoregressive and diffusion baselines.

OptiMer: Optimal Distribution Vector Merging Is Better than Data Mixing for Continual Pre-Training

OptiMer decouples data mixture ratio selection from model training in continual pre-training (CPT) by independently training single-corpus models, extracting their distribution vectors, and optimizing continuous merge weights post-hoc via TPE Bayesian optimization, achieving superior multi-domain capability with 15–35× lower search cost on Gemma 3 27B.

SpinHash: An O(1) Encoder for Any Vocabulary

SpinHash introduces a compact, vocabulary-independent encoder that maintains a strictly constant \(O(1)\) parameter budget of \(\sim 0.2\text{M}\) parameters across arbitrary vocabulary sizes and modality counts, projecting discrete tokens into geometrically structured binary spin codes and decoding them via a single shared feedforward network to rival standard embedding tables on text, extreme \(k\)-mer genomics, and clinical multimodality.

WRAP++: Web Discovery Amplified Pretraining

Addressing the intra-document information silos and extraction bottlenecks of single-document rewriting, WRAP++ leverages web hyperlink topologies to discover high-confidence entity relationship subgraphs, generating cross-document multi-hop QA pairs that amplify an 8.4B raw token corpus into 82.7B tokens and substantially boosting parametric factual knowledge recoverability on SimpleQA.


✏️ Knowledge Editing (4)

Editing Rules in LLMs Requires More Than a Localized Update

Addressing the failure of localized factual model editors on mathematical and physical rules, this paper reveals via fine-grained causal tracing that rule knowledge exhibits form-specific layer-wise separation (formulas and descriptions reside in early layers while numerical instances peak in middle layers), and proposes Distributed Multi-Layer Editing (DMLE) to perform coordinated cross-layer updates that dramatically boost rule understanding and instance portability.

Generalizable Lifelong Model Editing via Preference Optimization

To tackle superficial prompt overfitting and catastrophic forgetting in lifelong knowledge editing, GLIME formulates continual editing as a joint optimization problem combining factual injection, edit-independent preference optimization for behavior-level regularization, experience replay, and randomized orthogonal gradient projection, achieving superior generalization and downstream portability under realistic autoregressive decoding.

Measuring the Depth of LLM Unlearning via Activation Steering

Proposes the Unlearning Depth Score (UDS), a training-free, causal, and dataset-invariant metric that evaluates the recoverability of forgotten knowledge across internal layers via two-stage activation patching, outperforming observational probes and output-level metrics in detecting residual knowledge obscured by representation shifts.

Validator-Guided Decoding for Multi-Hop Knowledge Editing

To resolve the failure of edited language models to compose updated facts in multi-hop reasoning despite retaining strong fact-checking capabilities—termed the generate-verification gap—this paper proposes Validator-Guided Decoding (VGD), which employs a lightweight trained validator as an external correctness signal to early-commit high-confidence candidates and selectively revise inconsistent predictions at test time.


💬 LLM (Other) (49)

Jasper-Flash: Elastic Token Compression and Progressive Distillation for Inference-Scalable Text Embedding Models

Jasper-Flash proposes an inference-scalable text embedding framework integrating Elastic Token Compression (ETC) and Compression-Adaptive Progressive Distillation (CAPD), enabling dynamic sequence reduction on the fly; its 600M model slashes 1024-token encoding latency from 24.24 ms to 4.48 ms (a 5.4x speedup) while achieving an MTEB score of 74.75, outperforming substantially larger 4B models.

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

Addressing the severe computational and communication overhead in MUON caused by the ill-conditioned momentum spectrum during Newton-Schulz iterations, MUON2 applies Adam-style adaptive second-moment preconditioning prior to orthogonalization, achieving superior directional alignment with 40% fewer iterations and cutting pretraining wall-clock time by 25%.

TCC-Bench: Benchmarking the Traditional Chinese Culture Understanding Capabilities of MLLMs

To assess the vulnerabilities of multimodal large language models (MLLMs) in non-Western cultural understanding, this paper introduces TCC-Bench, the first comprehensive bilingual VQA benchmark spanning eight distinct knowledge domains with 770 curated images and 1,015 questions, revealing persistent deficits across both open- and closed-source models when grounding nuanced cultural knowledge in fine-grained visual evidence.

Can LLMs Reliably Self-Report Adversarial Prefills, and How?

Across 10 open-weight instruction-tuned LLMs and four safety benchmarks, models fail to reliably self-report whether a previous response was elicited by an adversarial prefill (claiming intent on 25.3% of compromised responses); this introspective signal is causally mediated by the internal refusal direction, and targeted finetuning to flag compromised responses counterintuitively increases initial attack success rates due to cross-turn parameter sharing.

TRACEPROBE: An Execution-Based Benchmark for Diagnosing Safety Failures in LLM Agents

TRACEPROBE is an execution-based sandbox benchmark for diagnosing LLM agent safety across 155 scenarios and three risk sources, using normalized traces and deterministic action replay to localize intermediate latent violations and pinpoint the exact step of failure.

ABDATABENCH: Evaluating AI Agents for Evidence-Grounded Antibody Record Construction

Addressing the bottleneck of dispersed and unstructured scientific evidence in therapeutic antibody literature, this paper presents ABDATABENCH, the first benchmark evaluating complete 22-field antibody record construction across 44 expert-annotated documents (279 records), alongside ABCURATOR, a record-centric multi-agent framework paired with an expert-driven skill accumulation workflow.

Complexity-Aware Routing: Detecting and Allocating Dynamic Compute for Efficient Mixture-of-Experts Large Language Models

CAR introduces an information-theoretic dual-signal framework combining Shannon Surprisal and Predictive Entropy within a decoupled perception-decision architecture, enabling dynamic expert allocation that cuts active parameters by ~10% while outperforming static Top-k and existing dynamic routing baselines.

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

The CASE framework constructs a sparsified bidirectional contradiction graph over sampled LLM candidate responses and computes its normalized spectral norm to quantify semantic disagreement, enabling dynamic early-stopping sampling that cuts generation costs by over 30% while combining token-level predictive entropy from a deterministic anchor to resolve failures on consistent errors.

Syntactic Complexity Convergence in Dialogue: Analysis of the Phenomenon and Application to LLM Detection

This paper conducts a systematic investigation of syntactic complexity convergence across full-length dialogues in three languages, discovering that while human dyads exhibit significant within-metric convergence in task-oriented settings, human-LLM dialogues exhibit a unique cross-metric convergence revealed by Linear-CKA, which enables an estimator-free Fourier-based detection of LLM-involved dialogues achieving 96% accuracy.

Advancing Fact-Checking for Large Language Models via Argumentative Reasoning

Addressing confirmation bias in retrieval and multi-source evidence conflict in real-world claim verification, this paper introduces ArgCheck, a neuro-symbolic framework coupling multi-round adversarial Pro-Con evidence collection with formal Dung abstract argumentation under grounded semantics, improving Macro-F1 by 5.39 and 11.10 points on MAVEN and AVeriTeC, respectively.

Browse all 49 LLM (Other) papers →


📖 NLP Understanding (49)

A Survey on Representing Linguistic Style: Challenges and Opportunities

This survey synthesizes linguistic theories and NLP representation learning to establish a unifying, content-independent definition of style, reviews predefined and neural style embedding architectures, and empirically shows that dedicated style embeddings achieve superior accuracy over frontier LLMs at thousands of times lower compute.

Alice: A Large-Scale German Benchmark for Rubric-Based Multi-Dimensional Automatic Short Answer Scoring

Addressing the limitations of conventional automatic short answer scoring that rely on reference answer overlap and single global scores, this paper introduces ALICE, a 16.5K German student response benchmark covering three distinct dimensions (learning performance, knowledge elements, and scientific skills), and formulates scoring as cross-encoder rubric retrieval to dramatically improve generalisation across unseen questions.

Are Non-English Papers Reviewed Fairly? Language-of-Study Bias in NLP Peer Reviews

This paper provides the first systematic characterization of language-of-study (LoS) bias in NLP peer review, presenting the expert-annotated LOBSTER benchmark alongside an LLM detection pipeline (87.37 Macro F1), and empirically demonstrates across 15,645 reviews that non-English papers face ~40x higher bias rates than English-only papers, dominated by unwarranted demands for cross-lingual generalizability.

BeyondAtoms: Benchmarking and Verifying Relational Factuality Beyond Atomic Facts

Addressing the vulnerability where conventional fact verifiers and atomic decomposition blindly trust entity/event overlaps while ignoring logical topologies, this paper introduces BEYONDATOMS, the first operation-centric relational factuality benchmark, and CARVE, a training-free verification framework based on claim-anchored extraction and bidirectional structural alignment.

CausalTAD: Injecting Causal Knowledge into Large Language Models for Tabular Anomaly Detection

Addressing the flaw where random column ordering disrupts feature causal dependencies in LLM-based tabular anomaly detection, CausalTAD proposes high-level latent factor causal discovery, column-level causal preference projection, linear ordering problem (LOP) relaxation, and factor-count reweighting, achieving state-of-the-art performance across 30+ benchmarks.

charting empirical patterns for llm fine tuning in scientific multi discipline l

A systematic empirical study of LLM joint fine-tuning across five naturally imbalanced scientific disciplines (math, chemistry, biology, medicine, and geography) distills four governing patterns and shows that an asymmetric shared-A LoRA-MoE rivals full fine-tuning while training only 0.902% of the parameters.

CodeTD: Topology of Attention Detects Hallucinations in Code LLMs

CodeTD models the attention maps of Code LLMs as weighted graphs, leverages persistent homology and Manifold Topology Divergence (MTD) to extract multi-scale 0-dimensional connectivity and 1-dimensional detour cycle features between prompt and generation tokens, enabling highly accurate, execution-free code hallucination detection and best-of-N reranking.

Deep and shallow biases in language models

This paper introduces the bias depth score \(\pi = \text{DR} \times \text{FR}\), decoupling repeated LLM preferences in open-ended choice tasks into persistent Deep biases inherited from pretraining and fragile Shallow biases driven by surface prompt wording across 4,442 grounded prompt families.

DiscoTrace: Representing and Comparing Answering Strategies of Humans and LLMs in Information-Seeking Question Answering

DiscoTrace represents long-form answers as structured sequences of question-related discourse acts paired with explicit question interpretations, revealing that human communities exhibit diverse rhetorical answering policies while large language models remain rhetorically homogeneous and systematically over-answer low-probability interpretations.

DiscoTrack: A Multilingual LLM Benchmark for Discourse Tracking

DiscoTrack establishes a standardized question-answering benchmark spanning 12 typologically diverse languages and 11 subtasks across four discourse tiers (salience, entity tracking, discourse relations, bridging inference), demonstrating that even frontier LLMs struggle with implicit document-level discourse reasoning.

