Skip to content

💡 LLM Reasoning

💬 EMNLP2026 · 601 paper notes

📌 Same area in other venues: 🧠 NeurIPS2026 (111) · 📷 CVPR2026 (16) · 🔬 ICLR2026 (241) · 💬 ACL2026 (82) · 🧪 ICML2026 (78) · 🤖 AAAI2026 (37)

🔥 Top topics: Reasoning ×122 · LLM ×104 · Multimodal/VLM ×31 · Agents ×31 · Alignment/RLHF ×25

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.

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

This work demonstrates that text-following errors in audio-language models during multimodal conflict stem from arbitration readout reversals rather than perceptual failures, and introduces GACL, a training-free gated counterfactual decoding rule that rescues audio-supported answers under strict faithfulness budgets.

LEXPATH: A Domain-Oriented Multi-Path Framework for Legal Article Retrieval

LEXPATH bridges the gap between lexical/semantic similarity and legal applicability by orchestrating an IRAC-guided sparse path, a hierarchy- and citation-aware dense path, and an intent-consistency reranking module, setting new state-of-the-art results across four comprehensive legal benchmarks.

CODA: Mitigating Over-Refusal and Jailbreaks in Large Language Models via Codebook-Orthogonal Direction Alignment

To resolve the critical tension between jailbreak vulnerability and benign over-refusal in inference-time activation steering, CODA projects hidden states onto a vector-quantized discrete codebook to extract stable semantic anchors, softly orthogonalizes the safety direction against benign technical intent, and dynamically modulates residual-stream intervention via risk-aware gating and exponential temporal decay without modifying model weights.

DeepSearch-World: Self-Distillation for Deep Search Agents in a Verifiable Environment

DeepSearch-World constructs a deterministic offline Wikipedia environment with reproducible tools and entity-level process verification, enabling the DeepSearch-Evolve framework to self-distill long-horizon web search and reflection behaviors into an open-source 9B agent without proprietary teacher models.

Hidden Threat in Synthetic Data: Covert Targeted Bias Injection through Benign Text

This paper demonstrates that targeted social biases can be covertly injected into aligned large language models via semantically benign synthetic datasets using subliminal learning, completely bypassing content safety filters while preserving general reasoning capabilities, and proposes Logit-Linear Selection (LLS) screening as an effective statistical defense.

TabSieve: Explicit In-Table Evidence Selection for Tabular Prediction

Addressing the lack of explicit evidence utilization and high vulnerability to noisy context in tabular in-context learning, TabSieve establishes a select-then-predict framework supported by the TabSieve-SFT-40K trajectory dataset and TAB-GRPO with dynamic task-advantage balancing, delivering significant performance gains across 75 classification and 52 regression tables.

Cube Bench: A Controlled Diagnostic for Spatial and Sequential Reasoning in Multimodal Large Language Models

This paper introduces Cube Bench, a fully controlled closed-loop diagnostic built on the Rubik's Cube with deterministic transitions and oracle distances, revealing that current MLLMs exhibit zero immediate error recovery (\(P(1)=0\)) and that strong single-step performance fails to translate into multi-step recovery.

Understanding the Effects of Distractors on Reasoning Vision-Language Models

This paper constructs Idis, a controlled multimodal benchmark of over 277k samples, discovering that visual distractors degrade reasoning VLM accuracy by vertically shifting down the length-accuracy curve without increasing reasoning length, whereas textual distractors lengthen reasoning traces and amplify test-time inverse scaling; causal analysis further reveals that the proportion of distractor attributes allocated during thinking directly drives reasoning collapse.

Why Act Without Asking? Evaluating Mobile GUI Agents under Ambiguous Instructions

Addressing the widespread "Act-without-Asking" failure mode where mobile GUI agents guess intents and proceed with actions under ambiguous instructions rather than asking for clarification, this work constructs MID-Bench—a step-level benchmark with strict user simulation and an Ask+Match protocol—and develops train-free confidence-gated cascades that boost clarification and execution accuracy without hurting standard action performance.

Predict, Don’t Iterate: Efficient Adaptive-Length Infilling for Diffusion Language Models

Addressing the sensitivity to initial length presets and substantial inference overhead in diffusion language model infilling, PILL freezes the backbone to directly predict target lengths via a single probing pass, decodes candidate spans in parallel using slot-wise attention and RoPE interpolation, and identifies the best span via single-pass causal scoring, requiring only two extra forward passes to outperform top baselines while running 1.82x faster.

StepGuard: Learning Step-Level Guardrails with Scalable Supervision and Safety–Utility Balancing

Addressing tool execution vulnerabilities and defense bias in autonomous LLM agents, StepGuard combines a prefix-aligned synthetic data engine (StepGen) with accuracy-gap-driven on-policy reinforcement learning (Balance-GRPO) to build a proactive 4B step-level guardrail, cutting attack success rates by 77.3% with only a 2.8-point sacrifice in benign task utility.

ANI: Adaptive Numerical Injection for Unifying Semantic and Arithmetic Representations in Numerical Reasoning

ANI introduces a plug-and-play adaptive numerical injection framework that identifies the layer of peak contextual maturity via linear probing, discriminates between nominal identifiers and quantitative operands via a context-aware gating network, and injects Fourier numerical topology exclusively at the terminal sub-token via an additive residual connection, boosting MATH by 9.5 points while preserving linguistic integrity.

OpenSkill: Open-World Self-Evolution for LLM Agents

Addressing the absence of curated skills and target-task feedback in realistic deployments, OpenSkill decouples open-world retrieval into domain knowledge acquisition and deterministic verification anchoring, synthesizes verifiable proxy pytest environments without target answers, and iteratively refines modular skill artifacts that transfer zero-shot across model families while attaining state-of-the-art pass rates.

InterPol: De-anonymizing LM Arena via Interpolated Preference Learning

This paper presents INTERPOL, a model-driven LLM identification framework that extracts target-specific response signatures from interpolated preference data by synthesizing controllable hard negatives via weight interpolation and training a detector with adaptive triplet preference learning, substantially outperforming statistical baselines and demonstrating that strategic voting can manipulate leaderboard rankings with up to 69.9% fewer interactions.

Language Chain in Alignment: Cross-lingual Ranking Preference Optimization

To alleviate language mismatch and degraded response quality in non-English alignment, CRPO reformulates cross-lingual alignment as a hierarchical ranking optimization problem via LambdaLoss, jointly modeling language consistency and generation quality over parallel response quadruplets to substantially outperform conventional binary DPO.

When In-Distribution Gains Fail: Evaluating Weak-to-Strong Reward Models under Preference Shift

This paper exposes that weak-to-strong reward models exhibit fragile out-of-distribution preference transfer despite high in-distribution gains, establishes a zero-shot domain-shift evaluation protocol with transfer-aware metrics, and proposes Representation Anchoring (ANCHOR) to regularize hidden representations toward a frozen pretrained reference.

CPR for LLMs: Critical-Point Routing against Catastrophic Forgetting in Domain Adaptation

Addressing the severe catastrophic forgetting of general reasoning capabilities during supervised fine-tuning (SFT), CPR (Critical-Point Routing) decouples the general base model from the domain expert at the model level, employing a lightweight hierarchical router that dynamically invokes the expert only on critical tokens where the base model fails, achieving a 1.4%–5.5% domain gain over vanilla SFT while suppressing general-capability degradation from up to 14.5% down to at most 0.5% with ~30% sparse expert invocation.

ToolForge: A Data Synthesis Pipeline for Multi-Hop, Multi-Turn, and Self-Reflective Tool-Use Data

ToolForge introduces an end-to-end data synthesis pipeline that constructs multi-hop, multi-turn, and self-reflective tool-use trajectories from QA seeds without live API calls, enabling Qwen3-8B fine-tuned on just 4,250 samples to improve by 21.6% over the TIR baseline across five out-of-domain QA benchmarks while demonstrating strong generalization on BFCL and API-Bank.

On the Robustness of Knowledge Editing for Detoxification

This paper presents a robustness-oriented evaluation framework for Knowledge-Editing-based (KE-based) detoxification in LLMs, revealing that reported defense gains are largely confounded by "pseudo-detoxification" (severe text repetition bypassing safety classifiers) and demonstrating that KE detoxification degrades sharply under joint multi-objective edits and cross-lingual transfer.

Adaptive Knowledge Routing for Complex Reasoning in LLMs

Addressing error propagation and local error reinforcement in revocable parallel decoding for masked diffusion LLMs, this paper proposes ASRD, a training-free embedding-space framework that leverages a temporal Anchor Tokens Cache to steer mask generation via entropy-weighted convex blending and verify tentative candidates via orthogonal residual probing, achieving up to a 6.4% accuracy gain and a 7.2x wall-clock speedup.

Towards Effective Structured Context Modeling for Conversational Recommender Systems via Dual-node Monte Carlo Tree Search

Addressing the vulnerability where unstructured free-form context induces severe preference tracking errors and noisy retrieval in conversational recommenders, this paper proposes DREAMS—a unified dual-node Monte Carlo Tree Search framework that explicitly navigates structured dialogue states for strategic preference elicitation (ELNode) and adaptive query refinement (EXNode).

ADORA: Training Reasoning Models with Dynamic Advantage Estimation on Reinforcement Learning

ADORA introduces a plug-and-play RL framework that dynamically calibrates advantage estimation weights based on live rollout statistics, categorizing training examples into Temporarily Advantageous Samples (TAS) and Temporarily Disadvantageous Samples (TDS) to suppress shallow shortcuts in weak models and amplify challenging deliberation in strong models, achieving state-of-the-art reasoning gains with exceptional sample efficiency.

Adaptive Knowledge Routing for Complex Reasoning in LLMs

This paper conducts systematic causal intervention and probing across four LLMs, revealing that propositional logical reasoning is not an opaque pattern-matching heuristic, but rather an internally organized system of four interlocking mechanisms: staged computation, boundary token aggregation, fact retrospection, and specialized attention heads.

LadderEdit: Edit-Level Residual Compression for Memory-Efficient Lifelong Editing of LLMs

LadderEdit introduces an edit-level residual compression controller for lifelong LLM editing that represents acquired exact LoRA updates along an adaptive rank ladder, utilizing a spectral-margin predictor and multi-dimensional behavioral auditing to reduce persistent storage by 5.2× to 12× while matching full-resolution accuracy up to 50,000 edits.

ToolLoop: Closed-Loop Tool-Use Data Synthesis via Decomposed Generation and Dynamic Self-Feedback

ToolLoop decomposes tool-use data synthesis into a three-stage pipeline (ground-truth function combination sampling, backward query derivation, and forward tool call instantiation) augmented with stage-specific dynamic self-feedback, achieving 86.40% non-reasoning accuracy on BFCL with only 11K synthetic examples.

LatentEvolve: Self-Evolving Test-Time Scaling in Latent Space

Inspired by Complementary Learning Systems (CLS) theory, LatentEvolve establishes a dual-phase latent test-time scaling framework that alternates between fast daytime episodic retrieval and slow nighttime parametric memory consolidation, empowering frozen large language models to continuously learn how to scale test-time compute in a fully unsupervised manner.

Can Retrieval Heads See Images? Multimodal Retrieval Heads in Long-Context Vision-Language Models

This work formalizes and detects multimodal retrieval heads (MMRetHeads) in long-context vision-language models, demonstrating that a sparse set of 4.4%–10.2% of attention heads causally drives multimodal evidence retrieval and reasoning, while their aggregated attention directly yields state-of-the-art zero-shot document re-ranking on MMDocIR.

PN-Mix: Pair-Norm Guided Mixed-Precision Key Cache Quantization for Efficient Long-Output LLM Inference

Addressing the severe KV cache memory bloat and reasoning accuracy collapse under uniform 2-bit quantization in long-output LLMs, PN-Mix demonstrates that query-key pair-norm products in RoPE-coupled channels dictate attention logit perturbation, proposing a mixed-precision key-cache quantization framework (2.26-bit Key) with one-time prefill salient-pair selection that maintains reasoning accuracy and unlocks up to 2.37x serving throughput.

Meta-Prompt Optimization for LLM-Based Sequential Decision Making

To resolve the failure of traditional prompt optimization caused by non-stationary rewards during LLM agent interactions, this paper proposes EXPO (grounded in adversarial bandits) and its exemplar selection extension EXPO-ES, which dynamically optimize task descriptions, meta-instructions, and exemplar sequences to significantly accelerate convergence and reduce regret.

Holmes-SQL: Benchmarking LLM Agents on Evidence-Aware Investigative Reasoning over Structured Databases

This paper introduces Holmes-SQL, the first benchmark evaluating evidence-aware urban crime investigation over relational databases across 500 cases, a 17-table schema, a 3-tier evidential hierarchy, and 10 logical-fallacy patterns. Empirical evaluations reveal that despite achieving over 97% SQL execution success, top LLM agents achieve at most a 49% solve rate while falsely accusing innocent fallacy-aligned suspects 39.6%–67.9% of the time, exposing evidence weighting under state uncertainty as the primary cognitive bottleneck.

Understanding Reliability in LLM-based Human Behavior Simulation

This paper introduces ReliMap, a structured diagnostic framework that decomposes LLM-based human behavior simulation into three generative layers (Base Distribution, Profile Conditioning, and Population Aggregation) and evaluates reliability at both the individual level (\(R_1\), Accuracy) and population level (\(R_2\), Total Variation Distance) across model capacity, profile completeness, and population coverage, uncovering a fundamental decoupling where individual accuracy gains do not reliably translate to population-level alignment.

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

This work introduces TAPIR, an evaluation protocol that distils a frozen, survey-grounded persuasion rubric into an anti-symmetric dual-head architecture, achieving exact algebraic speaker-order invariance (WPI=1), calibrated probabilities under proper scoring rules, a 39.41× inference acceleration, and a +7.68 pp accuracy gain over the same-backbone MS-PS baseline.

Tracing the Complexity Profiles of Different Linguistic Phenomena Through the Intrinsic Dimension of LLM Representations

By evaluating the intrinsic dimension (TwoNN) and directional information imbalance of residual stream representations across six prominent LLMs on controlled minimal pairs, this paper reveals that complex linguistic structures—particularly recursive subordination—systematically elevate representation manifold dimensionality, with mid-layer ID peak regions playing a causally essential role in building hierarchical syntactic structures.

Adaptive Knowledge Routing for Complex Reasoning in LLMs

To tackle the widespread failure of Large Multimodal Models in complex pathological reasoning—where models stagnate at surface-level descriptions and ungrounded hallucinations—this paper establishes MedInsightBench (332 cases, 3,933 insights) and proposes MedInsightAgent, a multi-agent framework combining visual root finding, grounded insight analysis, and adaptive follow-up questioning, boosting Qwen3-VL-32B to 0.5242 G-Eval F1 and 0.4995 Novelty.

Not All Queries Are Equal: Query-Type-Aware Mixture-of-Experts for Temporal Knowledge Graph Forecasting

QTI-MoE uncovers the severe performance drop of monolithic temporal knowledge graph forecasters in sparse and high-diversity historical regimes, categorizes queries into sparse, recurrent, and ambiguous types with offline LLM-refined soft labels, and routes queries to three specialized experts with complementary memory inductive biases using pure query embeddings at test time.

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

This paper reveals the bag-of-words degradation phenomenon in contrastive table dense retrieval, constructs a structure-adversarial benchmark with strictly lexical-equivalent negatives via Homologous Structural Swapping, and proposes GSR-Table, which applies orthogonal attention pooling and dynamic masking reconstruction to regularize dense representations with structural constraints.

Distributional Stabilization for Zero-Shot Relation Extraction

Under a full-vocabulary diagnostic on joint-encoding zero-shot relation extraction, this paper identifies two independent distributional pathologies—recall collapse and confidence ambiguity—and resolves them with SCAR via gradient-isolated Z-score normalized coarse-to-fine cascade augmentation and training-only interactive contrastive alignment.

EgoVoice: Proactive Spoken Assistance from Egocentric Multimodal Streams

Addressing the critical absence of proactive spoken guidance benchmarks and end-to-end training paradigms from streaming first-person inputs, EgoVoice reconstructs HoloAssist via sound source separation and duration-controlled speech resynthesis, presenting an end-to-end full-duplex multimodal framework optimized via SFT and length-normalized DPO that intervenes with timely speech instructions that conclude before user actions begin.

KItCAT: Knowledge Injection via Input Corruption for Auto-regressive Training

KItCAT presents a lightweight and compute-free data augmentation framework for decoder-only continued pre-training (CPT) by stochastically corrupting input token prefixes while preserving the original autoregressive targets, eliminating spurious lexical shortcuts in low-diversity niche corpora and achieving a 5–9× effective data multiplier over standard next-token prediction.

SSPC: Semantic-Structural Propagation Calibration for Multimodal Fake News Detection

Addressing the vulnerability where graph edges connect related posts with opposing veracity, SSPC shifts graph propagation from unconditional label smoothing to calibrated evidence transfer by repurposing noisy LLM pseudo-labels into learnable edge routing and cross-fitted node reliance gating.

Omnilingual MT: Machine Translation for 1,600 Languages

Meta AI introduces Omnilingual MT, scaling translation support to over 1,600 languages through a 256K extended tokenizer, grammatically diverse seed dataset MeDLEy, and a novel three-stage auto-encoding training scheme, allowing 1B–8B specialized models to consistently outperform 70B general-purpose LLMs in cross-lingual reasoning and long-tail generation.

CoMerge: Conflict-Driven Preference Optimization for Multi-Task Model Merging

CoMerge reformulates multi-task model merging as a self-supervised preference optimization framework, synthesizing hard negative samples from the degradation defects of naive task arithmetic merging to optimize lightweight tensor-wise coefficients (only 1,445 parameters for 8B models), achieving 0.9968 normalized performance on MergeBench while cutting GPU cost to peak by 60.1%.

Personalization over Generalization: History-Aware Sentence-Level AI Detection in Collaborative Writing

Addressing the limitation that conventional sentence-level AI detectors enforce a generic one-size-fits-all boundary that ignores individual writing styles, PHAS-ID grounds predictions in an author's labeled historical writing via dual macro-micro cross-attention and introduces GhostFill to faithfully synthesize collaborative interleaving dynamics.

Imag-Eval: a language-grounded framework for interpretable Text-to-Image instruction following evaluation

To tackle the confounding between surface prompt length and compositional complexity while addressing error propagation and overlooked structural defects, Imag-Eval establishes a language-grounded evaluation benchmark that disentangles instance counts from multi-constraint bindings, demonstrating via 8,842 annotated rules that generation breakdowns are primarily driven by compositional load rather than prompt length.

Adaptive Knowledge Routing for Complex Reasoning in LLMs

This paper proposes an evaluation-free benchmark analysis paradigm based on Sparse Autoencoder (SAE) feature coverage, transforming the computationally prohibitive multi-model performance-matrix problem into a lightweight feature-footprint diagnosis on a single indicator model, achieving a Spearman rank correlation of \(\rho \approx 0.85\) with true performance redundancy across 17 benchmarks and enabling deterministic, low-variance subset compression.

How Far Are AI Scientists from Changing the World? A Survey of Autonomous Scientific Discovery

This comprehensive survey systematically examines autonomous scientific discovery systems, establishes strict boundary criteria distinguishing true AI Scientists from single-task computational tools, structures the field around a four-capability loop (acquisition, hypothesis, validation, evolution), and proposes an E0–E4 cumulative evidence framework that prioritizes inspectable, auditable research trajectories over prompt-to-paper text generation.

MENTOR: Reinforcement Learning via Flexible Teacher-Optimized Rewards for Tool-Use Distillation

MENTOR resolves the exploration dilemma in distilling tool-use capabilities to small language models by replacing rigid trajectory imitation with a flexible process-aware reward (correctness, unordered set-level teacher alignment, and runtime validation) under GRPO, substantially boosting out-of-domain and unseen-tool reasoning performance.

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

Addressing severe acoustic distortion and linguistic fragmentation in speech affected by neuromotor and neurodegenerative disorders, AP-GRPO introduces an anchor-gated phonetic reinforcement learning framework that anchors reliable words and aligns corrupted inter-anchor spans against the patient's own speech signals via Soft-DTW, slashing word error rates on severe dysarthria and suppressing hallucinated insertions without requiring ground-truth reference transcripts.

Towards AI-Assisted Research Writing: Benchmarking LLMs for AI/ML Introduction Generation

This paper proposes SciIG, the first large-scale systematic benchmark for evaluating large language models on generating scientific paper introductions, curating 3,900 post-cutoff papers from NAACL 2025 and ICLR 2025 and employing a 30-metric multi-channel framework combining automated formula metrics, LLM judges, and expert human evaluations.

DVD: A Robust Method for Detecting Variant Contamination in Large Language Model Evaluation

Addressing the failure of traditional contamination detectors against semantically equivalent paraphrased variants, DVD introduces a reference-free framework that captures the bimodal variance of low-probability tail-token synthetic difficulty under temperature sampling, achieving up to 0.239 AUC gain over competitive baselines.

Want Better Synthetic Data? Steer It: Activation Steering for Low-Resource Language Generation

This paper introduces activation steering into low-resource synthetic data generation, deriving novel "Quality steering vectors" by contrasting human-authored and round-trip backtranslated text alongside Language steering vectors; steering early Transformer layers consistently enhances sample diversity and downstream classifier performance across 11 typologically diverse languages.

Canonical Color as a Lens into Concept Decodability in Vision Encoders and VLMs

By adopting "canonical color" as a controlled lens and removing chromatic and luminance cues via grayscale conversion and histogram equalization, this work demonstrates that object-level conceptual knowledge remains linearly decodable in frozen vision encoders and reveals that VLM post-training reformats rather than erases these visual semantic representations downstream.

How to Attack Stealthily? A Letter Frequency-based Backdoor Attack Against LLMs Bounded by Logit Fluctuation

This paper demonstrates that conventional word-level backdoor attacks are intrinsically vulnerable to token-masking defenses due to sharp output logit fluctuations (LF), formalizes LF as a principled stealthiness metric with provable sufficient bounds, and proposes BadFreq—a fine-grained letter-frequency backdoor based on the rare character 'z' that surpasses the runner-up baseline by 49% in LF-averaged stealthiness while maintaining near-perfect attack success.