Browse all 49 NLP Understanding papers →


✍️ Text Generation (15)

ArgSum-X: A Structure-Aware Benchmark for Argument Summarization and Key Point Analysis

Addressing the long-standing fragmentation between argumentative text summarization (ATS) and key point analysis (KPA) along with the absence of structural fidelity metrics, ArgSum-X presents the first unified multi-stage benchmark and structure-aware evaluation framework, revealing that while general language models achieve high semantic faithfulness, they suffer from substantial deficits in preserving argumentative flow and global reasoning topology.

Beyond a Single Story: Meta-Reviewing Sparse and Incomplete User-Generated Content

Addressing both macro review sparsity and micro attribute incompleteness in user-generated content, MOSAIC adapts academic meta-review principles to aggregate attribute-sentiment evidence from neighbor users, combining an MMoE multi-task backbone with personalized attention to deliver refined rating predictions and attribute-level explanations.

Beyond Prefix Answering: Parallel Infilling Reading for Long Contexts with Diffusion Language Models

Addressing the underutilization of native bidirectional infilling in diffusion language models caused by adopting autoregressive prefix-answering paradigms, this paper proposes Parallel Infilling Reading (PIR) to insert extraction slots between context chunks for parallel evidence denoising and subsequent answer reduction, while resolving cross-slot attention interference through slot-isolation masking and single-active-slot fine-tuning.

curp codebook based continuous user representation for personalized generation w

CURP proposes a plug-and-play personalization framework based on a discrete prototype codebook and a dual-aggregator, factorizing user interaction history into orthogonal subspace prototypes to capture query-adaptive dynamic preferences and stable intrinsic traits, achieving state-of-the-art personalized generation while training only 28M parameters (0.36% of the base LLM).

Evaluating Style-Personalized Text Generation: Challenges and Directions

This paper introduces a comprehensive benchmark spanning eight writing domains and three evaluation settings to scrutinize style discrimination metrics, empirically demonstrating that ensembles of diverse evaluation metrics consistently outperform individual evaluators.

Evo: Autoregressive–Diffusion Large Language Models with Evolving Balance

This paper introduces Evo, a duality latent flow model that unifies autoregressive next-token prediction and diffusion denoising into a continuous probability trajectory, adaptively balancing global planning and local refinement via learned per-token progression variables while preserving near-AR decoding throughput (52 tokens/s).

FAVoR: Measuring and Mitigating Author-Style Homogenization in Federated Personalized Generation

Addressing author-style homogenization in federated personalized text continuation—where central parameter aggregation erodes individual stylistic nuance toward a shared register—FAVoR proposes a decoupled shared-private PEFT residual mechanism that isolates local author-specific style adjustments outside aggregation, substantially improving cross-author style discriminability while maintaining continuation utility.

POLARIS: Guiding Small Models to Write Long Stories

POLARIS introduces a compute-efficient creative-writing RL recipe that integrates a structured 16-dimension LLM-as-judge Story Quality rubric with teacher-forced human-reference injection (HRI) inside GRPO groups, allowing a 9B model trained on 4 A100 GPUs for 48 hours to preserve narrative coherence and close length adherence across stories up to 12,000 words.

Provenance Before Prose: Claim-Locked Reporting for Statistical Text Generation

Addressing numerical drift, direction inversions, and rhetorical overstatements when LLMs draft statistical reports for functional neuroimaging (fMRI) and randomized controlled trials (RCTs), this paper introduces Claim-Locked Reporting, a protocol that locks reportable claims—including evidence sources, exact numbers, effect directions, and permitted language strengths—in a typed ledger before generation, achieving 98.5%–100.0% cross-seed reproducibility while eliminating harmful statistical fabrications.

SGTR: Learning Style-Generalizable Trajectory Reasoning for Zero-Shot Text Style Transfer

SGTR formulates text style transfer as the joint modeling of structured reasoning trajectories and final stylized outputs, decoupling the process into three interpretable stages alongside latent contrastive regularization to achieve superior zero-shot generalization on unseen styles.

Browse all 15 Text Generation papers →


🗣️ Dialogue Systems (21)

AI Psychosis: Does Conversational AI Amplify Delusion-Related Language?

This paper provides the first large-scale empirical study examining whether conversational AI amplifies delusion-related language during extended interactions, demonstrating that interactions with unconditioned conversational AI increase delusion-related language trajectories by an average of 233% in vulnerable simulated users, while conditioning AI responses on real-time inferred DelusionScore effectively reverses this harmful feedback loop.

Anticipating User Interpretations in Emotional Support Conversations via Inference-Time User-Side Simulation

Addressing the frequent conversational breakdown where sensible advice is misperceived by help-seekers as dismissive, judgmental, or pressuring, this paper presents Cognitive Look-ahead Simulation (CLS), an inference-time framework that infers structured user states, anticipates immediate user inner thoughts and state transitions via a lightweight cognitive world model, and selectively rewrites high-risk drafts before delivery, substantially reducing severe failure rates and improving multi-turn dialogue outcomes.

Beyond Single-Policy: Evaluating Composed Organization-Specific Policy Alignment in LLM Chatbots

Addressing the critical gap where real-world chatbot queries naturally span multiple organizational rules, COPAL introduces an automated evaluation framework driven by four empirical interaction patterns and target facets paired with explicit handling contracts, revealing an average composed-policy error rate of 33.1% across 9 frontier LLMs.

Client Simulation via Conversational Trajectory Modeling for Both Trainee Practice and Model Evaluation in Mental Health Counseling

PsyCLIENT tackles the overly compliant and behaviorally uniform nature of simulated counseling clients by decoupling static client profiles from dynamic conversational behavior trajectories, constructing the PsyCLIENT-CP benchmark of 120 Chinese clinical profiles and achieving superior authenticity and training utility.

From Chatbots to Confidants: A Cross-Cultural Study of LLM Adoption for Emotional Support

This paper presents the first large-scale cross-cultural empirical study of LLM adoption for emotional support across seven countries (\(N=4,641\)), disentangling individual demographics from country-level cultural variation via Cumulative Link Mixed Models (CLMM) and revealing a sharp divide between enthusiastic Anglosphere and skeptical Continental European users.

Know What You should not Know: Dynamic Attention Gating for Knowledge-Bounded Role-Playing

To tackle out-of-boundary knowledge leakage in role-playing conversational agents where Chain-of-Thought reasoning exposes anachronistic or cross-world knowledge, this paper proposes the Dynamic Attention Gating (DAG) Router—a lightweight network (<0.6% parameters of the host LLM) that keeps the LLM frozen and dynamically predicts per-layer FFN neuron gating masks from the system prompt, substantially boosting out-of-boundary compliance without compromising general conversational ability.

Learning an Efficient Multi-Turn Dialogue Evaluator from Multiple LLM Judges

Addressing the excessive inference overhead of multi-LLM judge ensembles and systematic biases in single judges, this paper proposes MTDEval, a lightweight evaluator trained on a new 10-dimensional multi-judge preference dataset (P2-MTD) using Thurstone's Case V model and Signal Detection Theory to jointly estimate judge reliabilities and dialogue quality, outperforming strong open-source baselines with over 10× faster inference.

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

Addressing the dual challenges of non-convergent clarification and strict rollout filtering in open-domain conversational QA, D4CQ introduces procedural scaffolding, contrastive intent alignment, response-diagnosticity loss reweighting, and compatibility-gated hindsight relabeling, substantially outperforming existing clarify-or-answer baselines under a comparable interaction budget.

PersonaDual: Balancing Personalization and Objectivity via Adaptive Reasoning

To resolve the double-edged sword dilemma where personalized user personas enhance subjective alignment but degrade factual objectivity and logical reasoning, PersonaDual unifies general objective reasoning and personalized reasoning within a single LLM, adaptively switching between modes via a novel DualGRPO reinforcement learning algorithm.

Preference-Aligned Data Synthesis for Fine-Grained Intent Detection

Addressing inter-class ambiguity and intra-class homogeneity in LLM-synthesized data for fine-grained intent detection, this paper introduces a parameter-level generator alignment framework that mines multi-centroid prototype-aligned diverse positives and confusability-weighted cross-intent hard negatives for Direct Preference Optimization (DPO), delivering significant downstream gains in zero-shot and few-shot regimes while improving generator reliability.

Browse all 21 Dialogue Systems papers →


🌐 Multilingual & Translation (24)

ALEE: Any-Language Evaluation of Embeddings via English-Centric Minimal Pairs

By applying controlled semantic perturbations on the English side of parallel corpora via Abstract Meaning Representation (AMR) with bidirectional NLI filtering, ALEE establishes a dynamic minimal-pair diagnostic benchmark across 275+ languages, exposing systematic cross-lingual alignment blind spots in structural and lexical semantics.

Assessing the Impact of Language Disparity on Multilingual Linguistic Ability in Large Language Models

Evaluating six LLM families across 101 languages on the MultiBLiMP syntactic minimal-pair benchmark under four paradigms reveals that post-training imposes a linguistic alignment tax borne disproportionately by low-resource languages, while native-language prompting crucially recovers latent grammatical competence masked by unprompted probabilities.

Bridging the Consistency Gap: A Noun Bank Augmented Framework for Low-Cost Long Text Translation

To resolve the severe tension between document-level lexical consistency and inference cost in long-text machine translation, NoBAT decouples translation into coarse serial Noun Bank construction with historical disambiguation and independent parallel sentence decoding via local noun constraint retrieval, improving entity consistency while slashing total token consumption by up to 6.3× and achieving an 8.8× wall-clock speedup.

Cross-Attention Eigenvalues Know the Word Order in Multilingual Machine Translation

By training linear probes on the normalized eigenvalue spectra of cross-attention Gram matrices in NLLB-200, this paper demonstrates that cross-lingual word-order typology (SVO vs. SOV) is near-perfectly decodable and broadly distributed across layers and heads when translating from English into diverse target languages, yet degrades sharply when translating into fixed English, proving that cross-attention geometry is fundamentally shaped by target-side generation demands rather than source-side structure.

Cross-lingual Biography Enrichment via Claim Extraction and Alignment

Addressing the pronounced knowledge asymmetry where non-English Wikipedia articles document long-tail figures with far richer detail than English Wikipedia, this paper presents a claim-based cross-lingual biography enrichment framework that extracts atomic claims into a shared English space, filters already covered information via fine-grained semantic alignment, and rewrites the English article with selected novel claims, achieving significantly higher supported additions with substantially lower hallucination rates.

DeepSEA: A Multilingual Evaluation Benchmark for Southeast Asia

Google Research introduces DEEPSEA, the first comprehensive human-translated benchmark covering 12 Southeast Asian languages across high- and low-resource tiers on 5 NLP tasks, uncovering severe cross-lingual performance collapse in open-weights models and an overwhelming reliance on English cognitive representations.

Discourse Dependency: A Continuous Criterion for Translation Difficulty

The paper reframes machine translation difficulty from traditional output-based quality estimates into a continuous, metric-free source-side criterion called Discourse Dependency (DDP) based on referential reach, demonstrating that frontier LLMs fail to exploit long-range document context and that standard automatic metrics suffer from systemic blindness on context-demanding segments.

FuzhouBench: A Benchmark for Eastern Min Chinese with Tone Sandhi Annotations

This paper introduces FuzhouBench, the first public NLP benchmark for Eastern Min Chinese (Fuzhou dialect) comprising 22.8k dictionary entries with 11.3k tone sandhi annotations and 3.06 hours of paired audio; empirical evaluations demonstrate that frontier LLMs fail on Yngping G2P and symbolic sandhi rules, while colloquial search queries trigger severe BM25 collapse that requires a two-stage hybrid and cross-encoder retrieval pipeline to recover.

IndicParam: Benchmark to evaluate LLMs on low-resource Indic Languages

IndicParam presents a human-verified graduate-level benchmark of 13,207 questions across 11 low- and extremely low-resource Indic languages derived from the official UGC-NET national examinations, systematically decoupling pure linguistic proficiency from general knowledge and demonstrating that even the top-performing model (Gemini-2.5) achieves only 58.0% average accuracy.

Length-Adaptive Decoding for Masked Diffusion Machine Translation

Addressing the structural defect in fixed-canvas masked diffusion machine translation where target canvas length must be pre-allocated before denoising (causing truncation or hallucination), this paper introduces Entropy-Valley (EV), a training-free test-time length selector that scores candidate canvas lengths by mean predictive entropy across all-mask forward passes, closing 33.0%–65.3% of the COMET-22 gap to reference lengths across En↔Zh and En→De while matching or exceeding matched-data autoregressive LLaMA-3-8B.

Browse all 24 Multilingual & Translation papers →


🔍 Information Retrieval & RAG (72)

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.

Browse all 72 Information Retrieval & RAG papers →


💻 Code Intelligence (8)

ChainSWE: Benchmarking Coding Agents on Multi-Bug Software Maintenance

Introduces ChainSWE, the first benchmark evaluating coding agents on sequential multi-bug software maintenance across shared codebases, demonstrating that accumulated repository state drift causes resolution rates to drop by up to 70% along the repair chain.

DSEffi-Bench: Demystifying Large Language Models’ Capabilities in Data Science Programming

Introduces DSEffi-Bench, the first execution efficiency benchmark for data science code generation (comprising 1,000 human-validated instances from real Stack Overflow pain points paired with stress-testing suites), revealing significant decoupling between functional correctness and runtime efficiency, uncovering that 79.1% of performance deficits stem from domain-specific root causes, and demonstrating up to +14.7% efficiency improvement and 13.0× cost reduction via taxonomy-guided optimization and library-conditioned routing.

Fast and Focused: Interactive Repository-Level Fill-in-the-Middle Code Completion with Hybrid Structural Retrieval

To tackle coarse retrieval granularity and suffix duplication in interactive repository-level code completion, ACAR combines lexical sliding-window and AST/LSP structural retrieval via line-level reciprocal rank fusion (RRF) alongside syntax-aware post-processing, outperforming strong baselines with only 87–95 ms non-decoding overhead.

Functional Entropy: Predicting Functional Correctness in LLM-Generated Code with Uncertainty Quantification

Addressing the collapse of NLI-based semantic entropy caused by its inability to distinguish functional code differences, this paper introduces a family of functional equivalence methods powered by LLM-based functional equivalence assessment to accurately predict code correctness across multiple programming languages without code execution.

PolySQL: Scaling Text-to-SQL Evaluation Across SQL Dialects via Automated Backend Isomorphism

PolySQL introduces a dual-execution evaluation methodology based on backend isomorphism and normalized comparison that eliminates the need for fragile query transpilation, uncovering a 25.2% relative performance degradation and predominant logical regression across SQL dialects with 100% execution coverage.

SABER: Benchmarking Operational Safety of LLM Coding Agents in Stateful Project Workspaces

SABER establishes a stateful benchmark of 716 executable Docker-sandboxed project tasks to evaluate the operational safety of LLM coding agents from multi-step execution traces and persistent workspace state deltas, revealing that even frontier models exhibit over 54.7% harmful safety-violation rates.

Verification Trap: Understanding Test-Time Selection Failures under False Premises in Code Generation

This paper demonstrates that test-time verifier selection in code generation fails systematically under misleading task premises because the generator and verifier become coupled through shared mistaken assumptions; verifier-written tests omit counterexamples and validate premise-consistent shortcuts over correct programs, a failure that can be predicted prior to hidden execution and mitigated via decoupled evidence channels.

XRepoTest: Benchmarking Multilingual Repository-Level Unit Test Generation for Large Language Models

XRepoTest establishes the first repository-level unit test generation benchmark across 5 high-impact yet underexplored languages (Rust, Go, Julia, PHP, and Ruby) spanning 3,642 focal methods, introducing Invocation Rate (IR) to expose misleading indirect test coverage and revealing that LLMs face severe bottlenecks under real project constraints with test pass rates peaking below 27%.


🎨 Image Generation (2)

Decomposing Visual Histories with Vision-Language Agents: Hierarchical Temporal Guidance for Compositional Image Generation

HTGF is a training-free framework for multi-reference compositional image generation that decomposes visual references into Subject, Structure, and Detail axes via a VLM agent, schedules their injection into SNR-aligned denoising windows derived from closed-form Tweedie sensitivity, and eliminates stage-transition artifacts using a manifold-aware corrector.

SpatialGuard: Harness-Guided Verifiable Spatial Reasoning for Text-to-Image Generation

SpatialGuard reframes complex 3D spatial text-to-image generation into an editable, verifiable layout state and introduces an agent-based Layout Harness (rule constraints, tool invocation, shared knowledge, and feedback loops) that orchestrates planning, visual realization, structured validation, and atomic repair to conquer multi-turn constraint decay, achieving a state-of-the-art 9.37 overall spatial faithfulness score.


🎬 Video Generation (1)

Helping Figures Tell their Story! Paper-Grounded Video Generation Explaining Complex Scientific Figures

To alleviate the high cognitive load of understanding dense scientific architecture diagrams, this paper establishes FigTalk, the first paper-grounded figure-to-video benchmark, and introduces MINARD, a three-team decoupled pipeline that coordinates narration generation, figure perception, and constrained visual grounding to synthesize faithful, temporally synchronized walkthrough videos.


🧩 Multimodal VLM (131)

Not What, But How: A Communicative Audit of LLM Response Framing

Introduces FRANZ, an automated communicative audit framework, and SQUARE, a 376k-sample subjective query corpus from Reddit, revealing that LLM framing exhibits persistent model-level signatures and a strong bidirectional coupling between insider cultural positioning and affective anthropomorphism.

AdaRetouch: An Interleaved Vision-Language Agent for Adaptive Multi-Turn Image Retouching

AdaRetouch introduces an interleaved vision-language agent for multi-turn photo retouching that replaces rigid, open-loop pipelines with dynamic intermediate visual feedback, curriculum-driven trajectory generation, and process-level reinforcement learning to achieve state-of-the-art non-destructive photo enhancement.

An Image Is Worth Ten Thousand Words: Verbose-Text Induction Attacks on VLMs

VTIA introduces a novel two-stage gray-box verbose-text induction attack that combines reinforcement-learning-driven adversarial prompt discovery with vision-aligned cosine and standard-deviation perturbation optimization, inducing target Vision-Language Models to generate up to 121.90x longer responses and drastically increasing inference energy and latency without requiring access to LLM backbone parameters.

AnaBench for Enhancing Scientific Table & Figure Analysis

This paper introduces ANABENCH, a large-scale benchmark of 63,178 instances across 9 scientific domains and 170 subdisciplines categorized along seven complexity dimensions, and proposes ANAGENT, a four-agent collaborative framework (PLANNER, EXPERT, SOLVER, CRITIC) with modular SFT and GRPO training that outperforms baselines by up to +42.12%.

Anchoring Speech with Semantics: A Multimodal Adapter Mechanism for Automatic Speech Recognition in Low-Resource Languages

Addressing severe transcript scarcity in low-resource speech recognition, SAMA-ASR inserts lightweight gated cross-attention adapters into frozen Whisper decoders to condition autoregressive generation on upstream auxiliary translations and speech acoustic anchors, cutting character error rates substantially without requiring oracle test translations.

Anchoring What Matters: A Dual-Level Learning Framework for Visually-Grounded Multimodal Reasoning

Addressing the dual-level optimization bottleneck in multimodal RLVR—where valuable visually-grounded trajectories are discarded and uniform token advantage allocation dilutes critical perceptual and logical steps—PIVOT introduces self-calibrated experience replay as retrospective anchors alongside vision-guided advantage allocation based on counterfactual support and entropy-gated future impact, delivering strong empirical gains on multimodal reasoning benchmarks.

AVSCap: Orchestrating Audio-Visual Synergy for Omni-modal Video Captioning

To tackle modality isolation and speech-centric bias in omni-modal video captioning, AVSCap constructs a 130K tri-modal corpus via a decoupled-then-fused pipeline, trains a 7B captioner with hybrid-reward GRPO, and introduces AVSCapBench with fine-grained atomic event recall.

Beyond Sparse Rewards: A New Benchmark and Structure-Aware Graph Alignment for Micro-Drama Understanding

Addressing character identity confusion and temporal hallucinations caused by sparse rewards in hyper-dense micro-dramas, this paper introduces M-Drama—the first large-scale bilingual benchmark spanning 35K+ instances—and proposes SAGA, a structure-aware graph alignment reward that decouples semantic triplet matching and transitive closure temporal order over narrative graphs, substantially outperforming vanilla DAPO and SFT baselines on Qwen3-VL-8B.

Beyond the Current Observation: Evaluating Multimodal Large Language Models in Controllable Non-Markov Games

Addressing the critical vulnerability where frontier MLLMs over-rely on current observations and fail to act on past unobserved context, this paper presents RNG-BENCH—a controlled suite of non-Markov games spanning Matching Pairs and 3D Maze navigation—introducing a pairwise duel protocol and a normalized Memory Gap diagnostic, while proving that SFT on synthesized trajectory rollouts significantly enhances in-context belief tracking without degrading general multimodal capabilities.

beyond visual memory mechanistic diagnostics of latent visual reasoning

By decomposing latent visual tokens into latent slots, boundary markers, and format, and evaluating on a state-of-the-art probe MVH, this paper reveals that latent gains do not stem from slot-content visual memory, but instead rely on boundary-marker control and visual attention routing.