LoMaR: Local Margin Preservation via Residual Subspace Adaptation for Noisy Prompt Robustness in Small Language Models

LoMaR tackles the noisy prompt vulnerability of on-device small language models by combining Stage I clean local logit margin expansion with Stage II SVD-decomposed residual subspace adaptation and margin alignment, significantly lowering performance drop rates and token rank flip rates without incurring any inference overhead.

HEAL: Anchored Hyperspherical Equilibrium for Multilingual Safety Alignment

Addressing the vulnerability where English-only safety preference data fails to transfer reliably into low-resource languages, this paper proposes HEAL—a lightweight, plug-and-play auxiliary objective that balances a stop-gradient English anchor shrink force with a pairwise non-English dilation force to establish a controlled shell-like equilibrium geometry on the unit hypersphere, substantially improving multilingual refusal and out-of-distribution jailbreak robustness without requiring any translated responses.

SaFeR-Steer: Evolving Multi-Turn MLLMs via Synthetic Bootstrapping and Feedback Dynamics

To combat safety decay and jailbreak vulnerabilities in multi-turn multimodal conversations, SAFER-STEER establishes a staged alignment framework combining synthetic bootstrapping for defensive cold-start with tutor-in-the-loop GRPO guided by trajectory-consistent summative rewards, achieving state-of-the-art bidirectional transfer across static and dynamic safety regimes.

From AI Reviewers to Evidence Assistants: Quantifying the Human-AI Responsibility Boundary in Peer Review

Leveraging 600 ICLR 2026 submissions, 2,231 official human reviews, and 3,600 AI reviews, this paper quantifies the human-AI responsibility boundary across review usability, score calibration, panel diversity, and grounded synthesis, showing that AI should serve as an inspectable evidence assistant while scoring and final evaluative synthesis must remain human-led.

Size Sound Symbolism Survives Genealogy Across 27 Languages

Addressing whether sound symbolism is merely an artifact of shared ancestry or language contact, this paper establishes an audio-validated dataset of 810 size adjectives across 27 languages and 13 families, demonstrating through normalized Levenshtein distance transfer tiers and adversarial feature scrubbing that phonological size cues survive genealogical controls and remain consistently above chance.

REINS: Refusal-Enhanced Inhibitory Steering with Sparse Autoencoder Features

To overcome the failure of single-direction sparse autoencoder (SAE) steering on harmful prompts embedded in complex disguises, this paper establishes the hierarchical GUISE benchmark and develops REINS—a coordinated steering method that suppresses dynamic harmful continuation features while amplifying calibrated refusal features, alongside a prompt-side REINS-Gate that safeguards general capabilities.

Learning to Remember: End-to-End Training of Memory Agents for Long-Context Reasoning

Unified Memory Agent (UMA) formulates query-agnostic memory management as an end-to-end Markov Decision Process using Task-Stratified GRPO, enabling a single policy to actively maintain structured key-value stores across long-horizon text streams and achieve substantial gains across diverse reasoning and state-tracking benchmarks.

Debias-SparseGPT: Bias-Aware Pruning for Large Language Models

Debias-SparseGPT introduces a bias-aware second-order formulation into post-training pruning by augmenting the layer-wise reconstruction objective with paired demographic contrasts, directly counteracting stereotype amplification across unstructured and 2:4 semi-structured sparsity regimes without compromising computational throughput or baseline language modeling accuracy.

3D-MoE: Towards Spatial Intelligence with Mixture-of-Experts for 3D Reasoning and Action Generation

3D-MoE establishes a sparse 3D multimodal foundation model by cultivating specialist and fusion experts via a two-phase routing curriculum with penalty-adaptive load balancing, seamlessly integrating a single-step rectified flow action head (Pose-DiT) to excel across 3D vision-language reasoning benchmarks and real-time embodied robot manipulation.

Adaptive Knowledge Routing for Complex Reasoning in LLMs

Through extensive cross-lingual and cross-architecture benchmarking, this paper shows that general-purpose thinking tokens in Large Reasoning Models (LRMs) bring no significant benefit to machine translation, and distilling human-inspired step-by-step CoT traces into student models fails to outperform standard input-output fine-tuning (IOFT); the performance gains from MT-specific modular prompting strategies stem strictly from intermediate translation drafts embedded inside the traces rather than the surrounding reasoning explanations, demonstrating that using teacher models to polish target translations and expand parallel corpora is far more impactful than building "thinking" translation models.

RATF: Robust Adaptive Tri-Encoder Framework for Fake News Detection against Style-Confusion Attacks from LLMs

To defend against LLM-driven style-confusion attacks where false claims mimic trusted journalistic tones within a mixed stream of authentic articles, RATF introduces a tri-encoder framework driven by conditional probability curvature rewritten scores, adaptively balancing objective content and stylistic signals via conditioned cross-attention and learnable parametric gating.

GUI-PRA: Process Reward Agents for GUI Tasks

GUI-PRA reformulates GUI process-level reward modeling from passive single-pass scoring into an active investigation agent, decoupling "what to verify" via experience-injected criterion synthesis from "how to verify" via criterion-guided autoregressive visual tool invocation, substantially boosting online execution success rates on AndroidWorld and Mobile-MiniWoB++.

Adaptive Knowledge Routing for Complex Reasoning in LLMs

This paper discovers that Matryoshka Representation Learning (MRL) retrieval embeddings inherently encode ontological granularity along their prefix dimensions, and proposes PEEL—a minimal 98K-parameter order-embedding probe that decodes high-fidelity taxonomic DAGs directly from concept vocabularies without needing document corpora or encoder retraining.

Adaptive Knowledge Routing for Complex Reasoning in LLMs

This work establishes a human-grounded benchmark for source-sensitive calibration in Turkish and English by holding propositional content constant while manipulating source trustworthiness and evidential markers, discovering that LLMs possess only weak, additive source awareness and severely fail to translate source credibility into appropriately cautious, human-like action-readiness judgments.

Adaptive Knowledge Routing for Complex Reasoning in LLMs (KNOWSIM)

KNOWSIM introduces a user simulator grounded in cognitive learning theory that tracks explicit knowledge state trajectories over a prerequisite Information Unit (IU) graph under cognitive load constraints, computing mechanistic metrics (Knowledge Gain, Delivery Calibration, Cognitive Overload) and uncovering aptitude–treatment interactions across frontier LLMs.

HoliTok: A Continuous Holistic Tokenization with Robust Dual Capabilities of Speech Generation and Understanding

HoliTok introduces a holistic continuous speech tokenizer that compresses 48 kHz audio into a compact 25 Hz sequence of 128-dimensional continuous latents, achieving high-fidelity waveform reconstruction while resolving the representation conflict between speech generation and understanding within a shared AR+DiT backbone.

Adaptive Knowledge Routing for Complex Reasoning in LLMs

To tackle retrieval noise accumulation and isolated evidence mismatch in knowledge-based visual question answering, this paper presents MaS-VQA, a selection-driven framework that performs joint cross-modal "Mask-and-Select" filtering over visual regions and text passages to construct compact explicit evidence, which subsequently guides a frozen MLLM to synthesize constrained implicit reasoning paragraphs without any reasoning-stage retraining.

LLM Fine-tuning with an Unconstrained Fine-tuned Teacher

Addressing the trade-off between target-task adaptation and general capability retention during LLM fine-tuning, UFT reconceptualizes the teacher from a static knowledge transferer into a value discovery mechanism, identifying high-value training samples via source-relative uncertainty budgets and guiding confidence-adaptive dual-teacher distillation to achieve superior Pareto performance on code and math reasoning benchmarks.

Benchmarking Patent Drafting from Inventor-Style Disclosures

Addressing the disconnect between existing patent benchmarks and real-world drafting practices, this paper establishes Dis2Pat, the first multimodal disclosure-to-patent benchmark from informal inventor-style inputs, and introduces Patent-MAF, a fully locally deployable multi-agent framework that outperforms open-source baselines and rivals state-of-the-art closed-source models in generating legally coherent claims and detailed specifications.

SWE-MEM: Learning Adaptive Memory Management for Long-Horizon Coding Agents

SWE-MEM introduces a proactive and on-demand memory management framework for long-horizon software engineering agents, equipping models with a flexible memory tool to autonomously decide compression timing, arbitrary span selection, and progressive summarization. Combined with curriculum fine-tuning and Memory-aware GRPO, it reaches 43.4% and 60.2% resolve rates on SWE-Bench Verified using 4B and 30B models under a strict 32K context budget, outperforming prior fixed-window and folding baselines while reducing overall token consumption.

Adaptive Knowledge Routing for Complex Reasoning in LLMs

This paper formulates self-corrected speech reasoning as an explicit condition-updating problem, constructs the controlled 13k-sample SCUR-Bench benchmark, identifies a widespread "final-condition adoption failure" across Speech LLMs, and proposes PACER—a GRPO-based reinforcement framework combining structured condition-aware rewards with speech-evidence contrast that slashes condition-related reasoning errors by 67.5%.

I Am Multitudes: A Survey and Roadmap for Rethinking Identity in NLP

This paper conducts a systematic PRISMA survey of 149 ACL Anthology papers to investigate how identity is operationalized in NLP, exposing a pervasive reduction of complex human identity to static sociodemographic proxies primarily used for post-hoc bias auditing, and establishes a comprehensive roadmap across definition, data collection, modeling, evaluation, and pluralistic alignment for identity-aware NLP.

Bridging Text and Structure via Latent Query Refinement for Knowledge Graph Completion

LQ-KGC establishes an EM-inspired alternating refinement framework between a dual-encoder language model and a relation-aware GNN over a semantically compatible, degree-balanced latent query set, enabling mutual reinforcement via pseudo-edge message passing and structural soft-label distillation.

OptiSkill: A Hierarchical and Evolving SkillBank for LLM-Based Optimization Modeling

OptiSkill establishes a hierarchical and evolving SkillBank for frozen LLMs—combining global formulation strategies with local step-level error-prevention experiences distilled from solver-verified trajectories and safely updated via batch-level admission checks—achieving superior formulation accuracy across eight diverse OR benchmarks.

Reviewing the Reviewer: LLM-Assisted Reviewer Feedback Generation for Guideline Compliance

Addressing pervasive lazy thinking and non-specific critiques in scientific peer review, this paper introduces an end-to-end framework integrating sentence-level BIO segmentation, neuro-symbolic multi-label violation detection, and template-grounded evolutionary feedback generation, reducing guideline violations in human rewrites by up to 92.4%.

Can Humans Dream of Electric Sheep? Human-Written Samples for Fine-Grained Vision-and-Language Hallucination Benchmarking

To overcome the severe model-dependence, sparse representation of rare error types, and rapid obsolescence inherent in modern LVLM hallucination evaluation, this paper presents SHEEP—a multilingual benchmark spanning 20,000 samples across four languages (Chinese, English, French, Italian) with character-level annotations over 5 hallucination categories, empirically demonstrating that human-written hallucinations provide higher inter-annotator agreement, finer category controllability, and faithful ranking proxies for model-generated outputs.

TriggerBench: Investigating Prospective Memory for Large Language Models

This work introduces TriggerBench, a five-dimensional benchmark comprising 1,265 tasks that systematically evaluates Prospective Memory (PM)—the capability to spontaneously recall and act on latent constraints without direct prompting—revealing severe attentional fragility and a Cognitive Cliff over long contexts, while establishing PM as a fine-grained behavioral probe of spare reasoning capacity.

TORe: Rehearsal-Free Continual Learning for Table QA via Operation Recipe Reconstruction

Addressing catastrophic forgetting caused by schema shifts and severe privacy regulations in evolving tabular environments, TORe introduces a rehearsal-free continual learning framework for Table QA that abstracts reasoning procedures into schema-agnostic Operation Recipes, identifies fine-grained structural deviations, and reconstructs both historical and consolidation knowledge via execution-verified LLM synthesis without storing raw historical data.

Revisiting Reinforcement Learning with Verifiable Rewards from a Contrastive Perspective

By revealing that GRPO is mathematically equivalent to a weighted positive-negative discrimination objective with likelihood-misaligned scores and score-insensitive credit assignment, ConSPO introduces length-normalized sequence log-probabilities into a group-wise InfoNCE formulation with a curriculum-scheduled margin, significantly outperforming GRPO across challenging reasoning benchmarks.

RePro: Proof-Verified Benchmark Rewriting for Reliable Evaluation of LLM Mathematical Problem Solving

This paper proposes RePro, the first verification-driven benchmark rewriting framework integrating Lean 4 neural automated theorem provers (ATPs) to filter candidate rewrites through staged feasibility screening, executable formalization, and kernel-checked proof verification, achieving 100% well-definedness, feasibility, and answer correctness on retained instances while exposing significant model sensitivity and memorization effects.

How Much is a Human Right Worth? ECtHR-NPD: A Benchmark for Predicting Non-Pecuniary Damage Awards

This paper introduces ECtHR-NPD, the first continuous monetary regression benchmark for predicting Article 41 non-pecuniary damage awards at the European Court of Human Rights across 14,575 cases, revealing that feature-based gradient-boosted trees and fine-tuned encoders consistently outperform frontier prompted LLMs and ReAct agents due to severe failures in zero-award identification and heavy-tail magnitude calibration.

Diagnosing LLM Fragility: An Empirical Study on Reasoning Entropy and Malicious Outputs

This paper introduces Bifurcation Entropy (HBE) to extend Kernel Language Entropy to intermediate reasoning steps, empirically demonstrating via In-Distribution Attacks (IDA) that top-quartile high-entropy decision nodes are exploited up to 6.2× more frequently than low-entropy steps, establishing an efficient triage signal for CoT safety monitoring.

Phrase-Localized Language-Contrastive Guidance: Training-Free Localized Accent Control for Code-Switching Text-to-Speech

Addressing the severe cross-lingual accent leakage in zero-shot multilingual TTS where foreign embedded phrases are flattened into the matrix carrier accent during code-switching, this work introduces Phrase-Localized Language-Contrastive Guidance (LCG). LCG extracts acoustic phrase masks via internal self-attention probing and decouples contrastive language steering from global text guidance, delivering authentic native-accented pronunciations without fine-tuning.

Refuse without Refusal: A Structural Analysis of Safety-Tuning Responses for Reducing False Refusals in Language Models

By structurally decomposing safety-tuning responses into boilerplate refusal statements and explanatory rationales, this work uncovers that initial refusal statements act as drivers of superficial cue reliance and false refusals; training models exclusively on request-specific rationales (Rationale-Only) substantially decreases over-refusal on benign queries while maintaining robust safety defense.

Forgetting Across Modalities: Unlearning in Multimodal Large Language Models

Addressing the severe gradient conflicts between forgetting and retaining objectives caused by cross-modal representation coupling in MLLMs, this paper proposes EASE (Efficient Alignment-aware Selective Erasing), which dynamically identifies conflicting parameters via opposite-sign gradient detection and resolves interference via manifold-constrained optimization in a 2D task-gradient subspace.

Knowing What Not to Answer: Selective Non-Compliance in Vision-Language Models

This paper introduces KoNA, a benchmark evaluating selective non-compliance in Vision-Language Models (VLMs) across paired single and compound queries, and proposes a two-stage fine-tuning framework integrating SFT and GRPO to enable models to refuse or correct invalid query components while reliably answering valid ones.

CAME: Company-Aware Evidence-Memory Experts for Interpretable Quarter-Ahead Revenue Forecasting

CAME reformulates quarter-ahead corporate revenue prediction as an anchor-relative log-residual refinement task over a no-leakage statistical baseline, combining typed semantic evidence cards, guidance-gated memory correction, and multi-channel experts to achieve superior numerical precision and source-linked auditability across 336 company-quarters.

Adaptive Knowledge Routing for Complex Reasoning in LLMs

Addressing the confound between substantive normative evaluation and generic multiple-choice answer-formatting machinery, this paper proposes a controlled locate–disentangle–intervene pipeline with non-normative MMLU-STEM controls that isolates 5–10 late-layer normative-specific (NS) attention heads (<1% of total) across Qwen3-8B, Llama3.1-8B, and Mistral-7B-v0.1, establishing bidirectional causal effects via knockout and amplification.

Auditing Chinese Web-scale Corpora via Sampled BPE Token Statistics

Inspired by marine pollution monitoring, SAMPLED-BPE introduces a lightweight token-level auditing pipeline that trains BPE on small streaming samples to profile web-scale Chinese corpora with a 148.4x speedup and 35.8x memory reduction, releasing a 630k-token hierarchical pollution dataset across 92k trees.

Cross-Version ∆-Transfer of QLoRA-Trained Adapters

This paper demonstrates that BF16 LoRA adapters trained on a 4-bit quantized base (NF4) transfer across model versions (e.g., Llama-3-8B-Instruct to Llama-3.1-8B-Instruct) as effectively as full-precision LoRA adapters across 20 downstream tasks (absolute score gap \(\le 0.054\)), while cutting post-load deployment VRAM by 49%–63% under NF4 runtime serving and showing that the empirical task-optimal merge scale \(\lambda^*\) systematically exceeds the weight-space Frobenius distance optimum.

DEPPO: Dual Experience Pool Policy Optimization for Long-Horizon Reinforcement Learning

Addressing the severe credit assignment mismatch caused by outcome-level reward broadcasting in long-horizon agents and complex multi-step reasoning, DEPPO maintains structured dual experience pools for historical successes and failures to compute state-conditioned success–failure preference log-ratios, adaptively reweighting step-level advantages under support-aware gating and asymmetric memory updates without any external critic or process supervision.

When Fairness Evolves: Value Orientation and Evolution in Multi-Agent Resource Allocation

This paper constructs a closed-loop survival-production multi-agent simulation using LLMs to demonstrate that deliberative negotiation endogenously resolves the fairness-efficiency tradeoff, uncovering systematic misalignments between subjective satisfaction and objective resource distribution while tracing the semantic evolution of fairness positions toward hybrid normative principles.

Adaptive Knowledge Routing for Complex Reasoning in LLMs

This paper introduces EFPL, a decoupled planner-executor framework that structures the agent decision space into explicit ASK, ANSWER, and ACT interaction modes and dynamically routes all-failed rollout groups to privileged self-distillation, substantially curbing tool over-calling, argument fabrication, and unfeasible task hallucination on BFCL and When2Call benchmarks.

Adaptive Knowledge Routing for Complex Reasoning in LLMs

Addressing the severe latency bottleneck of listwise LLM rerankers that ingest lengthy concatenations of full-text documents, this paper presents RRK, an efficient listwise reranker operating directly over multi-token soft-compressed representations that achieves 5×–20× latency reductions on a 7B backbone while matching or surpassing 0.6B–4B full-text rerankers across standard and long-document benchmarks.

Adaptive Knowledge Routing for Complex Reasoning in LLMs (Colla-Q: Toward Collaborative Experts in MoE Quantization via Minimax Precision Balancing)

Addressing the bottleneck where heavily degraded weak experts harm the aggregated output in low-bit MoE models, this paper proposes Colla-Q, an ensemble-inspired framework that estimates continuous expert capacity and uncertainty via label-free activation entropy and recursively compensates the weakest expert through minimax precision balancing, yielding superior multi-task reasoning accuracy and domain-invariant calibration at 1.57–2.54 bits.

In-Place Feedback: Reliable Refinement for Multi-Turn Expert-LLM Collaboration

Addressing the severe failure of LLMs to reliably absorb multi-turn conversational corrections due to error accumulation in dialogue history, this work proposes In-Place Feedback—an interaction paradigm where experts directly edit erroneous spans in the model's prior response and prune dependent reasoning, driving continuation from a clean corrected state while dramatically reducing token consumption and expert fatigue.

AdaPLD: Adaptive Retrieval and Reuse for Efficient Model-Free Speculative Decoding

AdaPLD is a training-free, model-free speculative decoding framework that breaks the candidate reachability bottleneck via lexical-first matching with semantic embedding fallback, and mitigates the brittleness of deterministic span copying through branched draft tree construction with successor extensions, delivering up to 3.10× end-to-end decoding speedup across diverse generation, editing, and reasoning benchmarks while strictly preserving the target model's output distribution.

SpecTr-GBV: Multi-Draft Block Verification Accelerating Speculative Decoding

SpecTr-GBV unifies multi-draft speculative generation and greedy block verification under optimal transport theory, rigorously achieving the theoretical upper bound of expected acceptance length under conditionally independent couplings while preserving exact target distribution fidelity and substantially improving inference throughput.

I³CL: Restoring Interactive Capabilities to Implicit In-Context Learning

This paper proposes Interactive Implicit In-Context Learning (I³CL), an elegant framework that restores inter-demonstration self-attention interactions and query-aware cross-attention routing to implicit task-vector representations, eliminating the representation bottleneck of isolated encoding and uniform average pooling while achieving a +3.54% average accuracy gain over existing task-vector baselines across eight classification benchmarks, outperforming LoRA fine-tuning and delivering near zero-shot inference efficiency.

GutenBench: A Taxonomy-Based Dataset for Evaluation of Information Retrieval Models

This work introduces GUTENBENCH, a fine-grained zero-shot IR benchmark built upon a four-level capability taxonomy (lexical, semantic, explicit aggregation, and latent conceptual bridging) spanning 2,170 expert-adjudicated queries and 272 parallel Bible translations across 141 languages, revealing that state-of-the-art retrievers systematically degrade on high-level conceptual queries and suffer acute performance drops in low-resource cross-lingual scenarios.

KairosAgent: Agentic Time Series Forecasting with Fused Semantic Reasoning

KAIROSAGENT introduces a reason-then-forecast agentic framework that dynamically invokes 23 statistical tools to deduce qualitative future morphology descriptions, fuses these semantic priors into a lightweight time series foundation model via gated cross-attention, and trains the reasoner using turn-level reinforcement learning on the 40k-trajectory T-STAR corpus for superior zero-shot forecasting.