Browse all 131 Multimodal VLM papers →


🧠 VLM Reasoning (17)

Ada-DeVL: Adaptive Decoupling of Visual and Logical Reasoning Pathways for Efficient VLM Inference

Ada-DeVL decouples multimodal test-time reasoning into an orthogonal \(2 \times 2\) taxonomy of visual perception depth and logical reasoning intensity, optimizing dynamic pathway routing via self-distillation SFT and GRPO with diversity/length penalties to drastically curtail token waste and overthinking.

Decompose, Look, and Reason: Reinforced Latent Reasoning for VLMs

The authors introduce DLR, a reinforced latent reasoning framework that dynamically interleaves textual premise generation, premise-conditioned continuous visual latent extraction, and grounded rationale deduction, coupled with a Spherical Gaussian Latent Policy (SGLP) under GRPO that unlocks active latent exploration.

FloCA: Towards Faithful and Logically Consistent Flowchart Reasoning

To tackle the severe hallucination and topology-inconsistent transitions in flowchart-oriented dialogue, FloCA decouples language understanding from graph execution via an external topology-constrained reasoning tool, achieving state-of-the-art task success and path coverage.

How Robust is OCR-Reasoning? Evaluating OCR-Reasoning Robustness of Vision-Language Models under Visual Perturbations

This paper introduces OCR-Robust, a controlled diagnostic benchmark evaluating vision-language models on OCR reasoning under visual perturbations, demonstrating that higher reasoning capability and larger scale do not guarantee robustness, and that structured visual representations such as charts and tables suffer pronounced worst-case degradation.

Keeping Eyes on the Image: Dynamic Visual Attention Recalibration for Vision-Language Reasoning

To tackle visual attention decay in Large Vision-Language Models (LVLMs) during long-form Chain-of-Thought reasoning—where models progressively ignore image evidence in favor of generated text—this paper proposes DVAR (Dynamic Visual Attention Recalibration), a training-free inference-time method that dynamically amplifies visual token value vectors via spatial semantic weighting and temporal shift scheduling, consistently boosting reasoning fidelity across six benchmarks.

Learning to Reason Before Masking for Referring Audio-Visual Segmentation

To address the lack of multi-hop cognitive reasoning over implicit clues in referring audio-visual segmentation, this paper introduces the Reasoning Ref-AVS task, the ReAVS benchmark, and Re2Seg, an omnimodal framework that reasons over candidate-entity evidence before conditioning mask decoding.

MedQA-MM: Shortcuts Behind Medical Visual Reasoning

Addressing the "reasoning inflation" crisis where medical multimodal multiple-choice benchmarks conflate surface accuracy with clinical visual reasoning, this paper establishes an evidence-ladder auditing framework with matched minimal repairs across six benchmarks, and releases MEDQA-MM, a 1,000-item shortcut-mitigated evaluation benchmark.

MeetingToM: Evaluating Multimodal LLMs on Theory-of-Mind Reasoning in Multi-Party Meetings

Addressing the gap where existing Theory-of-Mind evaluations focus on overt, externally verifiable video QA, MeetingToM establishes a hierarchical benchmark spanning individual mental states, dyadic addressee-attitude pairs, and group pseudo-consensus across 60 naturalistic multi-party meetings, revealing that leading MLLMs struggle with multimodal cue calibration and suffer severe bias when distinguishing genuine agreement from concealed dissent.

MemDreamer: Decoupling Perception and Reasoning for Long Video Understanding via Dreaming-Induced Dynamic Working Memory

MemDreamer decouples long video understanding into streaming perceptual memory construction and active agentic reasoning over a pure-text hierarchical graph, achieving state-of-the-art results (85.5 on LVBench) while cutting the active reasoning context window down to 5.9K–6.3K tokens.

MR^2-Bench: Going Beyond Matching to Reasoning in Multimodal Retrieval

MR^2-Bench introduces the first reasoning-intensive multimodal retrieval benchmark spanning 12 diverse sub-tasks across knowledge retrieval, illustration search, and visual relation reasoning, revealing that SOTA embedding models collapse from ~78% on standard benchmarks to under 10% Recall@1, while test-time CoT rewriting and multimodal reasoning rerankers unlock substantial headroom up to +15.22 nDCG@10.

Browse all 17 VLM Reasoning papers →


⚡ VLM Efficiency (1)

Learning to Zoom Efficiently with a Contrastive Curriculum

Addressing the heavy reliance of multimodal large language models on expensive human annotations and trajectory-level warm-start supervised fine-tuning (SFT) for acquiring zoom-in tool capabilities, this work presents a self-supervised intrinsic exploration reward formulated as an InfoNCE-style difference over subsequent reasoning logits coupled with an IoU-guided negative bounding box curriculum, achieving Pareto-optimal accuracy on high-resolution VQA benchmarks with just a single tool call.


🎵 Audio & Speech (31)

A Family Affair: Counterintuitive transfer effects across speech-language models

Challenging the long-standing dogma that closely related languages provide the best transfer donor in low-resource speech recognition, this paper demonstrates on seven Formosan endangered languages that two-stage fine-tuning with phylogenetically unrelated donors yields significantly larger WER reductions than related donors, frequently outperforming models trained with doubled in-language data.

Audio-Reward-R1: Reasoning Reward Model for Unified Audio Understanding and Generation

Audio-Reward-R1 is the first unified reasoning-based reward model for both Audio Understanding Preference (AUP) and Audio Generation Preference (AGP) tasks, combining two-stage fine-grained data curation with verifiable-reward GRPO reinforcement learning to achieve state-of-the-art results on Audio-Reward-Bench (87.90% AUP and 80.10% AGP), outperforming Gemini-3-pro-preview.

Better Retrieval, Worse Robustness: How Multi-hop RAG Amplifies Upstream ASR Errors

This paper empirically evaluates spoken multi-hop retrieval-augmented generation (RAG) pipelines across four English accents and three multi-hop QA benchmarks, revealing that while advanced extensions like entity-graph linking and iterative reformulation achieve higher clean-text F1, they amplify upstream ASR errors by expanding the clean-to-corrupted performance gap by 36% to 67%, with query-entity corruption identified as the primary catalyst of catastrophic retrieval breakdown.

Beyond Text Following: Repairable Arbitration Reversals in Audio-Language Models

Addressing the arbitration failure where audio-language models blindly follow conflicting text over clear acoustic evidence, this paper proposes Gated Audio Counterfactual Logit Correction (GACL), a training-free decoding rule grounded in causal activation patching that interpolates joint logits toward a same-audio reference branch to rescue audio decisions under strict faithfulness constraints.

BiMTokenizer: Preserving Semantic-Acoustic Balance in Low-Bitrate Speech Tokenization

BiMTokenizer challenges the recent architectural bloat of dual-tower speech codecs by revitalizing the single-tower paradigm with a bidirectional Mamba backbone and non-parametric Residual Spherical Leech Quantization (RSLQ), delivering state-of-the-art acoustic reconstruction and downstream semantic performance at ~1.1 kbps with only 253M parameters.

Bulbul: A Dataset for Dialectal Arabic Speech Recognition

BULBUL is a community-driven multi-dialect Arabic automatic speech recognition dataset spanning 275 native speakers across 11 Arab countries, providing 71.59 hours of speech covering 16 dialects/sub-dialects alongside accented Modern Standard Arabic and Classical Arabic subsets with two-level human verification.

CALM: Class-Conditional Sparse Attention Vectors for Large Audio-Language Models

CALM introduces a fine-tuning-free classifier that models internal attention heads of frozen large audio-language models as class-conditional experts, aggregating them via margin-weighted soft voting to achieve substantial gains over uniform voting across audio, audiovisual, and visual few-shot benchmarks.

CSAVocoder: A Causal Spatial Audio Vocoder Towards Real-Time Spatial Audio Generation

Addressing the failure of conventional neural vocoders to preserve multi-channel inter-channel acoustic cues and their non-causal streaming latency bottlenecks, CSAVocoder integrates an Attentional Mel Adaptor with dynamic pose FiLM modulation, a strictly causal stateful PixelShuffle generator, and an axial-attention Spatial Consistency Discriminator with format-aware physical losses, delivering superior spatial rendering fidelity alongside sub-real-time streaming efficiency (RTF = 0.1587).

DOA: Training-Free Decoder-Only Attention Policy for Long-Form Simultaneous Translation with SpeechLLMs

Addressing the absence of explicit cross-attention in decoder-only SpeechLLMs, DOA proposes a completely training-free streaming decision policy that slices a proxy cross-attention matrix directly from self-attention weights and pairs it with punctuation-aware history pruning, delivering offline-competitive quality and low latency on continuous long-form audio.

EmbodiChat: Towards Zero-Latency Text-driven Talking Face Generation for Embodied Dialogue

EmbodiChat resolves the trade-off between latency, lip synchronization, and hardware transferability in embodied dialogue by aligning facial animation directly to the intermediate acoustic latent space of two-stage TTS models without cross-modal adapters, delivering a ~150 ms Time-to-First-Audio on a consumer GPU and enabling robust expression transfer across heterogeneous humanoid robots via a Universal Facial Action Compiler (UFAC).

Browse all 31 Audio & Speech papers →


🧑 Human Understanding (3)

Beyond CTC Spikes: LLM-Driven Gloss-Temporal Supervision for Continuous Sign Language Recognition

Addressing the sparse spike phenomenon and weak visual-semantic correspondence caused by CTC loss in continuous sign language recognition, this paper introduces the Gloss-Temporal Module (GTM) to generate sustained gloss-to-frame interval supervision via LLM reasoning over multi-source video cues, substantially reducing WER on pure RGB baselines without extra parameters.

DiscoSign: Discourse-Aware Text to Sign Language Gloss Translation

DiscoSign is the first discourse-aware computational framework for text-to-sign language gloss translation, using explicit state registries and deterministic post-processing in a single LLM call to enforce spatial coreference, question-answer clause restructuring, and concept-gloss consistency across multi-sentence discourse.

sEMG Typing Transcription with Conformer Encoding and Beam-Aware Large Language Model Correction

This paper introduces a two-stage wrist-sEMG typing transcription framework that links a multi-scale bilateral Conformer CTC frontend with a fine-tuned Qwen3.5-9B corrector via beam-ordered candidate lists, slashing generic CER to 19.65% and personalized CER to 2.01% on the emg2qwerty benchmark.