INCLUDE-2.0: Disentangling Knowledge Gaps in Multilingual LLM Evaluation

INCLUDE-2.0 introduces a massive native multilingual benchmark covering 91 languages, 86 countries, and 127 language–country pairs with over 220K curated questions, establishing a four-way regional knowledge taxonomy and a paired-English evaluation framework that disentangles superficial linguistic failure from missing regional knowledge.

Targeted Token-Level Unlearning for Repairing Text-to-SQL Models

This paper proposes TTU-SQL, a targeted token-level machine unlearning framework that localizes the first structural divergence point between predicted and gold SQL using normalized ASTs, filters three forgettable error types, and applies localized Part-NPO with recovery SFT to eliminate recurring error generation tendencies without accessing the original training data.

JARVISGUI: Towards Cross-Device GUI Agents with Dynamic Task Composition

To overcome the single-device limitation of existing GUI agent benchmarks, this paper introduces JarvisGUI, the first large-scale dynamic benchmark spanning heterogeneous Android, Windows, and Ubuntu ecosystems; through slot-typing atomic task formalization and automatic task graph composition, JarvisGUI constructs 150 composite workflows and reveals that state-of-the-art open-source agents suffer severe performance collapses (falling as low as 0% success rate) on cross-device dependent tasks.

MARS: What Retrieval Signals Are Hidden in Multimodal Large Language Models for Text-Video Retrieval?

Addressing the severe information bottleneck caused by single-token single-layer embedding extraction in MLLM-based text-video retrieval, MARS introduces multi-layer adaptive representation slots, slot-aligned similarity aggregation, and hard-negative-aware slot specialization, achieving new state-of-the-art performance across four benchmarks within an efficient dual-encoder framework.

BEAR: Budgeted Evidence Allocation for Multi-Document Reasoning

Addressing the core tension between prohibitive compute costs in long-context LLMs and fragmented contextual isolation in flat retrieval, BEAR organizes multi-document collections into hierarchical semantic forests offline and dynamically balances high-level LLM exploration with structural TreeExpansion recovery under a strict token budget, outperforming prior retrieval baselines across nine benchmarks with an approximately 4K-token evidence context.

ToToFilter: Topic- and Topology-Aware Noise Filtering Framework over Medical KGs for Evidence-Based Medical Q&A

This paper proposes ToToFilter, a train-free topic- and topology-aware noise filtering framework that addresses entity-level noise from generic high-frequency hubs via Topic-Aware Entity Localization (TEL) and path-level noise from redundant or spurious multi-hop chains via Topology-Aware Path Filtering (TPF), significantly outperforming competitive graph-augmented baselines across four medical QA benchmarks and PrimeKGQA.

Breaking Temporal Ambiguity: Linguistic Temporal Knowledge Transfer for Weakly-Supervised Action Localization

Proposes the LTKT framework, which moves beyond treating actions as isolated visual snippets by using LLM-deconstructed three-stage physical priors ("preparation-execution-completion") coupled with asymmetric temporal causal gating to suppress static background noise, transferring structural temporal knowledge into a lightweight student backbone via asymmetric distillation and directional contrastive learning to establish new SOTA on THUMOS14 and ActivityNet.

ELITE: Evolving Localized and Informative Temporal Entity Alignment

Targeting the realistic scenario of continuously growing temporal knowledge graphs (TKGs), ELITE proposes an evolving continual alignment framework that leverages complementary relational-temporal seed selection, localized candidate inference, and margin-guided cross-snapshot score aggregation, matching full-retraining accuracy while cutting inference time by 3.7× and memory by 10×.

Linear Probing Provides Robust and Efficient Detection of Machine-Generated Text

Grounding machine-generated text detection in the Linear Representation Hypothesis, this paper reveals that machine-generated and human-written texts are inherently linearly separable in low-dimensional latent space, and shows that simple linear probes trained on PCA-reduced frozen hidden states achieve state-of-the-art out-of-domain robustness (+11 AUC) with only 10–100 samples while encoding a continuous spectrum of AI-editing degree.

EHR-Complex: Benchmarking Medical Agents for Complex Clinical Reasoning

This paper introduces EHR-Complex, the first large-scale interactive clinical database reasoning benchmark constructed on the complete MIMIC-IV substrate (365K patients, 31 tables, 500M+ records), featuring 52K executable tasks across 6 clinical intents and dual scopes, with queries averaging 31.93 SQL structural components, systematically exposing that frontier LLMs struggle with cohort-level temporal alignment and medical-code grounding (achieving at most 62.3% exact-match accuracy).

I Know What You Meme, Even If it Emerged Today: Understanding Evolving Memes through Open-World Knowledge Acquisition

This paper proposes Query-Retrieve-Conclude, a zero-shot framework that identifies missing knowledge gaps, retrieves external open-web evidence, and synthesizes evidence-grounded background knowledge statements, overcoming the failure of closed parametric VLMs on fast-evolving memes and lifting downstream detection Macro F1 to 0.71.

AnyCap: Omni-Modal Captioning with Instruction Alignment

Addressing the lack of fine-grained instruction controllability in existing multimodal captioners and the risk of catastrophic forgetting caused by direct fine-tuning, this paper proposes AnyCap—a plug-and-play revision-based alignment framework that jointly serves images, videos, and audio using a single model, supported by a 300k instruction triplet corpus AnyCapData and the decoupled AnyCapEval benchmark.

QReason: Query-Focused Decoupled Chain-of-Thought for Efficient Passage Reranking

QReason addresses the severe Chain-of-Thought redundancy and high latency bottleneck caused by sliding-window strategies in listwise LLM rerankers by decoupling global query-focused reasoning from window-specific passage relevance assessment, leveraging a two-stage training scheme (Relevance-Grounded Supervised Tuning and Rank-Aligned Dual-Reward Reinforcement Learning) to generate a one-time reusable reasoning query that matches or outperforms full reasoning rerankers at up to 10.75x speedup.

C³T: Counterfactual Causal Reasoning for Sentiment Shifts in Social-Media Conversation Trees

This paper addresses the lack of causal explainability behind emotional dynamics in social-media conversation trees by introducing the CASIRE benchmark and C³T (Counterfactual Causal Conversation Transformer)—a model integrating tree-temporal message passing, sparse ancestor attribution, and dual-stream forced counterfactual intervention—achieving superior event-OOD robustness and principled estimation of discourse-move treatment effects on downstream negativity.

FullFront: Benchmarking MLLMs Across the Full Front-End Engineering Workflow

FullFront introduces the first comprehensive benchmark assessing MLLMs across the full front-end development workflow—spanning conceptual webpage design, fine-grained visual perception QA, and functional code generation with interaction authoring—revealing critical grounding failures and a striking decoupling between visual perception and code implementation.

Adaptive Knowledge Routing for Complex Reasoning in LLMs

This paper proposes MARS (Multi-Agent Relay of Specialized LLMs), a prompt-only framework replacing stage-aligned planner-coder-debugger roles with topic specialists grounded by algorithmic-theory RAG, achieving a 0.624 pass rate on CodeContests with 3.3× lower wall-clock latency than heavy-search baselines.

EmoStance: Response-Side Affective-Orientation Control for Empathetic Response Generation via Emoji Weak Supervision

EmoStance formulates empathetic dialogue generation as response-side affective-orientation control, leveraging multi-annotator emoji distributions to induce a name-free continuous stance space that steers frozen instruction-tuned LLMs via continuous prefix embeddings, achieving a 62.2% decisive win rate in blind human evaluation.

ESGPT: A Data-Centric ESG Language Model with Unified Instruction and Benchmark Resources

To tackle fragmentation in ESG datasets and non-standardized evaluation criteria, this paper develops ESG-Instruct (a 30K balanced instruction tuning dataset) and ESGU-Bench (a benchmark spanning 6 core understanding tasks), training an 8B model ESGPT that outperforms large models like GPT-4.1 and DeepSeek-V4 in ESG reasoning.

MedVision: Benchmarking Quantitative Medical Image Analysis

The paper uncovers severe metric and spatial measurement failures across general and specialized medical vision-language models, curates MedVision as a large-scale quantitative benchmark across 22 multi-modal datasets with 24.3M single-instance annotations, and introduces MedVision-V0, an end-to-end VLM trained with chain-of-thought SFT and multiplicative process-reward GRPO reinforcement fine-tuning that dramatically reduces numeric measurement errors and aligns with real-world clinical staging criteria.

A^2Agent: Action-Aware Reinforcement Learning for Repository-Level Code Localization Agents

Addressing the issues where trajectory-level sparse rewards cannot attribute per-turn action credit and agents frequently explore gold code regions without committing them, A2Agent couples a dense per-turn discovery sequence with a terminal Discovery-Commit reward and estimates action-level advantages via exploration history grouping, allowing a 4B model to outperform baselines up to 8× larger across SWE-Bench Verified (+1.58% F1) and SWE-Bench Pro (+8.55% F1).

MemPro: Agentic Memory Systems as Evolvable Programs

MemPro treats the entire Memory Construction–Retrieval (MCR) pipeline as an evolvable program combining prompts and executable code, maintaining a runnable version tree where an Evolving Agent performs failure-mode-guided edit-debug refinement to surpass static architectures and prompt-only evolution methods within a few iterations.

How Reliable are AMR Parser Outputs? Uncertainty Estimation for AMR Parsing

Addressing the lack of confidence calibration in neural semantic parsing, this work presents the first black-box uncertainty estimation framework for AMR parsers by constructing graph-spectral and lexical estimators over pairwise beam candidate similarities and evaluating reliability via the Prediction Rejection Ratio.

It's Not RoPE that Creates Sinks: The Role of Self-Concentration and Value-Non-Mixing in Attention

By systematically decoupling BOS tokens, RoPE positional encodings, and causal attention constraints, this paper demonstrates that forced self-concentration from causal masking and the resulting "Value-non-mixing" state in early layers drive Attention Sinks and Massive Activations, decisively refuting the prevailing belief that RoPE is the culprit.

LoopCD: Loop-wise Contrastive Decoding for Improving Reasoning in Looped Language Models

LoopCD addresses the non-monotonic loop instability and early overconfidence of Looped Language Models (LoopLMs) through a training-free loop-wise contrastive decoding objective that penalizes premature intermediate distributions against the final refined distribution, delivering substantial reasoning gains across math, code, and QA benchmarks with virtually zero extra overhead.

MedSPA: Self-planning Medical Agents

MedSPA addresses the tension between uncontrolled chain-of-thought hallucination and rigid clinical template traversal in medical vision-language reasoning by decomposing diagnosis into an iterative planning-and-reasoning agent (PR Agent) and an evidence-grounded summary agent (SUM Agent), achieving state-of-the-art clinical efficacy (+0.086 CE-F1 on MIMIC-CXR) through soft template distillation and role-specific reinforcement alignment.

When the Strongest Teacher Is Not the Best Teacher: Student-Centric Answer Selection

Challenging the widespread assumption that the strongest teacher model consistently provides the optimal supervision, this paper introduces Student-Centric Answer Sampling (SCAS), a forward-only, gradient-motivated learning-cost proxy that adaptively samples verified correct teacher solutions best matched to the current student model.

Don’t Guess, Ask Wisely: Benchmarking and Training Proactive Elicitation in Open-Domain LLM Conversations

Addressing the pervasive habit of LLMs silently guessing unstated user needs, this paper introduces VagueBench with known hidden information sets and proposes DialogGRPO, an offline RL framework that formulates multi-turn elicitation as a dialogue-state DAG with turn-level dense rewards, achieving >88% proactive asking rate without requiring runtime user simulators.

Beyond FLOPs: Benchmarking Real Inference Acceleration of LLM Pruning under a GEMM-Centric Taxonomy

Addressing the disconnect where nominal FLOPs reduction fails to reflect physical GPU speedups in large language model pruning, this paper introduces a unified GEMM-centric taxonomy across M, N, and K logical dimensions and delivers an implementation-consistent benchmarking suite, charting the first empirical throughput-quality Pareto frontier and exposing the critical role of non-GEMM execution overheads.

DecSelfMask: Leveraging Unlabeled Text via Self-Relevance-Guided Masking for Decoder-Only Classification

Addressing the severe scarcity of expert annotations in specialized domains, DecSelfMask leverages attention-aware layer-wise relevance propagation (AttnLRP) on task descriptors to locate and mask task-critical text spans under a "two peaks" causality criterion, transforming raw hospital notes into self-supervised training signals that surpass standard continual pretraining by +6.3 F1 and baseline SFT by up to +27.8 F1 across 136 clinical classification tasks.

VideoSearcher: Empowering Video Deep Research with Multi-Tool Agentic Reasoning via Reinforcement Learning

Addressing the disconnect between open-world evidence seeking and video perception, VideoSearcher presents a closed-loop agentic framework that unifies temporal localization, spatial zooming, and multimodal web search, optimized by a novel Bi-branch Sequence Policy Optimization (BiSPO) algorithm that decouples answer accuracy from tool invocation behavior.

From Unverifiable Domain Text to Cross-Lingual RLVR via Code-Switched Task Synthesis

Addressing the severe scarcity of non-English verifiable reward tasks in specialized knowledge-intensive domains (medicine and cybersecurity), this paper proposes code-switched task synthesis to transform unverifiable raw English corpora into multi-view verifiable FIM-MCQA variants with identical answer bindings, significantly boosting target-language reasoning by bridging non-English inputs into dominant English-side reasoning pathways without sacrificing English proficiency.

MARBO: Relational Belief Grounding for LLM Agents in Social Deduction Games

Addressing the strategic inconsistencies, self-revealing speech, and blind voting of LLM agents in partially observable social deduction games, MARBO introduces explicit relational belief modeling into preference optimization, reinforcing speech only when it favorably shapes others' beliefs and rewarding actions only when grounded in correct target beliefs.

Speak for Me: Giving LLMs the Situational Awareness to Participate in a Meeting

Addressing the pervasive "silent abstention" problem in online meeting delegation where LLMs fail to recognize when to speak, this paper proposes CAPA (Collaborative Agent Predictive Architecture) with a perceive-act-recalibrate loop over explicit multi-party state representations, slashing the silence rate from 51.4% to 2.5% and doubling credited idea recovery to 52.2% across 137 AMI meetings.

How Well Can LLMs Simulate Real Learner Evaluations of Educational Feedback?

Using a real-world evaluation dataset from 321 Japanese high school students across multiple biology questions, this study systematically benchmarks how well six representative LLMs simulate subjective educational feedback ratings at both group and individual levels, uncovering a critical bias where models overestimate verbose, symbol-heavy feedback while failing to capture nuanced learner preferences.

PROBE: Prospective Retrieval with Oracle-Based End-Goal Distillation for Multi-Hop Question Answering

PROBE decouples the multi-hop reasoning state into factual summaries and prospective queries with context denoising, and optimizes dense retrievers via final-hop cross-encoder Margin-MSE distillation, enabling rerank-free, label-free multi-hop reasoning.

SMADE-IE: Sparse Multi-Agent Framework with Evidence-Driven Debate for Zero-Shot Information Extraction

SMADE-IE addresses the high token overhead, noisy context, and unstructured disagreement in zero-shot information extraction by introducing an adaptive mode selector for complexity-guided sparse routing and an evidence-driven debate module combining Toulmin argument decomposition, external evidence scoring, and Bayesian Beta confidence updates with dual-track early stopping, reducing token consumption by over 80% on large ontologies while outperforming existing baselines across 9 benchmarks.

Answer Probing-Guided Search for Diverse Solution Exploration of LLMs

Addressing the issue where test-time tree search using response-level semantic embeddings fails due to stylistic and linguistic confounding, this paper proposes Answer Probing-Guided Tree Search (APTS), which probes potential intermediate answers and uses their hidden state similarity and perplexity to steer search toward genuinely diverse and valid reasoning solutions.

Trip+: Benchmarking Agents in Personalized Interactive Travel Planning

This paper introduces Trip+, a comprehensive benchmark for evaluating LLM agents in personalized, multi-turn travel planning across a 40-city sandbox with 11 traveler profiles and stateful simulation, revealing that state-of-the-art models systematically generate technically feasible yet physically exhausting itineraries with weak profile alignment.

RIGOR: Atomic Semantic Normalization and Declarative Constraint Reasoning for Geometry Problem Solving

Addressing the two foundational bottlenecks of deductive database systems—coupled textual descriptions and eager intermediate variable elimination—RIGOR normalizes complex problem texts into atomic statements via an instruction-tuned GSN and solves non-linear geometric constraints declaratively via SMT query-and-refutation, setting new state-of-the-art results across GeoQA, Geometry3K, and PGPS9K.

Cognition on Graph: Navigating Massive Knowledge Space via Cognitive Cycles and Bidirectional Graph-Text Synergy

Addressing the structural sparsity of knowledge graphs and the severe noise of global corpus retrieval, this paper introduces CoG (Cognition on Graph), a training-free cognitive framework that performs continuous plan-explore-synthesis-reflect cycles and establishes bidirectional graph-text synergy, achieving state-of-the-art performance across seven multi-hop QA benchmarks while reducing accessed documents by 12.4×.

Rethinking Gradient Flow Through Frozen Blocks for Memory-Efficient Block-Coordinate Training

Addressing the activation memory bottleneck caused by frozen blocks in block-coordinate descent (BCD) training, FROST discards inactive MLP activations during the forward pass and selectively reconstructs token gradients via upstream-gradient norms and effective-rank budget allocation, cutting peak memory by 20%–34% without downstream performance degradation.

What Drives Paper Acceptance? A Process-Centric Analysis of Modern Peer Review

Drawing upon 28,358 submissions from ICLR 2017–2025 on OpenReview, this paper establishes a four-dimensional, process-centric framework spanning reviewer disagreement, interaction logs, external metadata, and linguistic patterns, revealing how non-content dynamics and optimal interaction windows shape final acceptance decisions.

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

AGORA establishes an archive-grounded document reasoning benchmark spanning 8 professional domains, 9,664 authentic workplace documents, and 372M tokens; its 362 deterministic numeric tasks require active file exploration, unit and terminology reconciliation, and multi-hop arithmetic, on which leading frontier models reach only 59.4% accuracy.

SecOPD: Mitigating Adaptive Prompt Injections by On-Policy Distillation

SecOPD proposes a secure on-policy distillation paradigm that leverages the unperturbed base model under clean context to provide fine-grained token-level advantage feedback on student rollouts under injected inputs, effectively resolving the credit assignment ambiguity of sequence-level rewards and slashing the PISmith adaptive prompt injection attack success rate on Qwen3.6-27B from 94.0% to 9.0% with virtually zero loss in general reasoning capability.

CacheSpec: Finding the Sweet Spot for Small Models in Large Language Models

CacheSpec converts one-time Program-of-Thoughts (PoT) reasoning artifacts into reusable parameterized program caches, employing the same lightweight small model for semantic variable extraction on cache hits and speculative drafting on cache misses, achieving up to 3.1x latency reduction and 2.8x throughput improvement while preserving complex reasoning accuracy.

Lung-R1: A Knowledge Graph-Guided LLM for Pulmonary Diagnostic Reasoning

Addressing the "Pulmonary Knowledge-to-Diagnosis Gap" where LLMs memorize textbook facts but struggle with patient-specific EMR deduction, this paper introduces LungKG (a 59k-node structured pulmonary graph) and Lung-R1 (trained with graph-constrained CoT and graph-guided RL), achieving state-of-the-art clinical EMR diagnosis performance.

Identify, Classify, Act: Per-Instruction Controlled Compliance for LLMs over External Data

To break the fundamental safety-utility tradeoff where models either over-refuse or over-comply when parsing untrusted external inputs, this paper formalizes the External-Data Instruction Classification (EDIC) task, introduces a scene-grounded counterfactual data synthesis pipeline, and develops a generative judge that delivers decomposed rewards to guide CoT policy models via GRPO to selectively follow beneficial instructions while rejecting adversarial prompt injections.

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

Using causal mediation analysis and path patching, this paper demonstrates that large language models do not maintain a global world state across updates, but instead rely on a compact attention circuit (3–10% of heads) that locally remaps binding identifiers during retrieval.

Beyond Outliers: A Data-Free Layer-wise Mixed-Precision Quantization Approach Driven by Numerical and Structural Dual-Sensitivity

Addressing the flaw of existing calibration-free mixed-precision quantization methods that treat modules uniformly and rely solely on scalar outlier statistics, NSDS mechanistically decomposes layers into Detectors and Writers, estimates sensitivity through both Numerical Vulnerability and role-aware Structural Expressiveness, and aggregates scores via MAD-Sigmoid and Soft-OR for robust, data-free bit allocation.

APEx: Distillation of Agent Procedural Experience for Adaptive Deep Research Question Answering

APEx introduces a hierarchical experience utilization framework coupled with a three-stage alternating GRPO training paradigm, organizing interaction traces into instance-level memories and category-level procedural skills that serve as regularization priors for online test-time reinforcement learning without ground-truth labels.

FlowMM: Cross-Modal Information Flow Guided KV Cache Merging for Efficient Multimodal Context Inference

FlowMM reveals that multimodal large language models exhibit distinct layer-wise divergence in cross-modal attention flows, and proposes a tuning-free KV cache merging framework combining layer-adaptive merging domains with sensitivity-protected token matching, cutting cache memory by 80% to 95% and boosting decoding throughput by 1.3-1.8x while suffering only a 0.12% accuracy drop on InternVL2.5-8B.

Towards a Comprehensive Scaling Law of Mixture-of-Experts

Addressing the failure of classical dense scaling laws in Mixture-of-Experts (MoE) architectures, this work systematically decouples five core factors—data size, total model size, activated model size, active experts, and shared expert ratio—to establish a comprehensive joint MoE scaling law across 450 controlled experiments, deriving theoretically invariant structural optima and efficiency-aware sparsity trajectories.

VIRA-Bench: Diagnosing Pedagogical Contingency Understanding in Online Tutoring Videos

This paper introduces VIRA-Bench, a diagnostic benchmark of 1,005 clips and 1,322 multiple-choice questions grounded in the IRF exchange structure and contingent scaffolding theory, exposing a 29.7-point gap between leading MLLMs (Gemini 3 Pro at 61.3%) and human experts (91.0%), with paired video-removal and temporal-shuffle controls systematically isolating visual perception from order-sensitive pedagogical contingency reasoning.

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

Addressing the pervasive bias where large language models default to figurative readings of idiomatic expressions, this paper introduces the IdioSteer benchmark and establishes a white-box residual-stream activation steering framework that leverages an asymmetric cross-layer regime to systematically boost literal interpretation rates by 2.4x to 4.5x without degrading generation fluency.

Stop Before You Fail: Operational Capability Boundaries for Mitigating Unproductive Reasoning in Large Reasoning Models

Revealing that Large Reasoning Models expose early failure signals when facing unsolvable questions, this paper introduces black-box reasoning expression monitoring and white-box prefill hidden-state probing to curtail unproductive reasoning, cutting token expenditure by 62.7%–93.6% with negligible accuracy loss.

Speaking While Listening: A Survey and Empirical Audit of Full-Duplex Spoken Dialogue Systems

This paper introduces three unified formal frameworks—the L0–L3 architectural hierarchy, the \(T \times I \times R\) interaction ontology, and a 5-state decision FSM—to resolve the taxonomy crisis in Full-Duplex Spoken Dialogue Systems (FD-SDS), demonstrating through independent empirical audits on FDB v1.5 that a substantial "Realization Gap" exists between architectural reachability and actual learned behaviors due to training data bottlenecks.

Diagnose, Then Refine: A Closed-Loop TTS System with AudioLLM-Guided Correction

LOOPTTS establishes a closed-loop TTS framework featuring a three-stage Filter–Judge–Refiner pipeline, where an AudioLLM diagnoses subtle word-level prosodic defects to produce structured instructions, and a dedicated Refiner trained with position-aware objectives performs guided expressive re-synthesis to resolve sporadic failures.

AdiDhwani: Preservation-Oriented ASR for Endangered Bhili and Mundari Languages

Addressing the severe lack of evaluation standards for tribal and vulnerable Indic languages, this paper introduces AdiDhwani, a preservation-oriented ASR benchmark for Bhili and Mundari (68.2 hours of verified speech), alongside novel diagnostics (HNB/HNBA, MFR, SDR) that reveal how standard edit-distance metrics (WER/CER) silently reward dominant-language lexical replacement and grammatical suffix deletion.

ADORA: Training Reasoning Models with Dynamic Advantage Estimation on Reinforcement Learning

ADORA dynamically calibrates policy gradient advantage estimation during online rollout sampling by evaluating length and difficulty advantages, applying attenuation to denoise weak reasoning models (VLMs) and amplification to break plateaus in strong reasoning models (LLMs), achieving state-of-the-art reasoning performance with extreme sample efficiency (2k samples).

Advancing Fact-Checking for Large Language Models via Argumentative Reasoning

Addressing confirmation bias in single-agent retrieval and the lack of principled conflict resolution over heterogeneous web sources, this paper presents MAVEN, a real-world benchmark with fine-grained error taxonomy and ruling-style justifications, alongside ArgCheck, a neuro-symbolic framework coupling multi-round adversarial retrieval with Dung's abstract argumentation grounded semantics, boosting Macro-F1 over the strongest baselines by 5.39 on MAVEN and 11.10 on AVeriTeC.

AGENTEHR: Advancing Autonomous Clinical Decision-Making via Retrospective Summarization

Addressing the critical information loss and fractured reasoning continuity caused by unidirectional context compression in noisy real-world EHRs, this paper introduces the AGENTEHR benchmark covering six patient lifecycle tasks and proposes RETROSUM, a framework uniting retrospective summarization with an evolving dual-experience memory bank to retain cross-temporal clinical dependencies while reducing systematic interaction errors by up to 92.3%.

AGENTQ: Quantization-Conditioned Backdoor Attacks on LLM Agents

Addressing the widespread reliance on low-precision deployment for on-device and latency-critical LLM agents, AGENTQ establishes the first quantization-conditioned backdoor attack (QCA) in the agentic regime. By combining layer-banded LoRA injection with partial-PGD repair over a multi-codebook equivalence class (NF4, FP4, INT8), it ensures that full-precision weights pass standard audits while the quantized artifact triggers malicious structured tool execution at up to 100% success rate without degrading benign utility.

AgentTAC: Mining Temporal Action-Context Error Patterns from Agent Traces

AgentTAC shifts language model agent evaluation from isolated per-trace diagnosis to corpus-level temporal action-context error pattern mining, accurately distinguishing newly introduced errors from propagated cascades, dynamically expanding an attribute library across tasks, actions, and environments, and extracting recurring sequential failure patterns that support effective test-time agent monitoring and repair.

Agreement Is Not Truth: Characterizing and Controlling Consistent-but-Wrong Reasoning in LLMs

The paper proves a theoretical lower bound demonstrating that self-consistency answer histograms alone cannot distinguish modal-correct from modal-wrong reasoning instances, and introduces RISC (Risk-Controlled Self-Consistency)—a framework that fuses verifier signals with majority ratios and calibrates acceptance thresholds via Learn-then-Test to guarantee finite-sample risk bounds while cutting API spend by ~9x.

AI Writers Have a Consistent Stylometric Footprint, but AI Editors Do Not

By introducing the GEN benchmark of 45,000 fully AI-generated and 273,420 AI-edited texts across 5 domains and 8 LLMs, this work demonstrates that AI text generation possesses a parsimonious and consistent stylometric footprint driven jointly by elevated entropy and lexical diversity; in stark contrast, AI editing does not mirror this footprint, instead showing an entropy drop and a steep decline in lexical density, proving that AI text is not a monolithic phenomenon.

Apples to Apples? Towards Comparable Crosslingual Language Model Evaluation

This paper demonstrates that popular normalized intrinsic metrics (PPL, BPB, BPC, Token-NLL, and MRR) introduce severe systematic crosslinguistic biases rooted in tokenization granularity, byte encodings, and orthographic conventions, proving theoretically and empirically that unnormalized sentence-level negative log-likelihood (Sent-NLL) evaluated on parallel texts is the only unbiased metric for crosslingual model comparison.

APPSim-Bench: Bridging Real-world Apps and Reproducible Evaluation for Mobile GUI Agents

AppSim-Bench introduces a reproducible, high-fidelity dynamic benchmark comprising 557 tasks across 17 simulated mobile apps from Chinese and English ecosystems; evaluating 19 frontier and specialized GUI agents shows that the best model solves only 50.27% of tasks, highlighting long-horizon degradation and numerical constraint maintenance as the core bottlenecks.

AR-Bench: Benchmarking Legal Reasoning with Judgment Prediction and Precedent Analysis

This paper shifts legal AI from forward prediction to retrospective diagnostic reasoning (APPELLATE REVIEW), proposing a controlled benchmark of 8,700 criminal judgments across charge, prison term, and fine errors (AR-BENCH) and revealing that 16 state-of-the-art LLMs suffer from severe deficiencies in unguided anomaly detection and rule-based penalty calculations.

Are Online Skill and Memory Modules Always Worth Their Tokens? A Budget-Constrained Study of Web Agents

Under a strictly controlled total test-time token inference budget, a budget-matched vanilla actor with extended interaction steps (Vanilla-IB) matches or surpasses online workflow memory (AWM), programmatic skill induction (ASI), and reasoning memory (ReasoningBank) across WebArena and WorkArena-L1, exposing the substantial explicit and implicit token overhead of online auxiliary modules.

Are We There Yet? Assessing Computer-Use Agents for Blind Users’ Accessible Interaction with Desktop Applications

Through a three-week longitudinal diary study with 8 blind users and controlled cross-model re-execution over 1,258 authentic desktop commands, this paper benchmarks computer-use agents in nonvisual workflows, revealing a top success rate of only 52.5% (GPT-5) and uncovering critical breakdowns in grounding, hidden path discovery, and termination.

Ask, Condition or Abstain: Reinforcement Learning for Missing-Premise Reasoning

Addressing the failure mode where answer-only reinforcement learning causes models to hallucinate definite answers on underdetermined queries, this paper proposes ACA-RL and the human-verified MPB benchmark. ACA-RL constructs gap-annotated problems via reasoning-DAG-guided condition perturbations and optimizes a structured preference reward over five terminal behaviors, teaching models to proactively ask, formulate conditional answers, or abstain while preserving competitive standard reasoning.

Ask Now, Use Later: Benchmarking the Proactivity Gap in Long-Lived LLM Agents

This paper formalizes the Proactivity Gap in long-lived personal assistant agents and presents the Ask-to-Remember (ATR) paradigm along with ATRBench, a controlled benchmark decoupling acquisition from application across 8 frontier LLMs, revealing that prompt-based interventions recover at most 15.5% of the 62.1–76.6 point gap to an oracle.

AURA: Intent-Directed Probing for Implicit-Need Surfacing in Situated LLM Agents

AURA introduces an IntentFrame inference stage between scene perception and tool execution to estimate the gap between literal queries and underlying implicit needs, dynamically regulating probe budgets to improve implicit-need coverage by 7.1% over ReAct-style probing while cutting probe calls by 82% with zero forbidden-tool violations on privacy-sensitive factual tasks.

AVERT: Audio-Verified Adjudication for Spoken Dialogue State Tracking

Addressing persistent ASR entity corruption and the acoustic blind spots of text-only state editors in spoken dialogue state tracking, AVERT proposes a value-level adjudication pipeline that marries cross-turn consensus with a lightweight audio verifier and a lexical attestation gate, raising the JGA of a 1B system to 40.13% on SpokenWOZ without retraining either the base model or editor.

Bag-of-Tokens Tuning: Calibrating LLM Logits with Contextual Probabilities

Addressing the severe Context-Logit Probability Gap where large language models fail to align their output logits with probability distributions explicitly specified in context, this paper proposes Bag-of-Tokens (BoT) Tuning—a lightweight, synthetic soft-label alignment method that directly supervises output distributions via KL-divergence minimization, delivering dramatic sampling fidelity gains alongside consistent improvements in downstream QA accuracy and uncertainty calibration.

Bayesian Control for Coding Agents

Formulating coding agent orchestration as a cost-sensitive Partially Observable Markov Decision Process (POMDP) enables dynamic Bayesian belief tracking of candidate correctness and value-of-information action selection, drastically reducing verification overhead while maximizing expected task utility.

BenchBrowser: Retrieving Evidence for Evaluating Benchmark Validity

BenchBrowser is an interactive retrieval infrastructure that searches item-level evaluation evidence across 20+ benchmarks for practitioner-defined use-cases, systematically uncovering pervasive threats to content validity and convergent validity in language model benchmarks.

Beyond Adapter Expansion: Packing Multiple Tasks into Spare LoRA Capacity

This paper challenges the common belief that continual learning requires expanding separate adapters for every task, demonstrating that a single LoRA module contains substantial per-task parameter redundancy and heavy-tailed update profiles. The proposed CAST framework adaptively allocates task-specific subspaces using an unsupervised geometric knee point criterion within a single shared LoRA adapter, freezes historical parameters, and routes inputs via lightweight semantic vectors, achieving accuracy on par with independent dedicated adapters across 15 tasks while maintaining a constant adapter footprint and over 24% unused capacity.

Beyond Conditional Computation: Retrieval-Augmented Genomic Foundation Models with Gengram

Gengram establishes a conditional memory module for genomic foundation models that explicitly stores and retrieves multi-scale (\(k=1\) to \(6\)) motif embeddings via collision-free hashing across a local causal window (\(W=21\) bp), injecting gated contextual representations into Transformer layers to eliminate redundant neural motif inference and stabilize MoE routing.

Beyond Dense States: Sparse Transcoders as Causally Testable Operators for LLM Latent Reasoning

Addressing the mechanistically opaque, black-box nature of dense representations in latent reasoning, this paper introduces LSTR, an in-loop framework that parameterizes latent state transitions via a Latent Transition Transcoder (LTT) combining a linear transport path with a Top-k sparse innovation path, preserving compression performance while exposing causally testable features for direct intervention.

Beyond Fresh Starts: Stateful Inference for Streaming ASR in Conversational Voice Agents

Addressing the dilemma where cascaded low-latency conversational voice agents reset ASR state at turn boundaries causing severe turn-onset word errors while continuous streaming corrupts acoustic memory via long pauses and backchannels, this paper proposes inference-time State Carry-over and State Rollback strategies that achieve a 15%–21% relative WER reduction at turn onsets with negligible memory overhead and zero retraining.

Beyond Input Understanding: Diagnosing Multilingual Mathematical Reasoning with Directed Acyclic Trace Graphs

This paper demonstrates that the severe multilingual mathematical reasoning gap in large reasoning models stems not just from non-English input misunderstanding but from trace-side execution breakdown; it introduces the Directed Acyclic Trace Graph (DATG) diagnostic framework alongside lightweight dynamic test-time thought supervision (Loop-Retry and Formula-Retry) that substantially recovers target-language reasoning performance.

Beyond Planning: Sparse Context Reset for Long-Horizon GUI Agents

This paper proposes SCOR (Sparse Context Reset), showing that long-horizon GUI agent failures stem primarily from context pollution and misaligned execution granularity (the Task Granularity Gap). By deploying a lightweight Refiner trained with turn-level GRPO to reset the frozen executor's accumulated context with locally executable subtasks and visual crops at sparse boundaries, SCOR boosts frozen EvoCUA-8B from 42.39% to 53.16% on OSWorld.

Beyond Scores: Understanding LLM-as-a-Judge Mechanisms in Summarization Evaluation

By introducing an eight-attack perturbation taxonomy across Readability and Adequacy dimensions alongside token-level activation patching, this paper reveals that LLM evaluators compute ratings via a two-stage mechanism—early attention detects and routes errors below layer 15, while upper-layer MLPs integrate signals to crystallize decisions at a sharp late layer—with fine-tuning acting as a sculptor on base model representations rather than building new circuits from scratch.

Beyond Seeing: Evaluating Multimodal LLMs on Tool-Enabled Image Perception, Transformation, and Reasoning

Addressing the saturation and passive nature of static multimodal benchmarks, this paper introduces VisualToolBench, the first evaluation suite of 1,204 expert-curated tasks with atomic multi-competency rubrics requiring active image manipulation (cropping, rotation, enhancement), revealing that frontier models reach less than 30% overall accuracy with visual perception serving as the primary bottleneck.

Beyond Static Grounding: Behavioral Contracts for Trustworthy Retrieval-Augmented Generation

Addressing the blind spot of static RAG evaluations that operate on fixed contexts and cannot detect whether systems causally depend on retrieved passages, this work introduces Dynamic Evidence Contracts with the Intervention Compliance Score (ICS) and presents DOVE-RAG, a single-call reference pipeline using NLI-based verification and deterministic three-tiered reorganization to achieve explicit, structured evidence adaptation.

Beyond the Illusion of Consensus: From Surface Heuristics to Knowledge-Grounded Evaluation in LLM-as-a-Judge

This work exposes the "Evaluation Illusion" in LLM-as-a-Judge frameworks—where high inter-judge agreement is an artifact of shared surface heuristics and common rubric instruments rather than substantive quality consensus—demonstrating that independent knowledge grounding via the MERG protocol slashes consensus by over 62% and revealing a severe mismatch between aggregate rankings and fine-grained preference judgments.

Beyond the Stability-Exploration Dilemma: Environmental Regularization for LLM Policy Optimization

Addressing the stability-exploration trade-off where action-side Policy-KL penalization throttles exploration in LLM reinforcement learning, this paper proposes Environment-Regularized Policy Optimization (ERPO), shifting regularization to the input side via Query-KL and reference-prior reweighting to decouple exploration while quelling environment non-stationarity and reward hacking.

Bias Before Disclosure: A Behavioral Audit of LLM Assistants Under User-Identity Uncertainty

This paper demonstrates that LLM assistants systematically assign demographic identities to users even on completely neutral information-seeking queries (unsupported demographic commitment), skewing heavily toward default privileged groups across 18 audited models, and proposes a dual explicit/free-form audit pipeline alongside an identity-abstention prompt that cuts implicit presupposition from 41% to 16%.

Bias in the Tails: How Name-conditioned Evaluative Framing in Resume Summaries Destabilizes LLM-based Hiring

By decomposing nearly one million resume summaries into factual content and evaluative framing, this paper demonstrates that race-gender name perturbations induce subtle, tail-concentrated framing variations that transform directional demographic advantage into severe symmetric decision instability in downstream LLM hiring.

BORA-BUDGET: Learning When to Think Under a Token Budget

Addressing the excessive token cost and harmful override risks of uniformly invoking expensive reasoning in hybrid-thinking models, BORA-BUDGET introduces a seed-preserving control framework that predicts marginal gains and risks from cheap seed outputs, allocates thinking via AutoThreshold under explicit token budgets, and filters updates through an adoption gate.

Bottom-up Policy Optimization: Your Language Model Policy Secretly Contains Internal Policies

By decomposing the LLM policy across the Transformer residual stream into layer and modular policies, this paper identifies a universal exploration-to-convergence transition and introduces Bottom-up Policy Optimization (BuPO), which aligns critical internal boundary layers during early RL training to reconstruct foundational reasoning representations and substantially boost complex reasoning performance.

Bottom-up Policy Optimization: Your Language Model Policy Secretly Contains Internal Policies

By projecting Transformer residual streams with the unembedding matrix, this paper decomposes an LLM into internal layer and modular policies, uncovers divergent exploration-to-convergence entropy dynamics across model families, and proposes Bottom-up Policy Optimization (BuPO) to pre-align foundational boundary layers before global RL training, substantially boosting mathematical reasoning at negligible diagnostic cost.

Branch Later, Compute Less, Think Better: Adaptive Depth Parallelism for Efficient Parallel Prefix-Tuning

Adaptive Depth Parallelism (ADP) identifies the functional division of Transformer depth, executing a single-path shared backbone in shallow feature-extraction layers and expanding into multi-path prefix-tuning branches only after an adaptively detected bifurcation layer, achieving 104.83% of full-stack parallel performance with only 39.4% of the computational load.

Branch Later, Compute Less, Think Better: Adaptive Depth Parallelism for Efficient Parallel Prefix-Tuning

Addressing the severe computational redundancy in full-stack model-internal parallelization frameworks (e.g., ParScale), this paper proposes Adaptive Depth Parallelism (ADP), which leverages the functional hierarchy between shallow feature extraction and deep reasoning in Transformers to share a single-path backbone in early layers and adaptively branch out into parallel prefix paths only after a diagnosed bifurcation layer, achieving 104.83% of baseline performance with only 39.4% of the computational load and 47.5% of the memory.

Bridging Text and Structure via Latent Query Refinement for Knowledge Graph Completion

Addressing the rigid, one-way, and static coupling between textual semantics and relational topology in knowledge graph completion (KGC), LQ-KGC introduces an iterative latent-query refinement framework on a compact, type-aware query set—alternating between injecting high-confidence text pseudo-tails as graph edges and distilling graph structural distributions back into the language model—achieving notable gains on WN18RR and FB15k-237.

Bridging the Temporal Gap: A New Benchmark and Dynasty-aware Adaptation Framework for Ancient Chinese WSD

Addressing the long-standing diachronic semantic evolution across two millennia of Ancient Chinese, this paper introduces AC-DIAWSD—a large-scale benchmark covering 12 dynasties and ~24K instances from the Twenty-Four Histories—and DynaSense, a model-agnostic adaptation framework combining hypernetwork-based temporal conditioning and diachronic contrastive alignment that achieves state-of-the-art accuracy and macro-F1 across diverse PLMs and fine-tuned LLMs.

Bridging Time Series and Language via Query-Adaptive Multi-Granularity Routing for Temporal Reasoning

Addressing the mismatch between static time-series representations and query-dependent temporal evidence scopes in Time-Series Question Answering (TSQA), TiMER freezes both the LLM and the multi-granularity time encoder, employing a Query-Adaptive Granularity Router (QGR) and a Granularity-aware Temporal Experts module (GTE) to dynamically synthesize local, global, and cross-granularity features, achieving 81.02% overall accuracy on TSAQA with only 0.08% external trainable parameters.

Gendered Prompting and LLM Code Review: How Gender Cues in the Prompt Shape Code Quality and Evaluation

This paper investigates how gender-related linguistic cues in prompts influence LLM code generation and automated code review across three complementary studies, demonstrating that while gender-coded prompting does not impair functional code correctness, automated LLM reviewers exhibit significant approval bias favoring female-authored code despite identical objective quality.

Can Crowdsourcing Survive the LLM Era? A Community Survey on Human Data Collection

Addressing the crisis of undisclosed LLM usage in crowdsourced free-text data collection, this paper presents the first comprehensive community survey of 155 researchers across NLP and adjacent fields, identifying that 44% observe LLM use while facing severe methodological gaps in mitigation, systematizing 28 qualitative impact themes, and deriving an actionable lifecycle framework spanning restricted, controlled, and unrestricted usage regimes.

Can Large Language Models Handle Discourse Particles? A Case Study of Colloquial Malay

To tackle the struggle of LLMs in capturing subtle pragmatic nuances of discourse particles in low-resource Southeast Asian languages, this paper introduces the MALAYPRAG benchmark along with a theoretically grounded five-attribute scaffolding framework, substantially elevating model accuracy across pragmatic function inference and particle slot-filling.

Can Persona-Conditioned Language Agents Simulate Counterfactual Online Discourse Dynamics?

Addressing the failure of prompt-driven language agents to capture longitudinal discourse evolution, this work presents PRISM, a state-space framework decoupling latent affective SDE evolution, POMDP participation control, and controlled text generation to deliver high-fidelity discourse forecasting and counterfactual interventions on 47k social media threads.

Can Reasoning Models Detect Changes to their Chains of Thought?

The paper presents a comprehensive empirical evaluation of whether contemporary reasoning models can detect and localize modifications made to their chains of thought (CoTs), demonstrating that unalerted models perform at chance level, self-awareness provides no measurable advantage over inspecting foreign CoTs, and fine-grained localization fails outside of blatantly irrelevant sentence insertions.