📹 Video Understanding (17)

CapRiCorn-1K: A Comprehensive Benchmark for Video Captioning and Subject Referential Consistency Across Temporal Scales

Addressing the critical limitations of existing video captioning benchmarks—short durations, static scenes, and neglect of identity tracking—CapRiCorn-1K introduces a 1,000-video benchmark spanning 15 seconds to 10 minutes with rich scene transitions and clothing variations, proposing a novel Rand-Index-inspired metric to evaluate subject referential consistency alongside fine-grained keypoint coverage across audiovisual and visual-only settings.

Caption-once, Frames-on-Demand: Visual-Need Routing for Budget-Aware Agentic Long Video Understanding

Caption-once, Frames-on-Demand (CFD) introduces a budget-aware edge-cloud agentic framework for long-video understanding that constructs reusable dual-track narrative memories offline (event skeleton plus clip micro-logs) and deploys an online Visual-Need Router to fetch bounded keyframes only when perceptual questions require direct pixel evidence, reducing per-query frame consumption to 5.8 frames on average while achieving competitive accuracy on Video-MME (67.5) and InfiniBench.

CausalChapter: Improving Long-Form Video Chaptering with Interventional Dependency Modeling

To tackle boundary error propagation and cross-segment context fragmentation in scalable segment-then-caption long-form video chaptering, CausalChapter introduces intervention-inspired predictive dependency modeling via Local Dependency Shift (LCDS) to detect smooth boundary transitions and Cross-Segment Support Selection (CSSE) to rerank historical contexts by output KL divergence shift, substantially outperforming prior state-of-the-art systems on AVLecture and VidChapters-7M.

CoFIE: Coarse-to-Fine Evidence Selection for Efficient Streaming Video Understanding

To tackle high front-end vision encoding costs and the limited latency gains of downstream token pruning in streaming video understanding, CoFiE decouples evidence selection into a pre-encoder coarse filter using grayscale histogram novelty and an LLM prefill fine ranker using text-to-vision attention, achieving up to a 2.54× end-to-end speedup while outperforming the full-frame baseline.

Emotion Understanding in Streaming Video with Trajectory-Aware Reliability

Addressing the vulnerability of isolated prefix predictions and deceptive single-prefix confidence in streaming video, this paper proposes TRACE, a trajectory-aware reliability framework that continuously tracks emotion belief trajectories from lightweight audio streams, evaluates decision risk via a two-head reliability model, and selectively triggers multimodal contextual reinterpretation, matching full-context performance while reducing reinterpretation calls to 55.58%.

HIPPO: Accelerating Video Large Language Model Inference via Holistic-aware Parallel Speculative Decoding

Targeting the visual token redundancy and draft latency bottlenecks in video-LLM inference, HIPPO integrates a multi-dimensional semantic-aware token preservation strategy with an overlapping video parallel speculative decoding algorithm, achieving up to 3.51× lossless inference speedup across diverse video benchmarks.

Long-to-Short Video Evidence Reasoning for Grounded Question Answering

Addressing temporal evidence duration imbalance and the optimization conflict between strict boundary overlap and answer reasoning in grounded video QA (GQA), this paper introduces LOVER, an RL post-training framework combining a long-to-short evidence curriculum, an evidence-spotting IoP reward, and adaptive timestamp rendering to achieve new state-of-the-art results among open-source models on NExT-GQA and ReXTime.

LongSpace: Exploring Long-Horizon Spatial Memory from Perception to Recall in Video

To tackle the challenge of sparse, scattered spatial cues being diluted by redundant visual tokens in long videos, this paper proposes LongSpace—a unified framework combining 3D Spatial Structure Perception (SSP) and Hierarchical KV Memory (HKM)—alongside LongSpace-Bench, a benchmark spanning seconds to hours of indoor room tours, setting strong state-of-the-art results in long-horizon spatial recall.

MemDreamer: Decoupling Perception and Reasoning for Long Video Understanding via Agentic Hierarchical Graph Memory

MemDreamer decouples long-video understanding into streaming construction of a pure-text Hierarchical Graph Memory and tool-augmented agentic retrieval via an Observation-Reason-Action loop, establishing new SOTA across four benchmarks with a 40–124× smaller reasoning context.

MemoryCard: Topic-Aware Multi-Modal Clue Compression for Long-Video Question Answering

Addressing the challenge that key evidence in long videos is sparse, brief, and temporally scattered while isolated frames lack coherent semantics, MemoryCard segments videos into topic-coherent sessions via VLM self-reading, renders event gists, speech transcripts, and representative frames into self-contained "Memory Cards", and performs question-conditioned retrieval with adaptive resolution allocation, boosting QA accuracy by up to 21.8% under comparable visual token budgets.

Browse all 17 Video Understanding papers →


🚗 Autonomous Driving (1)

PlannerForge: LLM Agents for Scenario-Based Testing of Motion Planners in Autonomous Driving

PlannerForge introduces the first unified LLM-agent framework spanning the entire lifecycle of scenario-based testing for autonomous vehicle motion planners, combining router dispatch, OpenStreetMap synthesis, layered retrieval, SUMO-grounded physical modifications, and zero-shot cost tuning.


🤖 Robotics & Embodied AI (7)

Inferring the Unspoken: Aligning Embodied Agents with Implicit Preferences

Addressing the failure of embodied agents to satisfy idiosyncratic user preferences under underspecified natural language commands, this paper introduces the 290-preference PBP benchmark and the InTU framework, demonstrating that verbalizing latent preferences prior to planning cuts alignment error roughly in half and localizes the bottleneck to preference acquisition rather than downstream execution.

IntentVLA: Short-Horizon Intent Modeling for Aliased Robot Manipulation

Addressing the issue where current-frame observation aliasing under partial observability causes chunked policies to resample conflicting intents across adjacent steps, IntentVLA extracts compact camera and register tokens from recent visual history via a frozen VGGT encoder, injects them into the Qwen3-VL backbone via gated cross-attention along with a global intent summary token, and conditions a DiT flow-matching action head, substantially improving rollout stability and task success across AliasBench, SimplerEnv, LIBERO, and real-world manipulation.

LIBERO-Para: A Diagnostic Benchmark and Metrics for Paraphrase Robustness in VLA Models

This paper introduces LIBERO-Para, the first controlled diagnostic benchmark that independently decouples action and object expressions to evaluate paraphrase robustness in Vision-Language-Action (VLA) models, alongside PRIDE, a difficulty-grounded deviation metric. Across seven diverse VLA configurations, the study exposes a pervasive 22–52 percentage point performance drop, demonstrates that object lexical variation acts as the primary bottleneck, and proves via DTW trajectory analysis that 80–96% of failures stem from planning-level task identification divergence rather than motor control execution.

LookStep: Efficient Vision-Language Navigation with Linguistic Foresight and Event Driven Memory

Addressing the low data efficiency of next-step action imitation and the memory explosion from accumulating visual histories in continuous vision-language navigation, LookStep introduces linguistic foresight for candidate action consequences and an event-driven rolling episodic memory, achieving a 49.7% success rate on R2R-CE using only monocular RGB while cutting peak GPU memory by over 55%.

MineExplorer: Evaluating Open-World Exploration of MLLM Agents in Minecraft

Addressing the entanglement of game-specific mechanics and short-horizon objectives in existing benchmarks, MineExplorer establishes a knowledge-controlled open-world exploration benchmark in Minecraft with ReAct-style implicit multi-hop task graphs and multi-agent scene synthesis, revealing severe degradation in frontier MLLMs when coordinating hidden prerequisites.

Mini-BEHAVIOR-Gran: Revealing U-Shaped Effects of Instruction Granularity on Language-Guided Embodied Agents

Addressing the lack of cross-task quantification and controlled evaluation for instruction granularity in language-guided embodied AI, this paper introduces MINI-BEHAVIOR-GRAN, formalizes planning-width as a robust cross-task metric, and reveals a non-monotonic U-shaped performance pattern across VLA architectures driven by vision-dominant shallow grounding under coarse guidance.

when search becomes memory accelerating robot design discovery with self evolvin

Addressing the memoryless limitation of traditional robot morphology search where costly simulation traces are discarded after each iteration, this paper introduces AUTO-ROBOTIST, a self-evolving LLM agent that distills evaluation traces into an explicit natural-language skill library (featuring archetypes, positive/negative rules, and empirical observations), accelerating cold-start discovery and enabling effective cross-scale transfer from \(5\times 5\) to \(10\times 10\) design spaces.


🎮 Reinforcement Learning (7)

Beyond SFT-Optimality: Cold-Start Data Selection for Reinforcement Learning via Gradient Subspace Coverage

Addressing the empirical rank inversion where SFT-optimal data subsets fail to translate into effective downstream reinforcement learning, this paper introduces RL-Proxy-Coverage, which constructs a low-rank reward-relevant gradient subspace from RL anchor prompts and selects SFT candidates via greedy log-determinant coverage to maximize early RL sample efficiency.

EDGE: Experience-Distillation for Guided Exploration in Agentic Reinforcement Learning

EDGE reframes retrieved external experiences from persistent inference-time retrieval dependencies into temporary training-time scaffolds, estimating online marginal gains without extra sampling and distilling beneficial exploration behaviors into the parametric policy via reverse-KL divergence.

Expanding the Solvable Frontier: Policy-Adaptive Curriculum Learning for Reasoning RL

Addressing the mismatch between static teacher-assigned difficulty and the evolving capability of reasoning models, PAD-Curriculum adaptively delineates the student's solvable frontier using greedy self-generated trajectories, ordering currently solved problems via normalized Scorer-to-Student KL discrepancy without token-level imitation, and substantially boosts mathematical reasoning under GRPO.

HIPO: Instruction Hierarchy via Constrained Reinforcement Learning

HIPO establishes an alignment framework that casts hierarchical instruction following (HIF) as a Constrained Markov Decision Process (CMDP), using decoupled dual evaluators and a group-relative primal-dual safe RL algorithm to maximize user utility strictly within the feasible region of system prompt compliance.

How Much Supervised Fine-Tuning Is Enough? Dissecting the Impact of SFT Extent on Downstream Reinforcement Learning Performance

This paper dissects how the extent of Supervised Fine-Tuning (SFT) governs downstream Reinforcement Learning (RL) efficacy, proposing the Rigidity Index and an accuracy-rigidity inflection point criterion to pinpoint the optimal checkpoint for transitioning from imitation to exploration without irreversible distribution collapse.

Long Document Reconstruction Unlocks Scalable Long-Context RLVR

This paper introduces an unsupervised Reinforcement Learning with Verifiable Rewards (RLVR) framework that masks paragraphs in raw long documents and trains LLMs via GRPO to reconstruct their original order, deriving objective, automated verifiable rewards and complexity curricula directly from intrinsic text structures without any human QA annotation or teacher model distillation.