Can Segmentation Models Understand the World? Proactive Affordance Reasoning via Vision-Grounded Chain-of-Thought

Addressing the limitation that conventional segmentation models passively wait for target-referential queries, this paper proposes SegWorld, a proactive affordance reasoning framework that generates an instruction-free scene prior before unrolling an eight-level visual chain-of-thought to ground high-level conversational intents into affordance-bearing object parts on the novel Intent2Part benchmark.

Cascade-SC: A Pareto-Efficient Cross-Model Cascade for Test-Time Reasoning

Cascade-SC introduces a containment-gated cross-model cascade for asymmetric test-time reasoning that accepts the strong teacher's answer when matched by the weak generator's candidate set and falls back to teacher-only majority voting otherwise, achieving Pareto-optimal accuracy–compute trade-offs.

Cast a Wider Net: Coordinated Pass@K Policy Optimization for Code Reasoning

Addressing the mode-collapse and redundant rollout issue of standard independent repeated sampling in code generation, this paper introduces Coordinated Pass@K Policy Optimization (CPPO), an RLVR framework that autoregressively generates complementary strategy tuples and updates a joint planner–solver policy using a validity-gated execution outcome reward.

Causal Tongue-Tie: LLMs Can Encode Causal Direction, But Their Yes/No Outputs Fail to Express

The paper reveals Causal Tongue-Tie, an empirical mismatch where LLM hidden states accurately encode evidence-supported causal directions (probe accuracy 0.97–1.00) while final Yes/No answers revert to commonsense priors (accuracy 0.35–0.53), and localizes the bottleneck to non-contrastive verdict interfaces via subspace erasure, activation patching, and 2×2 controls.

CharLOTTE: Orthographic Correspondence-Based Data Unification Enhances Cross-Lingual Transfer in Low-Resource NMT

This paper proposes CharLOTTE, a completely automatic and language-agnostic pipeline that mines cognates from parallel data to train character-level orthographic correspondence models, reshaping high-resource parent languages to match the orthography of related low-resource child languages and boosting downstream NMT by up to 5.53 BLEU.

ChildVox: A Speech, Audio, and Large Audio-Language Model Benchmark in Understanding and Characterizing Sound across Childhood

ChildVox establishes the first comprehensive child-centered acoustic foundation model benchmark spanning from birth through school age across 17 corpora and over 20 sub-tasks, systematically revealing complementary capabilities and significant blind spots among self-supervised audio encoders, ASR models, and large audio-language models on physiological sounds, early vocalizations, and disordered speech.

ChJGBench: A Comprehensive Benchmark for Chinese Legal Judgment Generation with Diagnostic Evaluation

This paper presents ChJGBench, the first large-scale Chinese legal judgment generation benchmark covering all three litigation types (civil, criminal, and administrative across 3,462 real judgments), paired with a case-type-aware diagnostic framework across 13 metric dimensions, demonstrating that domain-specialized legal models lack a universal advantage over frontier general LLMs in complex civil and administrative adjudication.

CirCA: Aligning Credit Assignment with Circuit Semantics via Atomic Logic Blocks for RTL Generation

Addressing the severe misattribution in RL-based Verilog generation caused by overly coarse sequence-level assignment or fragmented line-level splitting, CirCA introduces Atomic Logic Blocks (ALBs) as semantically complete optimization units, localizes faulty blocks through semantics-aware simulation and fingerprint-guided dependency cone matching, and integrates localized penalties into GRPO via asymmetric advantage correction.

CLASP: A Closed-Loop Behavior Alignment Framework for Social Media Agent Simulation

CLASP introduces a decoupled dual-model closed-loop reinforcement learning framework that dynamically refines explicit user personas via behavioral prediction discrepancies, achieving up to a 31.8% alignment quality gain and 32.1% variance reduction over long-horizon simulations on the newly constructed 20-million-action BlueTrack dataset.

ClueTracer: Question-to-Vision Clue Tracing for Training-Free Hallucination Suppression in Multimodal Reasoning

Proposes ClueTracer, a training-free and architecture-agnostic plugin for Multimodal Large Reasoning Models (MLRMs) that suppresses reasoning drift and hallucinations by tracing task-decisive evidence along a question-to-output-to-vision attention pathway, achieving up to 1.25× and 1.17× accuracy gains on HallusionBench and VMCBench.

Code-TVT: Synthesizing Test-Case-Driven Verifiable Tasks for Code Reinforcement Learning

Code-TVT reformulates existing program test cases from passive post hoc evaluators into intrinsic, execution-verifiable training environments—featuring edge test-case generation and program output inversion with gated anti-hacking checks—substantially boosting rejection-sampling fine-tuning and GRPO code reinforcement learning across LiveCodeBench benchmarks.

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

BiG-SURE addresses black-box LLM/VLM uncertainty estimation by pairing stable low-temperature responses as semantic anchors with high-temperature responses under meaning-preserving input perturbations as probes, constructing an NLI-weighted bipartite graph and quantifying uncertainty via the complement of normalized squared spectral energy, outperforming prior black-box estimators across text, multilingual, and multimodal QA abstention tasks.

CoDiQ: Test-Time Scaling for Controllable Difficult Question Generation

Pioneers depth-wise test-time scaling for question generation through CoDiQ, integrating six cognitive scaffolding strategies, hybrid difficulty-solvability verification, and breaking-point reinforcement learning to synthesize competition-grade reasoning problems with high fidelity.

CoFGraph: Interleaving Reciprocal Calibration with Confidence Feedback for LLM-Augmented Graph Learning

CoFGraph introduces a confidence-feedback graph learning framework that eliminates the noise propagation and severe inference latency of static, unidirectional LLM-to-GNN pipelines. By distilling a frozen LLM into a lightweight isomorphic Interpreter offline and establishing an instance-level confidence-driven reciprocal calibration between semantic and structural views, CoFGraph achieves consistent state-of-the-art accuracy across four benchmarks, cuts test-time LLM inference bottlenecks to zero, and exhibits remarkable robustness against semantic noise.

CoMAP: Co-Evolving World Models and Agent Policies for LLM Agents

Addressing the fundamental bottlenecks that textual world models freeze after training—causing severe distribution shift—and that policy self-improvement heavily relies on external rewards or verifiers, CoMAP presents a closed-loop co-evolutionary framework. In this framework, future-aware reflection refines agent actions prior to execution, and on-policy execution trajectories continually adapt the world model via teacher-student self-distillation, mutually enhancing both policy execution and environment modeling without external annotations.

CommunityBench: Benchmarking Community-Level Alignment across Diverse Groups and Tasks

This paper proposes community-level alignment as a scalable middle ground between universal and individual alignment, introducing CommunityBench—a benchmark spanning 6,919 communities and 12,149 instances across four tasks grounded in Common Identity and Common Bond theory—evaluating 17 foundation models to reveal critical failures in distributional reasoning and demonstrating that community signals transfer effectively to individual social simulation.

Compute Allocation in Evolutionary Search: From Depth–Breadth to Multi-Armed Bandits

This paper provides the first systematic study of fixed-budget compute allocation in LLM-guided evolutionary search, revealing capability-envelope collapse under effective FLOPs along with a bilinear depth–breadth surface, and proposes BaSE, a multi-armed bandit framework that dynamically schedules compute across trajectories to boost expected fitness by 12.3% without altering models, prompts, or evaluators.

Consensus Drift: Semantic Salience Can Override Majority Evidence in LLM Aggregation

This paper uncovers the "consensus drift" phenomenon in large language models: while symbolic majority voting exhibits near-zero minority override, introducing natural language statements allows semantically salient or authoritative minority evidence to systematically overturn a numerical majority.

Consensus Group Relative Policy Optimization for Text Generation

Consensus Group Relative Policy Optimization (C-GRPO) distills computationally expensive inference-time Minimum Bayes Risk (MBR) decoding into a single-pass policy by formulating within-group consensus utility as group-relative advantages in GRPO, matching or exceeding multi-sample MBR performance without gold references or explicit preference labels.

ContextRouter: Learning When to Use Context for Efficient LLM Machine Translation

ContextRouter formulates LLM-based machine translation as a test-time intervention routing task, employing group-relative preference supervision over lightweight semantic and structured features from a zero-context draft to dynamically choose between direct translation, context integration, or refinement—achieving superior quality at substantially lower token cost than static policies and QE reranking.

Controlling Creativity in Concept Space: A Framework for Generative Models

This paper introduces Concept-Level Temperature (Concept-\(\tau\)), a model-agnostic framework that identifies underspecified semantic attributes via structured concept graphs and maps a continuous creativity parameter into discrete surprise levels using convex renormalization functions, achieving monotonic linear, high-diversity creativity modulation without compromising input utility.

COPILOT: Controlling Fine-grained Provider Preference in LLM Recommendations

Addressing the severe cross-intent contamination and distractor clutter of LLM recommenders under complex requests, COPILOT establishes a training-free "Plan-Draft-Complete" pipeline that decomposes sub-tasks, injects preferred services as hard anchors into a diffusion language model's initial mask to generate structural drafts in parallel, and refines them via the target LLM, boosting preference adherence by up to 58.9% while cutting irrelevant mentions by up to 87.8%.

CORE-Bench: A Comprehensive Benchmark for Code Retrieval in the Era of Agentic Coding

CORE-Bench establishes a three-level repository-aware code retrieval benchmark tailored for coding agents—spanning decontextualized code understanding, issue-to-edit localization, and broader auxiliary context retrieval—revealing severe performance collapses in standard embedding models and providing an in-domain PR-supervised contrastive tuning recipe that significantly closes the agentic retrieval gap.

Correct Prediction, Wrong Steps? Consensus Reasoning Knowledge Graph for Robust Chain-of-Thought Synthesis

Addressing the pervasive problem where LLMs predict the correct final answer via flawed intermediate reasoning steps, this paper introduces CRAFT, a training-free framework that constructs a consensus Reasoning Knowledge Graph across candidate chains and synthesizes a purified, high-accuracy reasoning trace via topological traversal.

Counter with Evidence! A Multi-Agent Memory Efficient Reasoning Framework for Hate Category Informed Counterspeech Generation

Addressing the core flaws of monolithic LLMs that treat hate speech homogeneously and lack factual grounding, this paper introduces FIRE, a lightweight two-stage multi-agent reasoning framework that decouples hate intent diagnosis from counterspeech generation, leveraging a contrastive memory store and uncertainty-aware arbitration to achieve state-of-the-art factual accuracy and category alignment using sub-2B models with a 75% peak memory reduction.

CPJudgeBench: Meta-Evaluation of LLM Judges for Constraint Programming through Solution-Space Semantics

Addressing the absence of formal semantic ground truth in code evaluation for constraint programming, CPJudgeBench establishes a solver-supported meta-evaluation benchmark across 522 validated candidate models, systematically dissecting the capabilities and biases of 15 LLM judges through projected solution-space semantics.

Criterion-Activated Preference Synthesis: Attribution-Preserving Preference Data from Explicit Standards

To resolve preference attribution ambiguity where binary chosen-rejected pairs conflate style, length, and correctness, CAPS prospectively synthesizes contexts from explicit standards, decouples rubric adherence from holistic quality via dual-signal scoring, and purifies pairs through group-level Empirical Activation Tests (EAT) and pair-level Attribution Ratios (AR)—achieving superior performance across general, medical, and legal benchmarks with only 20k synthetic pairs compared to 35k human HelpSteer3 pairs.

CROG: Coverage-aware Reasoning Optimization on Graphs for Multi-hop Question Answering

Addressing the issue of evidence redundancy and incomplete reasoning coverage caused by independent path ranking in graph-based multi-hop QA, CROG formulates evidence organization as a constrained submodular coverage maximization problem and greedily constructs compact, complementary evidence subgraphs based on marginal coverage gain.

CROG: Coverage-aware Reasoning Optimization on Graphs for Multi-hop Question Answering

Addressing the issue of severe semantic redundancy and missing reasoning hops caused by independent path ranking in graph-based multi-hop QA, CROG formulates evidence organization as a constrained coverage maximization problem, greedily selecting complementary paths by marginal reasoning gains under node-budget and connectivity constraints to boost reasoning quality with fewer evidence units.

Cross-Lingual LLM-Judge Transfer via Evaluation Decomposition

This paper proposes a decomposition-based evaluation framework built around Universal Criteria Sets (UCS), which factors holistic quality assessment into language-agnostic, concept-level numerical representations and enables sample-efficient cross-lingual LLM-judge transfer using a lightweight predictor trained purely on English data without target-language supervision.

Cross-lingual Representation Learning via Centroid Intervention Fusion

To overcome the parameter explosion and knowledge fragmentation inherent in pairwise cross-lingual activation intervention, this paper proposes Centroid Intervention Fusion (CIF), which reformulates projection matrices as residual perturbations around identity and applies centroid-guided trimmed aggregation to yield a unified, language-shared operator, achieving consistent improvements across multilingual reasoning and generation with frozen LLM parameters.

Cube Bench: A Controlled Diagnostic for Spatial and Sequential Reasoning in Multimodal Large Language Models

Cube Bench provides an oracle-scored diagnostic built on the Rubik's Cube with deterministic transitions and exact distances-to-solved, demonstrating that current MLLMs fail to benefit from vision-text fusion, experience depth collapse in causal forecasting, and cannot self-correct after mistakes (with immediate recovery rates universally dropping to zero).

Cultural Evolution Drives the Emergence of Conventions in Neural Emergent Communication

This paper proposes a decentralized cultural evolution framework for bi-directional neural agents, demonstrating for the first time in an unsupervised and fully distributed multi-agent population that communicative conventions emerge from local in-context reference games, and proving that conventionalized languages substantially lower the learning barrier for cultural transmission.

CYPHERTURN: A Multi-Turn Benchmark for Conversational Text-to-Cypher Evaluation and the Autonomy Divergence

Introduces CYPHERTURN, the first multi-turn conversational Text-to-Cypher benchmark comprising 721 sessions and 5,927 turns across 7 knowledge graphs, revealing the "Autonomy Divergence" phenomenon where frontier model rankings reorder under autonomous operation due to cascading error management.

D3-Gym: Constructing Real-World Verifiable Environments for Data-Driven Discovery

This paper introduces D3-Gym, the first automatically constructed benchmark of 565 executable and verifiable environments drawn from 239 real scientific repositories across 4 disciplines. By synthesizing task-specific evaluation scripts through a two-stage planning-then-coding pipeline, D3-Gym enables rigorous trajectory verification and rejection sampling fine-tuning that improves Qwen3-32B by +7.8 points SR@3 on ScienceAgentBench, approaching frontier proprietary models.

Data Efficient Sample Selection for In-Context Learning

Addressing the generalization limits of static selection and the latency of online search, DEARICL recasts demonstration subset selection as an instance-level ranking problem, utilizing a multi-armed bandit top-\(m\) auxiliary objective to sample informative borderline arms and train a non-linear differentiable sorting surrogate with 63%–75% fewer LLM queries and up to 15.99% accuracy improvement.

DCGC: Draft-Conditioned Global Correction for Complex Reasoning with Masked Diffusion Models

DCGC introduces a draft-conditioned global correction framework using Masked Diffusion Models (MDMs), combining mixed-format supervised fine-tuning (SFT) with a novel Dynamic Dual-CFG mechanism that dynamically modulates the draft-conditioned residual using a token-level relative confidence gap to achieve robust, verifier-free post-hoc reasoning correction.

DCGC: Draft-Conditioned Global Correction for Complex Reasoning with Masked Diffusion Models

DCGC repurposes Masked Diffusion Models (MDMs) as verifier-free global correction modules for complex reasoning via dual-capability SFT and Dynamic Dual-CFG, dynamically scaling draft-conditioned residuals using a relative confidence gap to achieve 24.8% average accuracy on solver-failure hard sets across math, code, and knowledge reasoning.

Debating the Unspoken: Role-Anchored Multi-Agent Reasoning for Half-Truth Detection

To tackle subtle half-truths where statements are factually correct in isolation yet deeply misleading due to omitted context, RADAR establishes a role-anchored multi-agent debate between stance-neutral Politician and Scientist agents over shared noisy retrieved evidence, orchestrated by a dual-threshold early-stopping judge to uncover unstated assumptions with minimal inference cost.

DECENT: Distribution-aware Prototypical Calibration with Fine-grained Voting for VLM Misclassification Detection

Addressing the severe failure of conventional similarity-based confidence scores on boundary points and fine-grained visual categories, DECENT proposes a training-free dual-branch framework that couples multi-template cross-modal semantic alignment with neighborhood distribution distance via energy statistics and fine-grained prototypical soft-voting, delivering state-of-the-art gains for both classification robustness (DECENT-R) and misclassification detection sensitivity (DECENT-S).

Decoupled Physical Modeling and Execution for Physics Reasoning

Addressing conceptual breakdowns in multimodal LLMs caused by conflating physical modeling with mathematical calculations, this paper proposes an explicit two-stage framework that establishes structured physical modeling via curriculum SFT and refines physical grounding using rubric-guided GRPO.

DEEPCHART: How Far are LLMs from Faithful Data-Science Chart Generation?

Addressing the critical evaluation blind spot where visually plausible charts disguise severe underlying data errors, DEEPCHART introduces an expert-annotated benchmark of 1,482 instances across Academic, Finance, and Ecosystem domains that formalizes chart authoring as an Extract–Reason–Visualize pipeline, revealing that state-of-the-art LLMs suffer from severe hidden hallucinations and systemic failures during multi-step numerical derivation over long documents.

Demystifying Hybrid Thinking: Can LLMs Truly Switch Between Think and No-Think?

The paper reveals that current hybrid thinking LLMs fail to achieve clean separation between think and no-think modes by leaking implicit reasoning behaviors, and establishes a highly controllable two-phase training recipe by systematically dissecting four pivotal data and optimization factors.

DenseMixer: Improving MoE Post-Training via Precise Router Gradient

DenseMixer resolves the non-differentiability dilemma of standard TopK routers in Mixture-of-Experts (MoE) post-training by trading an extra forward pass across inactive experts for exact, unbiased router gradients via a Straight-Through Estimator (STE), delivering substantial gains across diverse 7B–30B MoE architectures without altering sparse inference efficiency.

DERELAB: Probing Defeasible Reasoning and Confirmation Bias in LLMs with a Generative Benchmark

DERELAB introduces a parameterized generative evaluation benchmark that generates multi-turn belief-updating dialogues with formally verified ground truth, revealing near-universal confirmation bias and a prominent "know-but-don't-output" metacognitive dissociation across current LLMs.

Detecting AI Impostors: How Do Middle Schoolers Identify LLM Agents in a Live Collaborative Setting?

This work introduces DoppelBot, a live cooperative social deduction game where middle schoolers detect LLM doppelgängers impersonating them and their peers, revealing that student detection accuracy increases from 79.3% to 86.7% over time as players transition from relying on superficial linguistic cues to leveraging shared contextual, group-level, and embodiment-based constraints.

DFlare: Scaling Up Draft Capacity for Block Diffusion Speculative Decoding

Addressing the capacity bottlenecks and rigid target feature conditioning in block diffusion speculative decoding, DFlare introduces lightweight layer-wise adaptive fusion, heterogeneous KV projections, and a progressive position-weighted loss, achieving up to 5.52× wall-clock acceleration across math, code, and chat benchmarks.

DIANOIA: Diagnostic Decomposition and Joint Optimization for Multi-Agent Reasoning

DIANOIA formalizes multi-agent reasoning gain into three measurable channels—coverage, fidelity, and synthesis—and implements them via a four-phase architecture (role-diverse proposers, grounded execution/verification, cross-review, and closed-loop synthesis) that matches or exceeds state-of-the-art multi-agent baselines while requiring approximately 5× fewer tokens on code generation.

Direct Diversity Optimization for Diverse Successful Trajectories in Preference Post-Training

DDO resolves policy collapse in sequential decision post-training by introducing Divergence-Tree Collection (DTC) to construct state-aligned multi-branch sets and the Reference-Relative Target-Odds (RTO) objective to calibrate relative probabilities across successful alternatives, achieving superior task success and strategy coverage across BabyAI, BabaIsAI, and WebShop.

directional structurality a lightweight proxy for reasoning data quality

This paper introduces directional structurality as a lightweight, training-free proxy for reasoning data quality, implementing a local paired negative log-likelihood gap metric (Local-DeltaAvg) via semantic reversal counterfactuals that consistently outperforms length, perplexity, problem difficulty, and LLM-judge heuristics under matched token budgets in continued pretraining and test-time reranking.

Discovering Hidden Gems in Model Repositories

This paper exposes the pervasive presence of overlooked, low-download "hidden gems" across open model hubs that consistently outperform popular foundation baselines, and introduces an accelerated Sequential Halving algorithm with correlated sampling and aggressive pruning that discovers top models with a 50× reduction in evaluation queries.

Disentangling Curriculum Learning in NLP: Towards a Unifying Taxonomy

Addressing more than a decade of empirical confusion and incomparability in NLP curriculum learning (CL), this paper establishes the first unified dual taxonomy that analytically disentangles difficulty evaluation (attribution source and task dependence across four quadrants) from training scheduling (formalized via expected training contribution \(\pi_t\), five retention regimes, and three monotonicity properties), backed by an extensive empirical survey of 107 strategies from 90 ACL papers.

Disentangling Semantics and Popularity: Adversarial Debiasing and Asymmetric Alignment for LLM-Enhanced Long-Tail Sequential Recommendation

Addressing popularity-induced semantic distortion and the erosion of semantic priors in LLM-enhanced sequential recommendation, ADAA introduces a decoupled two-stage framework combining adversarial popularity debiasing with head-tail asymmetric anchoring, improving long-tail recommendation accuracy with zero additional online inference latency.

Disentangling the Interpretive and Predictive Roles of LIWC: Controlled Substitution in Depression-Related Classification

Addressing the ambiguity surrounding the incremental predictive value of psycholinguistic dictionaries in modern multimodal depression detection, this paper introduces a controlled-substitution protocol across five English and Chinese corpora to isolate participant alignment, marginal distributions, and coordinate basis effects, revealing that LIWC provides negligible predictive gains over frozen speech and text embeddings but serves as an auditable, corpus-conditioned interpretive lens.

Distribution-aware Language Neuron Identification in Multilingual Large Language Models

This paper demonstrates that conventional methods thresholding positive activations discard essential density shapes and negative regimes, proposing Distribution-aware Language Neuron identification (DLN) via pairwise distribution overlap clustering to effectively isolate single-language neurons with up to 4.9x higher on-target damage per neuron and discover cohesive multi-language coalitions.

DLawBench: Evaluating LLMs Through Multi-Turn Legal Consultation

DLawBench addresses the reality that clients often present fragmented, subjective, and legally misframed narratives by creating a diagnostic benchmark grounded in 461 real Chinese and U.S. court cases with decoupled client-belief and court-record views, showing that 26 representative LLMs struggle severely with non-ideal clients and exhibit widespread, professionally packaged legal sycophancy.

Do Latent Tokens Think? A Causal and Adversarial Analysis of Chain-of-Continuous-Thought

By employing probe-guided orthogonal perturbations, cross-sample token swapping, and adversarial shortcut manipulations, this paper reveals that latent tokens in continuous chain-of-thought (COCONUT) do not faithfully encode multi-step reasoning, but instead act as non-semantic placeholders that substantially exacerbate shortcut exploitation.

Do LLMs Cross Alphabets? A Cross-Script Audit of LLMs Across Four Kazakh Latin Proposals

Evaluating 13 open-weight and proprietary LLMs across four official Kazakh Latin proposals reveals an average accuracy drop of 16.8 points from the Cyrillic baseline, disproves the prevailing tokenization explosion hypothesis by showing Latin Kazakh is often more token-efficient, and demonstrates that prompt mitigations fail while targeted LoRA fine-tuning recovers up to +20.0 points when paired with reasoning-capable models.

Do LLMs Trust the Accuser or the Accusation? Measuring Belief Shifts in Werewolf

This paper introduces a belief-shift evaluation benchmark in Werewolf by annotating 1,224 accusation messages across 200 games and evaluating 40 open-weight LLM configurations, revealing that models up to 120B blindly trust accusation claims and struggle to reconcile source reliability with message content.

Do Vision-Language Models Understand Human Engagement in Games?

This paper presents the first systematic evaluation of vision-language models on inferring human engagement from gameplay video across 9 FPS games, discovering that VLMs achieve only ~57% zero-shot accuracy (lagging behind majority baselines) due to a fundamental perception-understanding gap and reliance on visual spectacle shortcuts.

Does Faithfulness-Guided Alignment Hurt Accuracy? Unlocking Accurate and Faithful Post-Retrieval Reasoning

Addressing the "right-answer-wrong-reason" shortcut dilemma in multi-hop retrieval-augmented generation (RAG), this paper presents CRAFT, a reinforcement learning framework that couples structured, machine-auditable reasoning traces with a dual-reward design (deterministic rules plus decoupled LLM semantic judge auditing), demonstrating that faithfulness alignment promotes rather than sacrifices task accuracy from 1.5B parameters upward.

Does Sensitivity to Pragmatic Norms Emerge with More Parameters, More Data, or More Conversations?

Evaluating autoregressive Pythia models across parameter scale (70M–2.8B), 18 developmental checkpoints, and four 10M-word matched conversational data regimes under a binary Gricean maxim choice paradigm, this paper reveals that parameter and data scaling deliver inconsistent, non-monotonic pragmatic gains, whereas targeted exposure to structured conversational data yields dramatic improvements of up to +17% average accuracy in low-budget settings.

Don't Read Everything: A Curvature-Conditioned Query for Linear Attention

Addressing the failure of linear attention models in long-context retrieval caused by the lack of Softmax competitive normalization, this paper proposes Curvature-Conditioned Query (CCQ), an anisotropic query contraction based on running key covariance that modifies only the read step while preserving constant \(O(1)\) inference cache.

DR^3-Eval: Towards Realistic and Reproducible Deep Research Evaluation

DR^3-Eval establishes a realistic and reproducible benchmark for Deep Research Agents by grounding report-generation tasks in authentic multimodal user files and pairing each task with a static, per-case sandbox corpus containing supportive, distractor, and noise web pages, systematically exposing the evidence retrieval and factual grounding bottlenecks of frontier language models.

Draft-OPD: On-Policy Distillation for Speculative Draft Models

Addressing the offline-to-inference mismatch where supervised fine-tuning (SFT) of speculative draft models quickly plateaus, Draft-OPD introduces on-policy distillation via target-assisted rollout, verification error replay, and acceptance-aware asymmetric loss, delivering over 5× lossless inference speedup.

DreamReasoner-8B: Block-Size Curriculum Learning for Diffusion Reasoning Models

DreamReasoner-8B reveals that naive large-block training causes severe reasoning degradation in block diffusion models, and overcomes this limitation via block-size curriculum learning (fine-grained block size 4 transitioning to 32) paired with a context-aware RelaxedConfidence decoding probe, matching state-of-the-art autoregressive reasoners like Qwen3-8B-Thinking on AIME and LiveCodeBench.

DSR-SQL: Dual-Schema Reasoning for Text-to-SQL and PI-MCE Optimized Schema Linking

DSR-SQL addresses the fundamental tension between pruned-schema information masking and full-schema contextual distraction in small open-source LLMs by deploying dual-path reasoning (precision-first vs. recall-first), a Permutation-Invariant Minimum Cross-Entropy (PI-MCE) loss for order-robust table selection, and execution-aware candidate merging, achieving 67.41% execution accuracy on the BIRD test benchmark with a 7B model.

DuQUAD: A Dual-View Framework with Quality-Aware Evidence Pruning for Multi-Document Question Answering

DuQUAD tackles the "Lost in the Middle" dilemma in multi-document question answering by pairing an SFS-guided Local Agent (leveraging second-order attention contrast, perplexity, and semantics) with a Global Agent, followed by LLM-based DocScore gating and sentence-level evidence compression, improving accuracy by 3.0% and gold evidence recall by 9.4% with substantial computational savings over existing multi-agent baselines.

DuraS2ST: Chain-of-Thought and Reinforcement Learning for Duration-Aligned Speech-to-Speech Translation

Addressing the critical challenge of acoustic duration mismatch in speech-to-speech translation, DuraS2ST reformulates duration alignment as a forward-planning reasoning task, generating an explicit chain-of-thought to deliberate lexical choices and phonetic sequence lengths, optimized via multi-modal GRPO with a smooth duration margin reward and modality-aware reward attribution.

DVD: A Robust Method for Detecting Variant Contamination in Benchmark Datasets

This paper tackles the critical problem of "variant contamination"—where training corpora contain semantically equivalent but lexically or structurally altered benchmark problems—by proposing DVD, a reference-free detector that quantifies the sample variance of synthetic difficulty across repeated temperature sampling, capitalizing on the bimodal oscillation between memory adherence and perturbation drift to achieve robust detection across models and domains (up to 0.771 AUC).

E3-Rewrite: Learning to Rewrite SQL for Executability, Equivalence, and Efficiency

E3-Rewrite is an end-to-end LLM-based SQL rewriting framework that breaks free from rigid rule-based paradigms by incorporating execution plan hints into context and optimizing executability, equivalence, and latency through a two-stage curriculum-guided GRPO framework, reducing query execution time by up to 25.6% over strong baselines.

Edit Knowledge, Not Just Facts via Multi-Step Reasoning over Background Stories

This paper reframes knowledge updating as a multi-step reasoning problem rather than an isolated fact memorization task, introducing coherent background stories to generate intermediate-dependent multi-hop questions and distilling reasoning behaviors via teacher-student context distillation, which dramatically enhances knowledge portability across complex downstream reasoning tasks.

Edits in the Wild: A Multilingual Benchmark for Localizing On-Device Generative Image Edits on Consumer Platforms

This paper introduces CONSUMERGENEDIT, a large-scale multilingual benchmark covering 4,000 source images, 48,000 real on-device AI-edited images across Apple, Samsung, and Google ecosystems, and 144,000 learned image compression variants (JPEG AI precursor) annotated in English, Bangla, and Farsi across five VLMs, revealing critical vulnerabilities in current vision-language models when detecting, localizing, and classifying consumer generative edits.

Elastic Horizon: Discovering the Effective Interaction Frontier in Agentic Reinforcement Learning

Addressing computational redundancy and training collapse caused by monotonic open-loop interaction schedules in agentic reinforcement learning, this paper proposes Elastic Horizon—the first closed-loop controller that estimates the effective interaction frontier via the 90th percentile of successful trajectory lengths and EMA smoothing, attaining state-of-the-art success rates while cutting per-step trajectory tokens by up to 25%.

EmoDap: Dual-Agent Prompt Retrieval Framework for Zero-Shot Fine-Grained Emotion Recognition

Addressing semantic interference and counterproductive demonstration retrieval in zero-shot fine-grained emotion recognition, EmoDap dynamically compresses the candidate emotion space via contrastively learned dictionary prototypes and jointly trains a demonstration retriever and emotion predictor via GRPO, outperforming top commercial LLMs across four standard benchmarks.

EmoLASP: Emotion Recognition with Language Models and Answer Set Programming

This paper proposes EmoLASP, a neuro-symbolic framework coupling language models with Answer Set Programming (ASP) declarative reasoning for conversational emotion recognition, which selects optimal Valence-Arousal-Dominance (VAD) hypotheses via weak constraints and outperforms full-history prompt baselines without dialogue history or fine-tuning.

Energy-Based Transformers as Predictors of Reading Difficulty

This paper introduces energy-based transformers (NRGPT) grounded in dense associative memory to computational psycholinguistics, demonstrating that a single scalar energy metric naturally unifies discrete next-token predictability and continuous memory integration effort, subsuming both surprisal and attention entropy across controlled relative-clause processing and three reading-time corpora.

Enhancing Long-Context Reasoning via Verifiable Atomic Skills

This paper challenges the conventional monolithic treatment of long-context reasoning by decomposing it into a hierarchical spectrum of five verifiable atomic skills, proposing the Anchor-based Reasoning (AbR) framework to automatically synthesize strictly verifiable training data, and demonstrating that Reinforcement Learning (GRPO) on merely 4,000 samples improves DeepSeek-R1-distill-32B by an average of 7.7% across six leading long-context benchmarks.

ERR+: Sequential Entropy Resolution for Efficient and Decisive LLM Reasoning

To address within-group advantage collapse under binary rewards and the accuracy degradation caused by naive length penalties in RLVR, ERR+ introduces a two-phase framework that first rewards token-level entropy drops via Entropy Relief Reward (ERR) and subsequently refines conciseness using within-group tanh-standardized length penalties (RRER), simultaneously improving reasoning accuracy and reducing generation length.

ERR+: Sequential Entropy Resolution for Efficient and Decisive LLM Reasoning

Addressing the issues of within-group advantage collapse and exploration suppression in verifiable-reward reinforcement learning (RLVR), this paper proposes ERR+, a two-phase framework that rewards cumulative token-level entropy resolution in Phase 1 and applies a robust peer-relative length penalty in Phase 2, resolving early gradient conflict via a theoretically proven sequential optimization schedule.

Evaluating Coding Agents on Kernel Exploit Generation

The authors introduce KEX-bench, the first benchmark evaluating coding agents on deterministic kernel exploit primitive generation across 40 real-world Linux and Windows CVEs, revealing that while frontier models frequently trigger kernel faults, turning raw crashes into controlled exploit primitives remains a fundamental bottleneck.

EVAR: Evidence-Validated Hypothesis Admission for Budget-Aware Narrative Reasoning

EVAR addresses the issues of intermediate hypothesis contamination and uniform revision costs in non-interactive long-form narrative reasoning by compiling an immutable source-linked atomic evidence store, dynamically routing budgets via instance complexity, gating hypothesis admission through multi-faceted validation challenges, and triggering sufficiency-based early stopping.

Every Answer Has Its Own Path: Agentic Routing for Retrieval-Augmented Table Question Answering

To tackle the core mismatch between heterogeneous document modalities in table-text QA, KAIST introduces ATR (AgenticTableRAG), an agentic framework that decomposes questions into sub-queries dynamically routed across four execution primitives, leverages value grounding to eliminate SQL hallucination, and deploys a distilled 66M student router that delivers superior accuracy with zero per-sub-query LLM call overhead.

EvoGenUI-Bench: Evaluating LLMs as Multi-Turn Generative UI Assistants

EvoGenUI-Bench introduces a multi-turn generative UI maintenance benchmark of 150 five-turn tasks (750 turns), executing artifacts in browser sandboxes to gather multi-surface evidence bundles (screenshots, DOM, Playwright actor traces, source code, and runtime logs) and demonstrating that even the strongest model falls from a 74.9% Turn Pass rate to a 37.3% five-turn completion rate.

Expert-Aware Causal Tracing of Factual Recall in Sparse MoE Language Models

Using a two-stage activation patching protocol under conditional fixed routing, this paper reveals that successful MoE-block factual recall restoration exhibits striking architectural asymmetry: Qwen3 displays a partial, nonunique recurrent singleton contribution with positive specificity, whereas Mixtral exhibits routed-set specificity without singleton specificity.

Exploring Confirmation Bias: LLMs Perceive Democratic Claims as More Truthful Independently of Their Content

This paper formalizes political bias in truthfulness perception (PB-T) within LLM-based fact-checking, evaluating 20 instruction-tuned models to reveal an omnipresent left-leaning confirmation bias where 16 out of 20 models judge claims as significantly more truthful solely because they are attributed to Democrats.

Extending AI for Research to the Humanities: A Multi-Agent Framework for Evidence-Grounded Scholarship

Addressing the core challenge of qualitative interpretation and primary-source verification in text-centered humanities scholarship, this paper introduces SPIRE, a multi-agent framework operationalizing "Scholarly Primitives" over a multi-scale close-reading infrastructure to achieve transparent, evidence-grounded research inquiry.

F-GRPO: Don’t Let Your Policy Learn the Obvious and Forget the Rare

Addressing the non-monotonic tail-miss bias and distribution sharpening caused by finite rollout groups in Group Relative Policy Optimization (GRPO), this paper proposes F-GRPO, an empirical success-aware difficulty scaling method that mitigates sharpening and dramatically improves pass@256 and solution diversity without requiring larger rollout groups.

FaithMed: Training LLMs For Faithful Evidence-Based Medical Reasoning

This paper introduces FaithMed, a reinforcement learning framework that formalizes five evidence-based medicine (EBM) principles into 15 refined checklist rubrics, employing anchor-state advantage grouping and step-level process reward modeling to train LLMs for faithful evidence retrieval, critical appraisal, and patient-specific reasoning—achieving a 9.0% average accuracy improvement over agentic-search baselines across seven medical benchmarks and boosting EBM rubric compliance by 15.5%.

FAMPWQ: Fisher Information-based Adaptive Mixed Precision Weight Quantization for Effective LLM Inference

Addressing the steep performance cliff of conventional uniform-precision PTQ at sub-4-bit regimes, FAMPWQ quantifies layer-wise degradation via diagonal Fisher Information Matrix (FIM) shifts induced by explicit quantization rounding perturbations, and solves global storage-constrained bit-width allocation in minutes using a lightweight PPO-guided proxy optimizer.

Fast-Slow Thinking RM: Efficient Integration of Scalar and Generative Reward Models

Inspired by Dual Process Theory, F/S-RM integrates scalar-style first-token judgment (fast thinking) and chain-of-thought generative reasoning (slow thinking) within a single backbone model, adaptively activating slow reasoning via a dual-confidence mechanism to achieve state-of-the-art accuracy (84.3 average) while cutting token consumption by 22.5%.

FBHM: Functional Benchmarking and Steering of VLMs for Hateful Meme Detection

This paper introduces FBHM, an orthogonally decoupled benchmark of 5,000 memes spanning 25 rhetorical functionalities and 10 protected communities, reveals that state-of-the-art VLMs collapse to near-random performance under structural shifts, and develops Learnable Steering Vectors (LSV)—a 131K-parameter causal activation intervention trained on only 500 samples—achieving over 30 points Macro-F1 improvement with zero test-time token overhead while guaranteeing demographic fairness.

FIDES: Faithful Inference via Deep Evidence Signals for Retrieval-Memory Conflict in RAG

FIDES is a training-free contrastive decoding algorithm that estimates retrieval-memory conflict at every autoregressive step by combining three complementary multi-depth signals—output distribution opposition, cross-path hidden trajectory shift, and within-path prediction evolution—to dynamically calibrate contrastive strength, outperforming competitive training-free baselines by up to +13 points in context fidelity without degrading utility.

Fixing FOLIO and MALLS: Verified Annotations and an LLM-assisted Framework to Focus Human Relabeling

A rigorous audit of prominent NL-to-FOL benchmarks FOLIO and MALLS uncovers over 40% annotation errors and widespread ambiguity; evaluating on corrected ground truth boosts LLM translation accuracy by up to 23 percentage points, while a proposed Verdict-and-Refinement human oversight framework achieves 90% dataset accuracy with under 20% manual review.

Forget Without Compromise: Nexus Sampling for Streaming KV-Cache Eviction Under Fixed Budgets

Addressing the permanent information loss in long-context streaming LLM inference under fixed memory budgets, Nexus Sampling replaces deterministic top-K truncation with weighted reservoir sampling and pairs it with an iterative walk that amplifies hidden bridge tokens, preventing monotone marginal erosion while matching dense attention within ~1 point on LongBench at 80% eviction.

Format Collapse in Policy Optimization for LLM Reasoning

By systematically relaxing rigid chain-of-thought structural constraints in verifiable-reward reinforcement learning, this paper demonstrates that policy optimization consistently collapses into an answer-only format without explicit reasoning, driven by an intrinsic "Least Resistance Bias" toward the most learnable gradient signal.

Fractional Rotation, Full Potential? Investigating Performance and Convergence of Partial RoPE

This paper delivers the first systematic empirical pretraining study of partial Rotary Position Embeddings (RoPE) across 1B and 8B scales over 100B+ tokens, revealing that rotating merely 10% of per-head hidden dimensions matches the convergence speed, final loss, and downstream reasoning accuracy of 100% full RoPE while cutting RoPE cache memory footprint by up to 10× in extreme long-context regimes.

From Accuracy to Robustness: A Study of Rule- and Model-based Verifiers in Mathematical Reasoning

This paper presents the first comprehensive diagnostic study of rule- and model-based verifiers in reinforcement learning with verifiable reward (RLVR) for mathematical reasoning, discovering that rule verifiers suffer severe false negatives that throttle capable policy models, while fine-tuned generative verifiers collapse under reward hacking during RL—ultimately establishing discriminative verifiers as the superior paradigm for accurate and robust reward systems.

From Expressions to Needs: Hierarchical Interpretive Reasoning for Emotional Support Conversation

Addressing the premature intervention and psychological misalignment of existing emotional support conversation systems that directly react to surface-level distress expressions, this paper proposes HierINT, a psychologically grounded hierarchical interpretive schema, along with the HierINT-CoRe collaborative reasoning framework and the HierINT-Dlg dataset, substantially boosting conversational empowerment and personalization.

From Global to Regional Coverage: Hyperspherical Prototype Discovery for Storage-Efficient Multimodal Retrieval

To tackle the severe index storage bottleneck in multi-vector visual document retrieval models like ColPali and ColQwen2, CovMerge introduces a geometry-aware, train-free index compression framework on the unit hypersphere that combines Farthest Point Sampling (FPS) prototype discovery with spherical K-Means refinement, retaining up to 98.1%–98.6% of original nDCG@5 at a tight 10% storage budget.

From Poetry Parties to Poetry Critics: Benchmarking Classical Chinese Poetry Generation and Evaluation

This paper constructs Honglou-Poem, the first narrative-driven classical Chinese poetry generation benchmark derived from Dream of the Red Chamber, alongside Poetry-Tree, an axiomatically ordered quality benchmark built via systematic merit degradation; using Poetry-Tree signals, the authors train Poet-4B via GRPO with Spearman's rank correlation reward, achieving a Spearman's \(\rho\) of 0.8253 and closely rivaling hundred-billion-parameter closed-source models with over 100x fewer parameters.

From Reasoning to Generalization: Knowledge-Augmented LLMs for the ARC Benchmark

This paper formulates the ARC benchmark as a program synthesis task, establishes repeated-sampling planning-aided code generation (RSPC) as the best-performing solver backbone, and proposes KAAR—a framework that structures core knowledge priors into a dependency-aware ontology to progressively augment reasoning LLMs, achieving substantial accuracy and generalization gains.

When Seekers Are Hard to Help: Evaluating Emotional Support Dialogue Systems in Worst-Case Interactions

Addressing the superficial evaluation ceiling and optimistic bias resulting from cooperative, average-case simulated users in emotional support dialogue systems (ESDSes), this work grounds worst-case interactions through an expert simulation study with licensed counsellors, constructs a controllable worst-case seeker simulator equipped with emotional deterioration, introduces four robustness-oriented evaluation metrics, benchmarks 17 representative models to reveal severe performance drops, and demonstrates that fine-tuning with worst-case synthetic data substantially enhances smaller open-weight models.

From Specialization to Generalization: Instruction-tuned LLMs for Robust Harmful Content Mitigation

This paper introduces HIPPO, an instruction-tuned generalist LLM for multi-task harmful content mitigation, unifying 36 heterogeneous English hate speech datasets across 61 subtasks into standardized conversational prompts to outperform specialized task-specific BERT ensembles and zero-shot LLM baselines.

From Token Probabilities to Semantic Constraints: Towards Declarative Probabilistic Evaluation of Language Models

This paper introduces MODELLOG, a declarative probabilistic framework that translates local autoregressive token predictions into propositional variables governed by symbolic logical constraints, measuring satisfaction via calibrated weighted model counting and revealing systematic logical consistency failures alongside inverse-scaling phenomena across prominent language model families.