Source-Grounded Semantic Reinforcement Learning for Low-Resource Target-Language Generation

Addressing the scarcity of parallel data in low-resource target-language generation, SG-SRL introduces a train-reinforce-recover framework that turns abundant high-resource source monolingual corpora into reference-free RL supervision via a cross-lingual reranker reward, followed by lightweight supervised recovery to eliminate verbosity-induced reward hacking.


🎁 Recommender Systems (4)

Illuminating the Path: Reachability-Aware Supervision Curation for LLM-Based Recommendation

GEARec overcomes context-utilization bias and severe exploration collapse in LLM-based sequential recommendation through a reachability-aware data curation pipeline—extracting dual-perspective anchors, filtering unreachable reasoning trajectories via full-pool retrieval, and aligning with dynamic fine-tuning—achieving state-of-the-art accuracy across full-item candidate pools and boosting long-tail recommendation.

LLaDA-Rec: Discrete Diffusion for Parallel Semantic ID Generation in Generative Recommendation

Addressing the core weaknesses of autoregressive generative recommendation—unidirectional constraints that undermine global representation and early token mistakes that cascade along the decoding sequence—LLaDA-Rec introduces discrete diffusion for parallel semantic ID generation via Multi-Head VQ-VAE tokenization, dual-level masking, and adaptive-order beam search.

Re2A: Situated Conversational Recommendation via Rubric-based Preference Reasoning and Alignment

Re2A formulates situated conversational recommendation as a structured reason-then-align process, using automated rubric-guided policy optimization to infer explicit preference states and heterogeneous preference-conditioned DPO to resolve the dual alignment dilemma between user needs and physical scene constraints.

SafeGEO: Understanding Generative Engine Optimization Risks in Recommendation Agents

SafeGEO reframes Generative Engine Optimization (GEO) from visibility hacking into a utility-alignment failure for recommendation agents, evaluating 22 attack variants across 600 cases to show that realistic seller rewrites inflate flawed product recommendations by up to 83.2 pp while pre-ranking evidence breakdown mitigates this by up to 39.2 pp.


🔄 Self-Supervised Learning (2)

CWoMP: Morpheme Representation Learning for Interlinear Glossing

CWoMP reformulates interlinear glossing from unconstrained character sequence generation into generative retrieval over a mutable lexicon of atomic form-meaning morpheme units, eliminating morpheme hallucinations, enforcing strict 1:1 segmentation-gloss alignment, reducing inference FLOPs by 8.5x, and allowing zero-retraining inference-time vocabulary expansion.

Unified Work Embeddings: Contrastive Learning of a Bidirectional Multi-Task Ranker

To overcome task fragmentation and extreme label cardinality in labor market intelligence, this paper introduces the WorkBench ranking evaluation suite and Unified Work Embeddings (UWE)—a 109M-parameter bi-encoder that leverages bipartite many-to-many InfoNCE loss and bidirectional soft late interaction, surpassing 8B generalist models while exhibiting strong positive cross-task transfer.


📐 Optimization & Theory (1)

FormuEvo: LLM-Guided Evolution for Discovering Solver-Efficient Mixed-Integer Programming Formulations

FormuEvo reframes mixed-integer programming (MIP) formulation design as an LLM-guided evolutionary search over executable modeling programs, leveraging fine-grained solver statistics as verbal gradients alongside structured memory distillation to achieve up to 5.5x solver speedup over expert designs and prior LLM approaches.


🔗 Causal Inference (1)

From Text to Structure: The PACKaGE Benchmark for Bidirectional Causal Reasoning in Paragraphs and Graphs

Introduces PACKaGE, the first paragraph-level bidirectional causal reasoning benchmark comprising 4,566 fact-grounded text-DAG pairs, systematically uncovering a fundamental capability dissociation between causal extraction and text generation in modern LLMs.


🔬 Interpretability (19)

Attribute-Based Activation Steering of LLMs for Group-Specific Explanation Generation

This paper introduces an unsupervised, training-free attribute-based activation steering approach that extracts style and knowledge attributes of target groups from community Q&A texts and injects weighted activation vectors at inference time, achieving significantly improved group specificity while preserving factual accuracy.

Beyond Decodability: Reconstructing Language Model Representations with an Encoding Probe

Addressing the core limitations of classic decoding probes—namely non-comparable cross-feature metrics and vulnerability to feature collinearity—this paper introduces an Encoding Probe that reverses the probing direction to reconstruct internal model activations using multivariate ridge regression, enabling direct feature contribution quantification on a shared unexplained variance (UV) scale.

Confidence Before Answering: A Paradigm Shift for Efficient LLM Uncertainty Estimation

Pioneers a confidence-first paradigm and proposes CoCA, an end-to-end GRPO framework that dynamically grounds confidence targets in group-wise empirical success rates and decouples optimization via segmented credit assignment, reducing confidence inference token costs by over 92% while maintaining superior calibration.

Dissecting Transformers: A CLEAR Perspective Towards Green AI

Addressing the severe temporal mismatch between microsecond-scale component execution and millisecond-scale hardware energy sensors, this work introduces CLEAR—a three-stage profiling framework using activation caching and repetition amplification—to conduct the first comprehensive component-level inference energy dissection across 15 models and 4 Transformer architectures.

Diverging Transformer Predictions for Human Sentence Processing: A Comprehensive Analysis of Agreement Attraction Effects

Evaluating eleven autoregressive Transformers across comprehensive agreement attraction paradigms using surprisal theory reveals that while models align with human reading times in prepositional phrase modifiers, their predictions diverge sharply and contradict human cognitive asymmetries in object-extracted relative clauses, challenging the view that Transformers are faithful cognitive models of human morphosyntax.

Do Compounds and Phrases Follow the Same Combinatorial Principles? What a Multi-Dimensional Exploration Reveals

Constructs COMPPHRASE-7, a parallel Chinese compound-phrase benchmark across seven construction types, and establishes a three-dimensional probing protocol disentangling feature transfer, category geometry, and causal specificity to demonstrate that Chinese PLMs encode partially transferable construction cues while category geometry is layer-sensitive and localized causal circuits are absent.

Does Fine-Tuning Undo Activation Steering? Behavioural Recovery Without Weight-Edit Reversal

This paper investigates whether embedding activation steering directly into model weights survives downstream non-adversarial fine-tuning (SFT and RLHF), demonstrating that while fine-tuning can significantly restore steered behaviors (e.g., recovering 69% of ablated refusal under SFT), it reverses less than 0.8% of the weight edit, as the model routes around the intact edit via near-orthogonal weight updates.

From Profiles to Steering Vectors: Global Sparse Priors and Local Semantic Calibration for Interpretable LLM Steering

Addressing semantic residue and context overfitting in retrieval- and activation-based personalization, this paper proposes GLASS, a training-free framework that extracts global style sparse priors via Sparse Autoencoders and constructs local contrastive vectors across scenario clusters, enabling lightweight, interpretable, dual-stream activation steering during inference.

LSC-Pattern: Interpretable Lexical Semantic Change Analysis via Diachronic Sense Distributions

Addressing the black-box nature of conventional lexical semantic change detection that outputs only uninterpretable scalar change scores, LSC-Pattern connects LLM-based dictionary-grounded word sense disambiguation with multi-dimensional semantic properties (valence, semantic scope, metaphoricity), achieving high correlation with human judgments and transparent linguistic explanations across 150 years of Japanese corpora.

Observable Patterns Are Not Explanations: A Causal-Geometric Analysis of Latent Reasoning Models

Challenging the widespread assumption that observable latent-state patterns (such as BFS tree search and scratchpad computation) constitute evidence for internal reasoning mechanisms, this paper uses matched controls, causal tracing, and gradient-subspace interventions to show that these patterns are epiphenomena rather than mechanisms, proving that latent-thought influence is graded and concentrated within low-dimensional geometric subspaces.

Browse all 19 Interpretability papers →


📦 Model Compression (85)

A^3: Attention-Aware Accurate KV Cache Fusion for Fast Large Language Model Serving

To tackle accuracy degradation in cross-request KV cache reuse caused by position mismatch and missing cross-attention, \(A^3\) leverages shallow query-aware attention to guide selective token recomputation and pairs with adaptive token eviction, achieving a 2× reduction in time-to-first-token (TTFT) with near-lossless accuracy.

A Target-Centric Survey of Quantization-Aware Training

Addressing the limitation of prior quantization surveys that overlook the distinct numerical properties and lifecycles of different tensor targets, this paper presents the first target-centric survey of Quantization-Aware Training (QAT), deconstructing weight, activation, joint W+A, KV cache, and gradient quantization to formulate cross-target transferability rules, empirical evaluation protocols, and a rigorous minimal reporting checklist for real hardware deployment.

Accelerating Neural Architecture Search with Code Semantics

This paper introduces ECO-SCE, the first code-semantic transition estimator for neural architecture search that models parent-to-proposal evolutionary utility via semantic shifts, structural edits, and local evidence calibration, filtering low-gain code rewrites prior to costly evaluation and cutting downstream NAS GPU hours by 70.88% zero-shot.

AdaCode: Adaptive Lossless Compression for BFloat16 KV Cache

Addressing the severe memory bottleneck of KV caches in long-context generation without suffering from the approximation errors of lossy quantization, this paper reveals that BFloat16 exponent values are inherently concentrated and proposes AdaCode, a bit-exact lossless compression framework that encodes frequent exponents into short fixed-length codes while retaining outliers explicitly, achieving an average 71.78% compression ratio across diverse LLMs and 70.26% on vision-language models with zero accuracy loss.

Adaptive Layer Allocation for Efficient Transformer Fine-Tuning

Grounded in Moral Foundations Theory (MFT), this work uncovers a cross-lingual shared yet culturally nuanced moral subspace in LLMs, extracts causally steerable Moral Vectors, and introduces Adaptive Moral Fusion (AMF) to dramatically slash HarmBench jailbreak rates while minimizing XSTest over-refusals without degrading core utility.

Rethinking Stepwise Model Routing: A Cost-Efficient Table Reasoning Perspective

EcoTab presents a table-aware stepwise model routing framework that decouples table structure tokens from natural language text tokens via a word-level Trie, independently maps their uncertainties to next-step failure risks using offline counterfactual suffix replacement, and fuses the risks via Noisy-OR logic to guide small/large model selection without generating extra verification tokens.

HLoRA: Hierarchical Regularization on Combined Parameter Space for LoRA Fine-Tuning

Addressing the failure of conventional continual learning regularizers under LoRA's factorization ambiguity, HLoRA formulates path-integral importance directly over the combined weight space \(\Delta W = BA\) alongside element-wise and layer-wise hierarchical regularization, preventing catastrophic forgetting while speeding up importance calculation by ~20x and reducing storage by 85%–90%.

CulturePRM: A Process Reward Model for Mitigating Cultural Overriding in Cultural Reasoning

To tackle cultural overriding—a prevalent failure mode where reasoning models acknowledge local cultural norms but subsequently erode or substitute them with universal values or Western defaults—this paper introduces CulturePRM, a lightweight 1B process reward model that scores intermediate reasoning steps to guide Best-of-N reranking, yielding substantial accuracy gains and narrowing cross-cultural disparities.

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 due to their scarce attention heads causing premature subspace overlap, LightLoRA introduces a parameter-free sparse fine-tuning framework that scores attention heads via SVD parameter variation and directional gradient conflict penalties, selectively updating high-contribution orthogonal heads while preserving historical knowledge.

Popular but Wrong: Understanding and Mitigating LLM Overconfidence through Knowledge Popularity

This paper investigates the root cause of LLM overconfidence in factual question answering through the lens of knowledge popularity, revealing that hallucinated answers systematically drift toward globally popular or frequently co-occurring entities, and introduces a lightweight popularity-aware MLP calibrator that slashes average confidence on incorrect answers from 0.765 to 0.254 while reducing Expected Calibration Error (ECE) from 0.356 to 0.050.

Browse all 85 Model Compression papers →


🕸️ Graph Learning (10)

Dual-Prompt-Driven Textual Knowledge Graph Generation With Heterogeneous Information Embedding for CTR Prediction

Addressing the severe semantic loss of discrete feature encodings and the lack of fine-grained cross-modal alignment in CTR prediction, TKE4CTR designs a dual-prompt paradigm to transform tabular records into textual narratives and hallucination-mitigated local knowledge subgraphs, followed by bidirectional cross-attention gated fusion and a feature-gated MLP that achieve substantial AUC improvements across five diverse Amazon domains.

FBV-KGQA: Forward-Backward Verification for Knowledge Graph Question Answering

FBV-KGQA proposes a forward-backward verification framework for knowledge graph question answering that generates relation-level plans, quantifies the causal attribution of critical entity/relation pivots via schema-constrained counterfactual perturbations, and checks reverse question consistency to filter out superficially plausible candidate paths.

GraphLit: Learning Text-Enriched Dynamic Character Network Representations for Literary Study

GraphLit introduces Dynamic Heterogeneous Character Networks (DHCNs) to segment long novels into temporally localized heterogeneous graphs that align characters with their contextual text segments, training an 8.2M-parameter masked graph autoencoder with block ordering and global link prediction to systematically outperform text-only, graph-only, and prior hybrid baselines across 12 character-centric tasks and quotation attribution.

LLM-Powered Holistic Enhancement for Text-Attributed Graphs

Proposes EFS (Enhancing Features and Structures), a unified framework that enriches node features using LLM-extracted keywords, ranked category predictions, and explanatory contexts, while refining the graph topology with a lightweight edge refiner without requiring any per-edge LLM API calls, achieving state-of-the-art node classification across six benchmark datasets.

Logic Meets Language: Hybrid Query Answering over Incomplete Text-Labeled Graphs

Addressing the challenge of answering queries requiring both discrete logical constraints and flexible textual semantics over incomplete text-labeled graphs, this paper formalizes a hybrid query language and introduces HybridQE, an end-to-end framework that unifies query-conditioned entity representations with neural logical operators in a shared embedding space.

Ontology-Driven Structural Regularization for Document-Level Relation Extraction

Addressing pervasive and previously overlooked structural noise in distantly supervised Document-Level Relation Extraction (DocRE)—such as ontology constraint violations, missing inverse relations, and asymmetric contradictions—this paper formalizes structural well-formedness under OWL axioms and demonstrates that lightweight, model-agnostic data-centric structural regularization drastically eliminates logical contradictions in model predictions while consistently improving generalization performance.

PathSoft: Hybrid Soft Prompting with Ranked Reasoning Paths for Knowledge Graph Question Answering

PathSoft introduces a hybrid soft prompting framework for knowledge graph question answering that filters candidate reasoning chains with a lightweight PathMLP and encodes the ordered entity–relation sequences into continuous soft prompts via PathSoftAdapter, delivering state-of-the-art accuracy on frozen LLMs with single-call, ultra-low latency inference.

StruProKGR: A Structural and Probabilistic Framework for Sparse Knowledge Graph Reasoning

To tackle computational bottlenecks in random walk path collection and the isolated treatment of paths in sparse knowledge graph reasoning, StruProKGR introduces a training-free framework that leverages truncated distance-guided DFS pruning for up to 54.93x path collection speedup and incorporates intra-path diminishing returns and inter-path Bayesian odds updates to capture structural dependencies, consistently outperforming existing path-based and rule-based baselines across five sparse benchmarks.

TATK: Triple-Aware Top-K Learning with Knowledge-Grounded Verification for LLM-based Sequential Recommendation

Addressing the key mismatches in LLM-based sequential recommendation—weak context grounding, misalignment between generation loss and ranking metrics, and unverified candidate lists—TATK introduces a Triple-Aware framework that combines metadata-KG prompt grounding and position-aware Top-K rewards during training with single-pass knowledge-grounded verification at inference time, substantially boosting full-catalog Top-K ranking performance.

TKG-Thinker: Towards Dynamic Reasoning over Temporal Knowledge Graphs via Agentic Reinforcement Learning

TKG-Thinker addresses reasoning hallucinations and rigid static workflows in temporal knowledge graph question answering (TKGQA) by formulating reasoning as a sequential decision process, equipping LLMs with temporal tool primitives via a two-stage optimization pipeline combining rejection-sampled SFT cold-start and online reinforcement learning with multi-dimensional verifiable rewards.


🏥 Medical Imaging (1)

LENS: Exploiting the Cross-View Stability Gap for Hallucination Mitigation in Medical Vision–Language Models

This paper identifies an empirical cross-view stability gap where visually grounded clinical tokens maintain stable logits under semantics-preserving perturbations while hallucinated tokens diverge, proposing LENS to exploit this signal jointly via Semantics-Preserving Preference Optimization (SPPO) and stability-aware contrastive decoding.


🩺 Medical LLM (47)

AP-GRPO: Anchor-Gated Phonetic Alignment with Policy Optimization for Pathological Speech Reconstruction

Addressing acoustic degradation and severe hallucinations when speech language models transcribe dysarthric recordings, AP-GRPO introduces a group relative policy optimization framework that preserves reliable audible anchor words and scores corrupted inter-anchor spans via soft-DTW phonetic alignment, slashing word error rate in severe ALS from 0.75 to 0.29.

auditing evidence use in medical llm diagnosis

Addressing the limitation where diagnostic accuracy hides improper clinical reasoning, this paper presents a behavioral evidence-use audit via controlled subset interventions and diagnostic margin interaction mining, establishing a diagnosis-relative framework with multi-stage stability filtering to isolate true evidence failures.

Can MLLMs Perceive and Describe Medical Image Quality Beyond Scores?

MedQ-Bench establishes the first comprehensive multimodal benchmark for language-based medical image quality assessment across 5 modalities, 40+ degradation types, and 3,308 images. Evaluating 23 frontier MLLMs reveals substantial gaps against human experts, identifying severe performance drops in mild degradations and highlighting a pervasive diagnostic bias where medical-specialized models overlook catastrophic imaging artifacts.

PsyCLIENT: Client Simulation via Conversational Trajectory Modeling for Both Trainee Practice and Model Evaluation in Mental Health Counseling

PsyCLIENT decouples static client profiles from dynamic conversational trajectories, conditioning LLMs on 12 discrete client behavior labels and de-identified reference utterances alongside coherence soft constraints, achieving higher expert-rated authenticity and clinical training utility across 120 detailed Chinese client profiles spanning 60 counseling topics.

ClinHallu: A Benchmark for Diagnosing Stage-Wise Hallucinations in Medical MLLM Reasoning

ClinHallu introduces a benchmark of 7,031 validated multimodal medical VQA instances augmented with structured 3-stage traces (Visual Recognition, Knowledge Recall, Reasoning Integration) and stage-replacement counterfactual interventions, pinpointing task-specific cognitive bottlenecks and demonstrating that full trace supervision significantly mitigates stage-wise hallucinations.

Cognitive-Level Prefix-MoE Reward Modeling for Curriculum Alignment of Medical Language Models

To resolve the inherent reward conflict between exam-style multiple-choice questions and complex open-ended clinical reasoning, this paper proposes Prefix-MoE RM—a cognitive-level reward model that softly fuses trainable prefix embeddings across Bloom-inspired levels via a frozen router before scalar scoring—paired with prompt-type dispatch and four-stage curriculum GRPO.

Criterion-Activated Preference Synthesis: Attributed Alignment for Medical Dialogue Agents

This paper introduces Criterion-Activated Preference Synthesis (CAPS), a criterion-first preference synthesis framework that seeds context generation with explicit quality standards, decomposes evaluations into rubric-bound and general blind judges, and applies group-level Empirical Activation Tests (EAT) and pair-level Attribution Ratio (AR) filters to isolate clean, causal preference signals—outperforming a 22k SFT baseline and matching 35k human preference pairs on HealthBench with only 2k synthesized pairs.

CRU-Reason: Knowledge Graph-Grounded and Interpretable Clinical Prediction with Clinical Rationale Units

CRU-Reason integrates machine learning priors and constrained knowledge graph exploration to construct disease-centered Clinical Rationale Units (CRUs), guiding large language models to conduct fine-grained, decoupled clinical reasoning and delivering a 9.4% average improvement over strong baselines in drug recommendation and diagnosis prediction.

CRU-Reason: Knowledge Graph-Grounded and Interpretable Clinical Prediction with Clinical Rationale Units

CRU-Reason synergizes machine learning sequential statistical priors, medical knowledge graphs, and large language models by decomposing complex electronic health records into disease-centric Clinical Rationale Units (CRUs) via constrained graph exploration, driving fine-grained evidence-grounded clinical reasoning and outperforming top EHR baselines by an average of 9.4% on MIMIC-III and MIMIC-IV.

Dementia-R1: Reinforced Pretraining and Reasoning from Unstructured Clinical Notes for Real-World Dementia Prognosis

To address the challenge of reasoning over complex, non-monotonic disease progression trajectories under sparse delayed rewards in unstructured clinical notes, this paper presents Dementia-R1, a two-stage reinforcement learning framework that cold-starts via GRPO on verifiable intermediate clinical cognitive indices and fine-tunes on binary prognosis, achieving 84.02% AUROC on real-world clinical notes and outperforming models up to 10× larger.