From Value Conditioning to Behavioral Shift: Lightweight Value Alignment of LLMs

This paper introduces a lightweight value alignment approach that discards paired preference annotations and reinforcement learning, demonstrating that fine-tuning LLMs solely on scalar ratings from Schwartz value surveys—down-regulating target values while anchoring non-target values to baseline—consistently shifts probability distributions in contextual Reddit AITA moral judgments and sequential decision-making in MACHIAVELLI.

Frozen Cores Need Task Signal: Fisher-Whitened Cross-Covariance for Parameter-Efficient Adaptation

Addressing the challenge that low-rank adaptation methods obscure initial subspace quality due to movable factors, FCCA estimates signed input-error cross-covariance and diagonal Fisher moments during calibration, truncates singular directions in the local Fisher metric, and applies thin QR to construct rigid orthonormal bases, allowing fine-tuning of only an \(r \times r\) core with 36.9K parameters (roughly \(200\times\) fewer than LoRA) while matching full trainable adapter accuracy.

Fully Decoupled Reinforcement Learning with Adaptive Data Curation for GUI Agents

To tackle severe rollout idle periods and long-horizon sparse-reward signal imbalances in desktop GUI agent training, this work presents DART, a fully decoupled agentic RL framework separating environment clusters, rollout services, data management, and training. Combined with multi-level adaptive data curation (dynamic task allocation, cold-start experience replay, high-entropy step filtering, and truncated importance sampling), DART-GUI-7B achieves a 43.28% task success rate on OSWorld, setting a new SOTA for open-source 7B-scale agents.

Generative Models Enhanced by Sequence Labelling and Aspect-Code Switching Improve Cross-lingual Aspect-Based Sentiment Analysis

This paper introduces a generative cross-lingual aspect-based sentiment analysis framework that combines an encoder auxiliary sequence labelling task (SeqLab) with alignment-free aspect-code switching (ACS) data augmentation, leveraging token-level word boundary supervision and dual-path confidence verification to achieve new state-of-the-art results across eleven languages and three domains.

GENIE: A Fine-Grained Measure for Novelty

Addressing the inability of black-box holistic metrics to explain where novelty originates and their confusion between quality and diversity, GENIE leverages question-based alternative semantics to automatically decompose tasks into orthogonal feature axes, computing population-grounded Q&A dissimilarity to deliver explainable, fine-grained, and paraphrase-robust novelty evaluation.

GNN-CB: A Graph Neural Network Competition Benchmark for Human and LLM Evaluation

GNN-CB introduces the first competition-based benchmark for evaluating both humans and LLMs on end-to-end Graph Neural Network coding tasks, revealing through 18 diverse competitions that current LLMs lag behind expert humans on over 88% of task setups primarily due to failures in graph topology construction and relational feature modeling.

GRACE: Step-Level Benchmark for Faithful Reasoning over Context

This paper proposes GRACE, the first step-level benchmark for context-grounded textual reasoning, featuring a data-driven two-track eight-category error taxonomy and a rigorous human-adjudicated evaluation set. The benchmark reveals a critical disconnect where 49.5% of traces with unfaithful intermediate steps still produce correct final answers, and shows that incorporating step-level context faithfulness into reinforcement learning as process rewards substantially boosts both reasoning reliability and task accuracy.

Gradients Know What Outcomes Don’t: Unlocking Reinforcement Learning for LLM Reasoning with Gradient-Aligned Rewards

Addressing the flat-reward pathology in reinforcement learning from verifiable rewards (RLVR) where correct reasoning paths are indistinguishable, this paper proposes Gradient-Aligned Reward (GAR), an online process reward mechanism that measures cosine similarity between truncated output-layer gradient-activation vectors of policy rollouts and expert chain-of-thought anchors, delivering consistent gains on competition-level math benchmarks with under 9% wall-clock overhead.

Grammar Control for Language Learning with Sequential Monte Carlo and Small Language Models

This paper introduces an inference-time framework utilizing Sequential Monte Carlo (SMC) and decoupled symbolic potential functions (combining fast regular expression prefix heuristics with final dependency parsing) to steer small language models toward target pedagogical grammar constructions, boosting a 0.5B model's grammar coverage to 98% and outperforming models 140 times larger without any fine-tuning.

Evading Fake News Detection: Self-evolving Social Manipulation via LLM-driven Agent Orchestration

This paper introduces EvoSM, a self-evolving social manipulation attack framework against propagation-based fake news detectors, which coordinates heterogeneous semantic and topological actions through an LLM-driven orchestrator and distills reusable cross-target priors via reflective experience memory to achieve state-of-the-art black-box evasion and transferability.

Grounded Token Initialization for New Vocabulary in LMs for Generative Recommendation

Addressing the severe token-embedding collapse caused by standard mean-of-vocabulary initialization when extending language models with new domain tokens, this paper proposes Grounded Token Initialization (GTI)—a lightweight stage that freezes the LM backbone and trains only the new token embeddings via paired bidirectional linguistic supervision prior to supervised fine-tuning, preserving a rich, high-rank geometric manifold and yielding substantial retrieval gains across both industry-scale and public benchmarks.

GSEM: Graph-based Self-Evolving Memory for Experience-Augmented Clinical Reasoning

To resolve boundary failures and collaboration conflicts in experience reuse for clinical agents, this paper proposes GSEM, a graph-based self-evolving memory framework that models experiences as indications and contraindications organized in a dual-layer graph, achieving state-of-the-art clinical reasoning via applicability-aware traversal and feedback-driven recalibration.

GSR-LoRA: Group-Sparse Routing-aware Low-Rank Adapter Compression for Multi-Task Mixture-of-Experts Fine-Tuning

Addressing the rigid parameter allocation in multi-task PEFT where conventional fixed-rank LoRA fails to account for heterogeneous rank demands across tasks, layers, and modules, GSR-LoRA models low-rank rank components as selectable expert groups, suppresses redundant components via Group Lasso regularization, and conducts router-aware non-prefix compression to slash average rank from 16 to 5.5 (65.6% parameter reduction) on GLUE with a negligible 0.1-point drop.

GTS: Inference-Time Scaling of Latent Reasoning with a Learnable Gaussian Thought Sampler

This paper introduces the Gaussian Thought Sampler (GTS), which reformulates heuristic noise perturbations in continuous latent reasoning into a context-conditioned, learnable Gaussian policy optimized via GRPO while freezing the backbone, substantially improving candidate-set exploration quality and pass@N scaling across diverse reasoning benchmarks.

HachiWare: One-Shot Selective SVD in a Dual-Branch Framework for KV Cache Compression

HachiWare proposes a dual-branch KV cache compression framework that selectively allocates reconstructable low-rank SVD storage to offline-identified semantic-critical heads while applying lightweight eviction to the remaining heads, incurring only a 3% accuracy drop at 16× compression on RULER and delivering up to 16.50× decoding throughput speedup.

HalloMTBench: A Large-Scale Human-Verified Benchmark for Diagnosing Hallucination Triggers in LLM Translation

Addressing the failure of traditional NMT benchmarks to elicit hallucinations in modern multilingual LLMs, this work introduces a fine-grained taxonomy distinguishing Instruction Detachment from Source Detachment and establishes HalloMTBench—a human-verified benchmark of 6,908 authentic hallucination cases across 11 language pairs mined from 16M+ translations—revealing clear U-shaped length dependencies, cross-lingual proximity confusion, and model-dependent reasoning behaviors.

HARP: Hadamard-Preconditioned Adaptive Rotation Processor for Extreme LLM Quantization

HARP introduces a learnable, structured two-sided orthogonal processor based on sparse block-orthogonal butterfly stages that initializes exactly to the randomized Hadamard transform (RHT) and adapts the rotation basis to layer-specific curvature and downstream quantizer codebooks, delivering substantial perplexity and zero-shot improvements in 2-4-bit LLM regimes while retaining 90% of RHT inference throughput.

He Said, She Said, They Simulated: Simulating Situated Gossip with LLM Agents

Introduces the first LLM agent-based simulation sandbox framework for situated workplace gossip, combining socially grounded agent profiling, modular experimental strategies, and a 6-phase chain-of-reasoning to both replicate classic human psychological benchmarks and generate novel organizational hypotheses under peer-review competition.

HEMP: Hierarchical Entity-aligned Memory Prompting for CT Report Generation

HEMP presents a hierarchical entity-aligned memory prompting framework for 3D CT radiology report generation, resolving spatial ambiguity via two-stage graph-constrained attention and optimal transport refinement, enriching clinical patterns through a conditional dual-Hopfield memory with implicit and explicit prompt injection, and regularizing diagnostic coherence across three complementary severity granularities.

Meerkat2: Bridging Global Reasoning and Local Knowledge for Efficient Medical AI

Meerkat2 decouples clinical reasoning transfer from high-fidelity local statute learning across an 837K-example mixture and introduces Verbosity-Decoupled Distillation (VDD) to cancel internal gradient updates toward excessive output length, matching frontier models on U.S. and Korean clinical licensing benchmarks while slashing token generation costs by over 50%.

HiTeC: Hierarchical Contrastive Learning on Text-Attributed Hypergraph with Semantic-Aware Augmentation

Addressing the issues of semantic-topological disconnection, perturbation noise from random augmentations, and poor long-range structural modeling in text-attributed hypergraphs (TAHGs), HiTeC introduces a decoupled two-stage contrastive learning framework. By combining structure-aware text encoder pretraining, semantic-aware dual augmentations, and s-walk-induced multi-scale hierarchical contrastive objectives, HiTeC significantly surpasses strong baselines across six benchmark datasets.

How AI Models Manage Epistemic Authority: A Taxonomy and Comparative Analysis of Responses to User Disagreement

Grounding its methodology in Conversation Analysis, this paper introduces a 6-type challenge taxonomy and a 4-layer response annotation framework, revealing that LLMs systematically dissociate social politeness from epistemic concession by validating users in 85% of responses while maintaining their original claims in 65%.

(How) Do MLLMs Report Bistable Images Like Humans?

Evaluating the canonical duck–rabbit and 50 synthetic Visual Anagrams, this paper demonstrates that Multimodal Large Language Models (MLLMs) report ambiguous images with human-consistent modulability and report exclusivity, revealing through circuit interventions that bottom-up and top-down modulations follow distinct computational pathways and that exclusivity is causally rooted in single-object construal.

How Do People Challenge Racial Stereotypes Online? Counter-Story Detection Across Reddit Communities

This paper introduces the first computational framework for detecting and characterizing counter-storytelling against racial stereotypes online, establishing a three-dimensional taxonomy grounded in narratology and Critical Race Theory, and uncovering 1,312 authentic counter-stories and seven latent rhetorical roles across 615 Reddit communities.

How Identity and Opinion Shape Political Sycophancy in LLMs

This paper introduces an open-ended factorial evaluation framework across 450 double-edged political dilemmas to disentangle political sycophancy triggered by explicit user opinions from that induced by demographic identity cues, revealing that these two vulnerabilities dissociate across 13 instruction-tuned LLMs and combine sub-additively.

How Well Do Vision-Language Models Understand Movies?

This work introduces MF2 (Movie Facts and Fibs), a fully manually curated benchmark consisting of 53 open-licensed full-length movies and 868 human-authored contrastive fact-fib claim pairs; evaluation across 22 modern VLMs reveals that top models (Gemini 2.5 Pro at 63.8% pairwise accuracy) lag substantially behind humans (84.1%), showing that while subtitles serve as the primary narrative backbone, models struggle to identify and integrate claim-relevant visual cues dispersed across extended multimodal contexts.

Hybrid Open-Ended Tri-Evolution Makes Better Deep Researcher

Addressing the limitations of static data and lack of verifiable ground truths in long-form deep research, this paper proposes Hybrid Open-Ended Tri-Evolution (HOTE), a tri-evolutionary framework co-evolving a proposer, solver, and judge via dual-mode reinforcement learning, enabling an 8B model to surpass strong 8-32B baselines with significantly less training overhead.

HybridThink: Learned Interleaving of Continuous and Discrete Reasoning in Large Language Models

Addressing manifold deviation and anchoring failures in continuous reasoning, HybridThink introduces an RL-trained lightweight gating policy to dynamically interleave continuous concept exploration and discrete symbolic anchoring, boosting accuracy on math and coding tasks while drastically cutting token consumption.

HybridThinker: Efficient Chain-of-Thought Reasoning via Compressed Memory and Transient Thought Steps

Addressing the premature loss of fine-grained local information in chain-of-thought compression, HybridThinker temporarily retains recent thought steps within a sliding window alongside persistent compressed memory tokens and introduces a hybrid attention training scheme to resolve shortcut learning, matching uncompressed accuracy on Qwen2.5-7B while reducing peak KV cache by 61.7% and inference latency by 20.3%.

HyQuant: Hybrid-Precision Quantization for LLM Attention

To tackle severe performance degradation in aggressive low-bit attention during long-context Chain-of-Thought (CoT) reasoning, HyQuant introduces a vertical-line-aware hybrid-precision quantization framework that preserves a tiny subset (<5%) of persistent vertical-line anchor tokens and a recent sliding window in full precision while compressing the vast majority of historical keys/values to 4-bit with fused FlashAttention-style operators, achieving near-lossless accuracy alongside 1.32×–3.58× decode-kernel speedups.

"I Strongly Suspect This Website Is a Scam": Benchmarking PII Leakage and Detection without Defense in Autonomous Web Agents

Introducing SCAMMER4U, a pre-registered benchmark of 91 adversarial environments with 60 single-axis paired siblings and 10 benign twins, this paper demonstrates critical-tier PII leakage reaching 54%–93% across frontier web agents and establishes an empirical 30.2 pp "detection–action gap" where agents identifying a site as fraudulent still submit sensitive credentials in 35.9% of sessions.

Identifying and Mitigating Bottlenecks in Role-Playing Agents: A Systematic Study of Disentangling Character Profile Axes

This work systematically disentangles character profiles along three axes—Familiarity (Known vs. Unknown), Structure (Structured vs. Unstructured), and Disposition (Moral vs. Immoral)—revealing that post-SFT alignment suppression on value-laden fields is the dominant bottleneck in role-playing agents, and introduces Field-Aware Contrastive Decoding (FACD) to significantly close this gap without parameter updates or sacrificing moral persona quality.

Identifying and Transferring Reasoning-Critical Neurons: Improving LLM Inference Reliability via Activation Steering

AdaRAS reveals that a tiny subset of MLP neurons in LLMs strongly correlates with multi-step reasoning validity; by contrasting trajectory-level activation differences with polarity-aware sign-flip filtering and gating interventions via a lightweight failure predictor, it rectifies erroneous reasoning traces at test time with negligible computational overhead.

Imagine-then-Plan: Agent Learning from Adaptive Lookahead with World Models

Addressing the "shallow grounding" and trial-and-error breakdowns caused by purely reactive perception in LLM agents, this paper introduces Imagine-then-Plan (ITP) within a newly conceptualized POIMDP formulation, adaptively balancing task progress against rollout costs to guide both training-free reflection (ITPI) and reinforced online actor-critic optimization (ITPR), achieving new state-of-the-art results across ALFWorld, ScienceWorld, and tool-use benchmarks.

InComeS: Integrating Compression and Selection Mechanisms into LLMs for Efficient Model Editing

InComeS compresses each individual edit context into the KV cache of a single gist token and integrates these representations through a specialized cross-attention module equipped with a learnable Zero-Gist token, breaking context window constraints and dramatically accelerating multi-hop reasoning and natural language model editing over large batches.

Incorporating Explicit and Implicit Chain-of-Thought as Hybrid Reasoning Paradigm with Guided Latent Space Modeling

Pre² introduces a phase-aware hybrid reasoning framework that decouples thinking and answering via a guided latent space, predicting continuous pre-thinking and pre-answering vectors to guide autoregressive decoding while retaining full explicit step-by-step supervision, achieving substantial reasoning improvements across domains with minimal inference overhead.

InFi-Check: Interpretable and Fine-Grained Fact-Checking of LLMs on Grounding Documents

Addressing the lack of interpretability and categorical oversimplification in document-grounded LLM fact-checking, InFi-Check introduces a controlled 5-stage synthesis pipeline yielding 15.6k training instances and the benchmark InFi-Check-FG, training 8B InFi-Checkers that simultaneously output grounding evidence, 6 fine-grained error types, chain-of-thought justifications, and corrections, outperforming GPT-4o and o3 by over 27% balanced accuracy.

InfoDensity: Rewarding Information-Dense Traces for Efficient Reasoning

InfoDensity addresses reasoning verbosity and length-penalty reward hacking by identifying that high-quality reasoning traces exhibit low uncertainty convergence and fast descent along their entropy trajectories, and introduces a single suffix-max envelope reward coupled with group-relative length scaling in reinforcement learning to simultaneously compress response lengths and enhance accuracy.

InfoSpec: Information-Guided Anchor Selection for Retrieval-Based Speculative Decoding

Addressing the severe retrieval ambiguity caused by fixed future-token anchors in retrieval-based speculative decoding, InfoSpec formulates anchor selection as an inference-time optimization problem under a fixed verification budget, estimating candidate utility via model confidence, ambiguity reduction, locality, and continuation consistency to construct high-value draft trees and achieve superior wall-clock speedup across multiple LLM backbones.

INSPIRE: An Internalize-Then-Improve Approach for Example-Driven Mathematical Reasoning

Addressing the core limitation where large language models rely excessively on abstract theorem invocation and outcome correctness rather than probing theorem preconditions, INSPIRE combines Reference-Guided Student Internalization to avoid distribution shift with a stage-wise rubric preference training strategy (M-DPO followed by C-DPO), substantially boosting counterexample-driven mathematical reasoning.

iPOE: Interpretable Prompt Optimization via Explanations

This paper proposes iPOE, an interpretable prompt optimization approach that derives modular, rule-like annotation guidelines from decision explanations and optimizes them via discrete editing operators, delivering up to a 39 percentage point gain over baseline prompts while maintaining high human-LLM alignment in decision rationale.

Is Convergence Inevitable? Tracing Output Homogeneity Back to Base Models

By systematically tracking full alignment stages, conducting sample-level controlled SFT intervention experiments, and directly probing unaligned base models via prompting, this work reveals that language model output homogeneity (semantic convergence) is not introduced by post-training preference alignment, but is an inherent latent property already encoded during pretraining, where SFT merely serves as an amplifying catalyst.

Is the Future Predictable? Modeling Future Predictability for Reliable Temporal Reasoning

Addressing the critical issue where large language models produce overconfident hallucinations on future temporal queries beyond evidential support, this paper formalizes the evidence-bounded ChronoRegime task and proposes Confidence-Aware Temporal Reasoning (CTR), a framework that estimates future predictability from historical dynamics and regulates answer commitment, boosting overall accuracy by 6.7 percentage points and cutting hallucination rates by 18.4 percentage points across multiple LLM backbones.

Knowledge Editing for Masked Diffusion Language Models

This work provides the first systematic study of locate-then-edit knowledge editing in Masked Diffusion Language Models (MDMs), demonstrating that factual recall causally localizes to the early-to-mid MLP of the last subject token across both ARMs and MDMs, identifying a severe multi-token editing breakdown during iterative denoising, and proposing an intermediate-state optimization correction that substantially restores multi-token editing efficacy.

KnowledgeGain: Evaluating and Optimizing Science News Generation for Reader Learning

Addressing the failure of conventional text generation metrics to capture actual reader learning, this paper proposes KnowledgeGain, a pre/post-testing comprehension metric, and introduces a calibrated reader simulator LLMSim to filter candidate articles, significantly improving downstream human learning outcomes.

Label Over Logic? How Source Cues Bias Human Fallacy Judgments More Than LLMs

Through a controlled double-blind experiment comparing 505 human evaluators against GPT-5.2, Gemini 2.5 Flash, and Claude Sonnet 4.5, this paper reveals that humans exhibit severe cognitive vulnerability to logical fallacies when labeled as authored by humans or human-AI teams (fallacy penalty dropping from 1.47 to 0.42), whereas modern LLMs remain remarkably source-agnostic and maintain rigorous critical evaluation.

LADA: Enabling Adaptation of Black-Box LLMs to Dynamic Domain Changes at Test Time

Addressing the failure of black-box LLMs under sudden and dynamic domain distribution shifts, LADA meta-trains a lightweight adapter across Cartesian products of multi-source datasets and composite reasoning error types, quickly adapts it at test time using merely three positive–negative sample pairs from the target domain, and adaptively guides black-box reasoning generation via step-level reliability scoring and local resampling at low latency.

LAK-LoRA: Layer-Wise Adaptive Expert Allocation and Principal Component–Based Feature Separation for Medical Multi-Task Learning

LAK-LoRA addresses conflicting objectives and capacity mismatches in medical multi-task LLM fine-tuning by introducing depth-adaptive expert capacity scaling alongside an SVD-based principal component feature separator and orthogonality constraints, outperforming existing multi-task PEFT methods without inflating parameter budgets.

Language Model Maps for Prompt-Response Distributions via Log-Likelihood Vectors

This paper constructs language model maps for conditional prompt-response distributions via conditional log-likelihood and PMI vectors, proves that squared Euclidean distances in this space approximate conditional KL divergence, discovers that prompt modifications induce approximately additive shift vectors (Prompt Shift), and accurately predicts downstream task scores and composite prompt effects without autoregressive generation.

Large Reasoning Models Learn Better Alignment from Flawed Thinking

Addressing the acute vulnerability where large reasoning models (LRMs) are easily hijacked by flawed chain-of-thought (CoT) premises, this paper introduces RECAP, an RL post-training framework that mixes counter-aligned reasoning prefills into rollouts, training models to self-reflect and override deceptive trajectories to achieve robust alignment without inference overhead or math capability degradation.

Last Layer Logits to Logic: Empowering LLMs with Logic-Consistent Structured Knowledge Reasoning

To eliminate logic drift in structured knowledge reasoning, this paper introduces the Logits-to-Logic framework that compiles graph topologies into an NFA and operates directly on last-layer logits via contrastive strengthening and topological filtering, achieving state-of-the-art KGQA performance with a single LLM forward pass.