Browse all 47 Medical LLM papers →


⚛️ Physics & Scientific Computing (1)

MDForge: Agentic Molecular Dynamics Pipeline Design under Sparse Simulator Feedback

MDForge frames molecular dynamics (MD) pipeline design as open-ended code generation driven by verbal reinforcement learning (Verbal RL), utilizing a multi-expert PRISM arbitration debate to densify sparse, expensive simulation feedback into stage-wise diagnostics and localized code patches, matching human expert pipelines on SAMPL benchmarks and prospectively discovering a picomolar binder verified by wet-lab competition NMR.


👥 Social Computing (39)

Affective Context Amplifies Sycophancy in LLM Responses

Drawing on social psychological ingratiation theory, this paper evaluates LLM sycophancy by contrasting independent third-party judgments against user-facing responses, demonstrating that models systematically soften or withhold negative feedback when facing users, and that user affective context—particularly negative vulnerability states like loneliness and distress—substantially amplifies this divergence while driving models into non-committal evasive sycophancy.

Beyond Good Intentions: When Does the Framing of Multilingual and Low-Resource NLP Research Become a Caricature?

This paper systematically examines how research framings in multilingual and low-resource NLP frequently disconnect technical outputs from grand claims of community empowerment and decolonisation, offering an epistemic analytical framework, paired empirical comparisons, a 25-year ACL bibliometric audit, and a four-pillar accountability checklist.

CogniDir: Combating Cognitive Malicious Comments via Adaptive Distributional Learning for Robust Fake News Detection

CogniDir reformulates robust fake news detection against LLM-synthesized cognitive malicious comments as a dynamic data-mixture optimization problem, adaptively balancing training exposure across heterogeneous attack mechanisms via an information-theoretic vulnerability metric and Dirichlet-mean resampling.

ContraGAND: Auditing and Repairing Gender Ambiguity Failures in LLMs with Neuro-Symbolic Counterfactuals

Addressing the vulnerability where machine translation and large language models fall back onto occupational gender stereotypes when source inputs lack gender cues, this paper builds CONTRAGAND, a 14,556-instance neurosymbolic contrastive benchmark, reveals an alarming performance drop of up to 69.1 percentage points when transitioning from closed to open referent evaluation, and demonstrates that contrastive fine-tuning paired with 31B teacher distillation boosts a 4B student to 0.976 accuracy, matching the frontier teacher at one-eighth the parameter count.

counter stereotypic evidence is not enough unequal uptake in llms

This paper introduces a controlled sequential evaluation framework to audit how LLMs take up counter-stereotypic context, revealing that evidence successfully incorporated in explicit forced-choice judgements largely breaks down or reverses in downstream free-text generation, with Asian and Hispanic female targets facing persistent under-correction.

Demographic Prompting at Scale: When More Attributes Hurt LLM–Human Agreement

By systematically enumerating all demographic attribute combinations across 5 subjective NLP tasks and 5 open-source LLMs, this paper shows that full-attribute prompts never yield optimal alignment and frequently degrade performance, proposing a three-level diagnostic framework based on word-demographic learnability and Fisher directional coherence alongside mechanistic neuron probing.

Detection is Not the End: Personalized Modeling of User Susceptibility to Misinformation

Addressing the fundamental limitation that conventional fact-checking focuses solely on objective truth detection while ignoring cognitive heterogeneity and early-stage evidence scarcity, this paper formalizes the Personalized Misinformation Susceptibility Prediction (MiSP) task, constructs multi-source survey and social-platform benchmarks across four seen/unseen generalization settings, and proposes VulneraSpotter—a graph neural network integrating dual-path heterogeneous interaction graphs, masked semantic homogeneous graphs, and self-supervised cross-modal alignment.

Do Large Language Models Adapt Politeness Like Humans?

By designing an artificial-language repeated request game that decouples payoff feedback from community exposure alongside hierarchical Bayesian cognitive modeling, this work shows that both humans and LLMs balance payoff and exposure via a policy-mixture mechanism, but LLM baselines remain markedly more payoff-oriented than human baselines.

Emotion Experience, Expression, and Perception: Emotion Analysis on Multimodal Social Media Posts

This paper introduces Mult2EMo, the first multimodal social media emotion dataset (7,200 posts) jointly capturing author experience, post expression, and reader perception, demonstrating that understanding the triggering event is decisive for reconstructing authors' emotions, while text remains the dominant modality over images.

Enhancing LLMs in Predictive Political QA with Semi-Structured Data

Addressing the fundamental gap where predictive political question answering cannot be resolved via direct factual lookup, PSL introduces a dual-view framework that extracts fine-grained issue stances from semi-structured profiles in a semantic view, aggregates high-order interaction graph representations in a vector view, and aligns them via collaborative filtering to significantly enhance LLM decision forecasting.

Browse all 39 Social Computing papers →


📂 Others (2)

Dissecting Transformers: A CLEAR Perspective Towards Green AI

Addressing the temporal mismatch between microsecond-scale operator execution and millisecond-scale power sensors, this paper proposes CLEAR, a software-based fine-grained energy measurement pipeline that profiles individual Transformer components across 15 models, revealing that the Attention block consumes significantly more energy per FLOP than the overall model.

Train Smarter, Not Harder: Switching Signal-Guided Training in Active Learning

Addressing the conflict in active learning between computationally burdensome retraining and warm-started fine-tuning that loses plasticity and degrades calibration early on, HYBRIDAL introduces an adaptive schedule that monitors model stabilization signals (\(\Delta\alpha\) or \(\Delta\mathrm{Acc}\)) and irreversibly switches from retraining to fine-tuning, preserving non-inferior F1 while slashing up to 49% of training time and recovering substantial calibration fidelity.


📄 ai_safety (1)

LLM-Based Adversarial Persuasion Attacks on Fact-Checking Systems

This paper introduces Persuasion Injection Attacks against automated fact-checking (AFC) pipelines, leveraging generative LLMs to inject 15 label-invariant persuasion techniques into claims, and demonstrates through decoupled evaluation that worst-case attacks collapse classification accuracy to near-zero while manipulative wording simultaneously cripples both evidence retrieval and veracity classification.


📄 financial_nlp (2)

FinDeepForecast : A Live Multi-Agent System for Benchmarking Deep Research Agents in Financial Forecasting

FinDeepForecast introduces the first end-to-end, live multi-agent system that continuously generates contamination-free forward-looking financial forecasting tasks under strict temporal isolation, benchmarking 13 state-of-the-art models across 8 major economies and 1,314 firms over a ten-week evaluation horizon.

Summoning the Oracle to Slay It: Mitigating Look-Ahead Bias in Financial Backtesting with Large Language Models

Addressing parametric look-ahead bias caused by LLMs memorizing historical market movements during pre-training, this paper proposes FINCAD—an inference-time contrastive decoding scheme that elicits parametric memory via an adversarially optimized prior prompt and dynamically discounts it using entity- and date-adaptive confidence signals without retraining.


📄 information_extraction (9)

EMER-Guard: Protocol-Grounded Verification for LLM Emergency Instructions

Addressing the severe trade-off between life-threatening instruction errors and excessive refusal in emergency NLP, this paper constructs EMER-GuardBench—a 69,000-example counterfactual auditing benchmark with instruction-group-disjoint splits—and presents EGIS, an evidence-conditioned multi-task verifier that delivers 89.9% Harm F1 while eliminating benign-stress false positives entirely under the Severity-Weighted Miss Rate (SWMR) metric.

All Relations at Once: Efficient Open-World Tabular Relation Discovery with Hierarchy-Aware Reward Shaping

Addressing the core bottlenecks of closed-set label reliance, rigid binary penalties on near-miss predictions, and quadratic pair-wise inference latency in column relation extraction, this paper introduces an open-world reinforcement learning framework with hierarchy-aware reward shaping and span-level advantage assignment that generates all column-pair relations in a single structured response, markedly outperforming both domain-specific models and frontier LLMs.

SSR: Symbolic-Neural Soft Reasoning for Robust and Verifiable Thinking Chains

Addressing the catastrophic brittleness of deterministic symbolic solvers under imperfect LLM auto-formalization, SSR introduces confidence-weighted MaxSAT relaxation, ambiguity-resolving targeted retrieval, and bidirectional deduction tree generation, boosting multi-step reasoning accuracy and producing replayable, verifiable thinking chains.

IdioSteer: Controlling Figurative-vs-Literal Interpretation via Residual-Stream Activation Steering

Addressing large language models' deep-seated bias toward figurative interpretations of idioms during open-ended generation, this paper proposes residual-stream activation steering to achieve precise inference-time control: by injecting a cross-layer mean-difference vector at the prompt's final idiom token, models are systematically shifted toward literal continuations while maintaining fluency, with the steering direction transferring seamlessly to unseen idioms.

MelonBench: A Benchmark for Coded Entity Resolution over Noisy Web Evidence

Addressing the reliance of existing browsing benchmarks on explicit queries and clean authoritative sources, this paper presents MelonBench, a 665-item benchmark instantiated in Chinese entertainment blind-item discourse to evaluate search agents under intentional semantic obfuscation and noisy user-generated content, where the best search-augmented system achieves only 49.9% accuracy and reveals critical bottlenecks in semantic de-obfuscation and open-ended entity recall.

ObSchema: From Local Observations to Global Entity Attribute Schemas in Scientific Literature

Addressing scattered, heterogeneous, and condition-entangled entity properties in scientific literature, ObSchema introduces a batch-wise observation-to-schema framework that decouples semantic classification from deterministic evidence-based promotion, integrating value-aware duplicate merging and adaptive grouping to induce high-fidelity hierarchical attribute schemas.

Personalization over Generalization: History-Aware User Modeling for Dialogue Agents

Existing sentence-level AI detectors treat all authors identically and fail to capture individual writing boundaries; this paper proposes PHAS-ID, a personalized detector that grounds sentence predictions in an author's labeled writing history via dual cross-attention fusion, alongside GhostFill, a style-aware synthetic generation pipeline mirroring real collaborative interleaving patterns.

Measuring Source-Resolution Distortion in LLM Output Arbitration: A Source-Bounded Evaluation Framework

Addressing the pervasive loss and misattribution of critical source structures when LLMs arbitrate among multiple candidate answers, this paper formalizes Source-Resolution Distortion (SRD), introduces a source-bounded benchmark, and proposes A2R—a framework that builds a sparse decision ledger before deterministic rendering to substantially cut conflict erasure and boost provenance recoverability.

The Harder Text Embedding Benchmark (HTEB): Beyond One-dimensional Static Robustness

Addressing the fundamental blind spots of static, single-scalar embedding leaderboards like MTEB, this paper introduces HTEB, a dynamic evaluation framework that stochastically injects 8 LLM-generated transformations along Lexical/Stylistic, Length, and Language axes, demonstrating that embedding robustness is multidimensional and that parameter scaling fails to close robustness gaps.