One Instruction Is a Benchmark: End-to-End Instruction-Following Evaluation with FlexBench

Addressing the disconnect between aggregate public leaderboards and concrete deployment needs, this paper presents FlexBench/FlexEval, an end-to-end framework that instantiates an instruction-specialized test suite from a single target instruction by deriving verifiable atomic dimensions, generating profile-conditioned multi-turn dialogues, and disentangling workflow progress (Coverage) from conditional correctness (Achievement) via precondition-aware tri-valued judging.

Latent Thoughts, Certain Gains: Bayesian Optimization for LLM Reasoning

BLTO is a training-free test-time reasoning framework for frozen LLMs that treats continuous latent thoughts as stochastic variables under a variational Bayesian inference formulation, optimizing latent vectors via an entropy-based intrinsic reward, reparameterized low-variance updates, and an uncertainty-guided adaptive controller.

Latent Thoughts, Certain Gains: Bayesian Optimization for LLM Reasoning

BLTO introduces a training-free framework that casts continuous latent thought generation in frozen LLMs as variational Bayesian inference over stochastic cognitive states, employing self-supervised output predictive entropy as an intrinsic reward and reparameterized gradient updates to achieve consistent gains across complex reasoning benchmarks.

LatticeMind: A Conflict-Aware Memory Primitive for Multi-Agent Systems

Addressing persistent failures caused by recurring contradictions and stale claims in multi-agent LLM systems, LatticeMind formulates contradiction handling as a write-time structured memory update primitive that couples cheap symbolic consistency checks with selective LLM reconciliation, lifting label-blind ConflictBank accuracy from a 0.525 baseline to 0.965.

Learning Concepts, Not Tokens: Self-Supervised Semantic Alignment for Language Models

Addressing the surface-level token competition in standard Next-Token Prediction where valid semantic continuations are penalized, this paper introduces a self-supervised concept-level training objective that optimizes the cumulative likelihood over contextual synonym sets, boosting human semantic alignment and reasoning while preserving general predictive performance.

Learning User Simulators with Turing Rewards

This paper introduces Turing-RL, a reinforcement learning approach for training personalized user simulators that replaces conventional ground-truth response matching with a discriminative Turing reward (scoring how indistinguishable a generated response is from real human behavior conditioned on user history). Optimized via GRPO with chain-of-thought reasoning, Turing-RL consistently outperforms baseline methods on both LLM-as-a-judge and human evaluations across multi-turn chat and Reddit forum discussions.

Learning What to Retain: Gated-Memory Routing for Efficient Collaboration in Multi-Agent LLM Systems

Gated-Memory Routing re-architects multi-agent coordination as a sequential decision process conditioned on a learned, gated execution memory, jointly optimizing write and retrieval gates, role allocation, heterogeneous backbone routing, and budget-aware early halting to exceed the strongest multi-agent baseline by 2.44% average accuracy while cutting HumanEval inference cost by 31.9%.

Learning What to Share and What to Personalize: Hierarchical Strategy Co-Evolution for Agent Memory

Addressing the rigidity and personalization failure of static, one-size-fits-all memory rules in long-context conversational agents, this paper introduces HiPS, a hierarchical framework that decouples memory management into a globally shared foundation and a user-specific adaptive tier, co-evolving declarative natural-language rules and policy models through on-policy evidence and cross-level rule flow.

Length and Meaning for Language Model Maps

Addressing the computational bottleneck where log-likelihood evaluation scales linearly with text volume when constructing macroscopic language model maps across domains, this work reveals that text length and semantic embeddings jointly determine inter-model likelihood variance. The authors propose Predictive Importance Sampling (PredIS), which achieves the KL divergence estimation accuracy of uniform sampling with only ~45% of the text volume while zero-shot generalizing across 24 unseen domains.

Less is MoE: Trimming Experts in Domain-Specialist Language Models

Addressing the catastrophic capability collapse of Mixture-of-Experts (MoE) models under expert-level pruning on reasoning and general benchmarks, this paper introduces Fisher-MoE, a fine-grained compression framework that retains all experts while physically trimming FFN intermediate dimensions via empirical Fisher information, preserving core capabilities at a 50% pruning ratio while slashing weight VRAM by ~45% and boosting inference throughput by 21%.

Less Uniform Discrete Diffusion is More Powerful and Scalable

Addressing the over-uniform training objective and condition-target confusion in uniform discrete diffusion language models (UDLMs), this work introduces LUDI, featuring a less-uniform (LU) loss and per-token corruption hints that scale a 7B diffusion model to complex reasoning tasks with a 3 token-per-step speedup.

Litmus (Re)Agent: A Benchmark and Agentic System for Predictive Evaluation of Multilingual Models

This paper introduces a controlled benchmark of 1,500 questions and LITMUS (RE)AGENT, a DAG-orchestrated agentic system that decomposes performance queries into hypotheses, retrieves citation-grounded evidence, and synthesizes predictions through linguistic feature modeling and code-based regression to infer missing multilingual evaluation scores under restricted literature evidence.

LLMs Learn Better In-Context from Rules than from Examples

By systematically developing five novel, complex reasoning tasks spanning games, arithmetic, and synthetic linguistic inference, this paper demonstrates that LLMs learn far more reliably from natural language rules than from input-output examples, while adding demonstrations to rules or scaling demonstration volume fails to produce consistent performance gains.

LoCoDec: Robust Decoding for LLM-based Sequence-Structured Prediction

Addressing structural hallucinations and catastrophic collapse in LLM-based structured prediction under long-context perturbations, LoCoDec reformulates prediction as constrained token paraphrasing, combining dynamic neighborhood context with KV-cache truncation and monotonic stack-based decoding to secure top-tier perturbation robustness (R-F1) across 16 benchmarks.

LongAct: Harnessing Intrinsic Activation Patterns for Long-Context Reinforcement Learning in LLMs

Reveals that Query and Key attention representations exhibit persistent high-magnitude outlier dimensions across long sequences, and proposes LongAct, a dynamic saliency-guided sparse reinforcement learning framework that selectively updates only the weight rows tied to these critical activation channels (top 30%), improving Qwen3-8B's LongBench v2 score by ~8% (36.73% vs. 27.04% SFT baseline and 32.80% full-parameter DAPO) with zero inference latency or architectural modifications.

LongNovel: A Multi-Scale Benchmark for Hallucination Detection in Long-Context Novel Summarization

To tackle the elusive challenge of detecting factual hallucinations in massive narrative contexts, this paper introduces LongNovel, a bilingual benchmark spanning 16k to 100k tokens across eight fine-grained hallucination categories constructed via multi-model arbitration and entity-referenced generation, uncovering severe performance collapse in long-range temporal and causal reasoning even among frontier LLMs.

Lost Between Scripts: Recovering Arabic Words from Noisy Learner Orthographies

Addressing the acute challenge where second-language learners produce distorted spellings across Arabic script and Romanized Arabizi, this paper formulates gloss-grounded intended-word recovery, demonstrating that combining noisy surface forms with English semantic glosses closes cross-script performance disparities and allows leading LLMs to approach human expert retrieval accuracy.

Lost in Delusion: Examining LLM Safety Under User Delusions and Distress

Using matched 16-turn multi-turn dialogue simulations based on 30 clinically grounded personas across six LLMs (4,200 conversations total), this paper uncovers a severe "recognition-intervention gap" where delusional framing suppresses safety interventions by up to 4.5× despite models accurately detecting distress, identifies an accumulated premise acceptance mechanism termed "narrative debt", and demonstrates that only explicit delusion assessment coupled with non-collusive response guidance reliably mitigates belief confirmation.

LTS-VoiceAgent: A Listen-Think-Speak Framework for Efficient Streaming Voice Interaction via Semantic Triggering and Incremental Reasoning

Addressing the bottleneck of system-induced post-turn-end silence in cascaded voice agents, LTS-VoiceAgent explicitly separates when to think from how to reason incrementally through a Dynamic Semantic Trigger and a Thinker-Speaker dual-role orchestrator, achieving sub-second streaming responsiveness with minimal invalid computation and zero audible rollbacks.

LUMINA: LongitUdinal MultImodal reasoNing Agent for Medical Imaging

Addressing the degradation and opacity caused by collapsing irregular longitudinal clinical histories into monolithic prompts, LUMINA formulates longitudinal reasoning as a structured single-agent state machine featuring intra-modality anchor comparisons, multi-timescale trajectory memory with Sequential Monte Carlo (SMC) compression, and bounded contradiction audits, establishing state-of-the-art results on Alzheimer's disease classification and 36-month progression forecasting across 9 MLLM backbones.

MA-Mesh: Resilient Multi-Agent Collaboration via Role-based Self-Healing

Drawing inspiration from cloud-native service meshes, this paper proposes MA-Mesh, a framework that cleanly decouples task-level collaboration logic from system-level resilience policies. By overlaying sidecar-style resilient proxies equipped with cross-agent reflection and role-based failover onto multi-agent topologies, MA-Mesh establishes an infrastructure-level self-healing layer that outperforms the competitive baseline C+I by an average of 32.51 percentage points in accuracy under diverse system chaos scenarios.

Manifold Embedding: A Point-to-Hyperplane Approach

This paper introduces the Point-to-Hyperplane Distance Map (PHDM), an auditable, parameter-free Euclidean-to-manifold mapping that interprets each Euclidean feature scalar as an exact signed geodesic distance to a canonical orthogonal hyperplane, providing closed-form solutions across four constant-curvature geometries that consistently outperform exponential map baselines in LLM arithmetic reasoning, image classification, and graph learning.

MathDebugger: A Benchmark for Type-Aware Error Detection in Synthetic Math Data

This paper introduces MATHDEBUGGER, the first text-only benchmark evaluating type-aware error detection across both mathematical question statements and solution traces (4,000 questions and 2,000 answers), revealing that reasoning-tuned models suffer from a substantial solving-vs-verifying gap and showing that fine-grained error-type supervision significantly boosts automated mathematical data correction.

Measuring Affinity between Attention-Head Weight Subspaces via the Projection Kernel

Addressing the direction-dependent scaling and norm distortion problems of the classical Composition Score, this work introduces the principal-angle-based Projection Kernel (PK) to measure the geometric overlap between attention-head weight subspaces directly, substantially improving head circuit detection and functional clustering across GPT2, BLOOM, and Qwen while identifying critical hub identity heads.

Mechanism Shift During Post-training from Autoregressive to Masked Diffusion Language Models

By employing Edge Attribution Patching with Integrated Gradients (EAP-IG) and logit-lens diagnostic probes across paired 7B architectures, this paper demonstrates that post-training autoregressive models into masked diffusion models induces a task-dependent mechanism shift: prefix-dominant tasks largely preserve inherited causal circuits, while globally constrained tasks systematically reorganize computation toward earlier transformer layers.

MELD: Mel-Spectrogram-Based Speech Language Modeling with Discrete Latent Variables

MELD introduces a variational autoregressive framework that bridges discrete latent sampling and continuous mel-spectrogram generation, enabling end-to-end joint optimization of the mel encoder and speech language model while seamlessly unifying zero-shot TTS and STT without relying on decoupled neural codecs or heuristic stop predictors.

Messier: A High-Resolution Corpus for Cross-Benchmark Agent Evaluation

Addressing the severe fragmentation of agent evaluation and the obfuscation of granular behaviors caused by aggregate scores, this paper introduces MESSIER, a unified corpus of 957,611 high-resolution records across 30 benchmarks, 745 agents, 11,891 tasks, and 74,263 verifiers, demonstrating that open granular verifier data can reveal starkly uneven frontier progress, expose distortions induced by all-pass scoring, and accurately reproduce latent capability indices without costly reruns.

MetaBench: Benchmarking Two-Stage Reranking and Top-N Screening for Meta-Analyses

Addressing the labor-intensive nature of medical systematic review citation screening and the lack of reproducible fine-grained eligibility benchmarks, this paper introduces MetaBench—a large-scale benchmark of 7,801 meta-analyses formulating study selection into candidate reranking and budgeted Top-N screening, demonstrating substantial screening effort reduction and precision gains via coverage tiering and PICOS-targeted synthetic hard negatives.

Metonymic Circuits for Abstract Concept Grounding in Vision Transformers

This paper investigates how Vision Transformers ground abstract concepts lacking direct physical referents by uncovering internal metonymic circuits via sparse Top-k Transcoders on CLIP and DINO, demonstrating that concrete intermediate anchors causally bridge perceptual primitives to abstract targets independently of training caption co-occurrence statistics.

MICE: Minimal Interaction Cross-Encoders for efficient Re-ranking

By systematically identifying and eliminating superfluous token interactions via causal attention masking, MICE reconstructs the cross-encoder into an asymmetric architecture featuring independent early contextualization, unidirectional cross-attention over frozen document representations, lexical exact-match heads, and top-layer pruning—achieving up to 2.5× FLOPs reduction while outperforming standard cross-encoders in out-of-domain retrieval.

MIDR: Enrichment-Augmented Indexing for Multimodal Document Retrieval

Addressing the excessive serving-time storage and computational overhead of visual multi-vector retrievers like ColPali and ColQwen2.5, MIDR introduces a training-free enrichment-augmented indexing framework that moves multimodal reasoning to offline ingestion via an extract–verify–refine loop, materializing visual layout evidence into structured text fields to enable fast, lightweight BM25F and dense text-centric serving.

MIND: An Evidence-State Decision Interface for Criteria-Grounded Psychiatric Inquiry and Diagnosis

Addressing unsupported assertions and inquiry goal drift caused by unstructured text in multi-turn psychiatric consultation, MIND constructs a typed evidence state \(E_t\) encompassing observed evidence, unresolved criterion checks, active differentials, and reliability-gated clinical supports, sharing this interface across policy decisions, 3D clinical process rewards, information-gain scoring, and trajectory rectification to achieve up to 8.8-point diagnostic gains under EMR-grounded simulation.

Mindscape-Aware Retrieval Augmented Generation for Improved Long Context Understanding

Inspired by human cognitive mindscapes, MiA-RAG introduces a unified conditioning paradigm where a hierarchical document summary serves as an external semantic scaffold for both retriever (MiA-Emb) and generator (MiA-Gen), enabling a 14B model to consistently surpass 72B RAG baselines in long-narrative comprehension and global sense-making.

Minimal Grammars as Supervision for DSL Generation

This paper introduces minimal grammars—the exact, non-redundant subset of DSL productions (typically 2%–20% of the full reference grammar) required to derive a target program—as an explicit structural supervision signal for fine-tuning open-weight models, coupled with an inference-time RAG+CoT grammar predictor that delivers massive accuracy leaps across semantic parsing and hardware description languages.

MIRAGE: Reliability Assessment via Grounded Evidence for Table Reasoning

Addressing the critical failure mode where large language models produce stable but structurally ungrounded tabular answers, this paper introduces MIRAGE, a structure-grounded reliability framework that estimates residual structural uncertainty across sampled execution traces and employs quantile-based calibration to achieve low-risk, interpretable selective table reasoning.

Mitigating Hallucination in Vision-Language Models through Barrier-Regulated Adaptive Closed-form Steering

Addressing visual grounding decay and over-correction in vision-language models, BRACS introduces a training-free barrier-regulated adaptive closed-form steering framework that monitors pre-softmax image-attention energy and derives minimum-norm hidden-state updates analytically without backpropagation.

Mitigating Memory Interference in Lifelong Model Editing through Multi-Head Hopfield Retrieval

To tackle memory interference where crowded codebook representations cause retrieval ambiguity under dense lifelong editing, MH-HoReN factorizes continuous Hopfield attractor dynamics across complementary feature subspaces, combining head-wise scores for robust routing alongside a semantic clustering scheme for scalable batched editing.

Mitigating Overthinking via Interleaved System-1/2 Reasoning in Large Language Models

To alleviate overthinking and computational waste in large reasoning models caused by uniform long Chain-of-Thought, this paper proposes InterSR, a framework that dynamically alternates between fast System 1 and slow System 2 thinking at step-level boundaries using a lightweight semantic-entropy predictor, reducing token consumption by up to 51% while maintaining or slightly improving problem-solving accuracy.

Modular Representation Steering for Multi-Attribute LLM Alignment

To overcome severe optimization interference in shared parameter tuning and directional cancellation when composing activation steering vectors, POLAR trains attribute-specific low-rank representation intervention modules on a frozen LLM and composes them at inference time via log-z demand calibration and greedy cosine compatibility filtering.

MoPLEx: Estimating Plackett-Luce Mixture Models for Multi-Objective Alignment

MoPLEx uncovers the fundamental unidentifiability barrier of mixture ranking models when ranking lengths are short (\(m \le 2k-1\)), breaks this barrier by appending unannotated base model generations as bottom-ranked responses, and cuts training complexity from \(O(km)\) to \(O(ka)\) via first-order input embedding gradient approximations, achieving a 43.7% clustering and 15.2% ranking accuracy gain across UltraFeedback and Persona.

MoralSteer: Safety-Preserving Fidelity Gains in Morally Complex Role-Playing via Theory-Guided On-Policy Self-Distillation

Addressing the systemic prosocial collapse and trait substitution exhibited by safety-aligned LLMs in morally complex villain role-playing, MoralSteer introduces Moral Foundations Theory (MFT) reasoning chains as training-time privileged information for a frozen teacher and employs a unified scalar moral complexity signal \(M(c)\) to govern bidirectional KL interpolation, trajectory mixing, and curriculum stages—improving L3+L4 role-play fidelity by +5.8% to +9.9% across four base models while keeping safety violation variations strictly within \(\pm 1\text{pp}\) across five major safety benchmarks.

More Test-Time Compute Can Hurt: Overestimation Bias in LLM Beam Search

This paper models test-time beam search for LLM reasoning as a noisy Top-\(b\) candidate retention primitive, proving from order statistics and extreme value theory that expanding candidate pools systematically induces overestimation bias; consequently, beam widening must match the scorer's signal-to-noise ratio, whereas blindly scaling compute under weak scoring signals degrades output quality.

Multi-Agent Evolve: LLM Self-Improve through Co-evolution

Multi-Agent Evolve (MAE) breaks the fundamental dependence of LLM reinforcement learning on human-labeled data and external deterministic verifiers (such as code interpreters or math engines) by instantiating an asymmetric Proposer-Solver-Judge triad from a single parameter-shared backbone, delivering a 4.86% average accuracy gain across 22 benchmarks on Qwen2.5-3B-Instruct.

MultiHashFormer: Hash-based Generative Language Models

MultiHashFormer encodes discrete tokens into collision-free multi-ID hash signatures and reconstructs them sequentially via a cascaded predictor, breaking the linear vocabulary parameter scaling, mitigating the softmax rank bottleneck, and enabling zero-parameter multilingual vocabulary expansion.

Natural Language to First-Order Logic LLM-based Autoformalization

Addressing the lack of unified task formulations and systematic surveys in natural language to First-Order Logic (FOL) autoformalization, this paper formalizes the problem as an ontology-relative tuple decoupling Ontology Extraction (OE) from Logical Translation (LT), while providing a comprehensive survey of benchmark datasets, syntactic/semantic evaluation metrics, and LLM enhancement techniques across fine-tuning, prompting, and verification-refinement loops.

LR2GS: New Intent Discovery from Local Relations to Global Structure

Addressing the challenge in New Intent Discovery (NID) where absence of explicit supervision hinders the reliable formation of well-separated global intent clusters, LR2GS starts from the foundational building blocks—local relations. By integrating inter-query semantic neighbor alignment, intra-query descriptive/prototype variant consistency, and cluster-boundary ambiguity calibration, LR2GS progressively shapes an optimal global intent representation space with superior intra-cluster compactness and inter-cluster separability.

NLP Meets the World: Toward Improving Conversations With News Media About Natural Language Processing Research

Addressing the escalating cognitive gap and distorted public perceptions surrounding LLMs, this paper adapts the US National Academies and EU QUEST science communication frameworks to formulate actionable recommendations for NLP researchers dealing with news media across cognitive terminology, unsustainable expectations, and ethical failures.

Not all Negation Cues are Equal: Affixal Negations Yield Better Negation Understanding

This paper constructs NegCue, the first large-scale (1.8M sample) corpus systematically categorizing 214 negation cues into affixal, single-word, and multi-word types, and reveals through Next Sentence Polarity Prediction (NSPP) pre-training that affixal negations yield substantially larger generalization gains across five downstream reasoning benchmarks than the ubiquitously studied single-word cues.

OBJECTION! Lawyer Agents Mitigate Guilty Bias in Legal Judgment Prediction

Addressing the critical "Guilty Bias" in Legal Judgment Prediction (LJP) caused by prosecutorial narratives in court records, this paper introduces OBJECTION—a training-free inference-time framework that embeds an adversarial Lawyer Agent into a three-step legal reasoning structure (Offense, Unlawfulness, Culpability)—alongside a new 3.4k real-world "Natural Innocent" benchmark, drastically slashing the False Guilty Rate from 82.93% to 16.69%.

On-Policy Distillation with Curriculum Turn-level Guidance for Multi-turn Agents

Addressing the failure mode where compounding errors in multi-turn On-Policy Distillation (OPD) drive rollouts out of the teacher's familiar distribution, this paper proposes Guided-OPD, which injects teacher turns via a cosine curriculum schedule, pairs asymmetric KL divergences according to the acting role, and eliminates training-inference mismatch.

On the Role of Supporting Evidence in LLM Reasoning: A Causal Analysis of Legal QA with Activation Patching

This paper applies activation patching and logit lens techniques to four open-source LLMs on U.S. MBE and Brazilian OAB bar exam benchmarks, demonstrating that retrieved supporting evidence causally steers legal answer selection primarily within the upper layers (exhibiting a late-binding signature), with causal integration heavily dictated by the discourse structure of the evidence rather than mere semantic relevance.

One Model, Many Graphs: Learning over Attributed Graphs across Heterogeneous Schemas

OMG-VLM presents a unified vision-language model (VLM) framework for attributed graph learning across heterogeneous modality schemas, utilizing a target-aware textual aggregation adapter and a graph-aware visual compression adapter to seamlessly process text-attributed graphs (TAGs), image-attributed graphs (IAGs), and multimodal-attributed graphs (MAGs) within a shared native token space for end-to-end node classification and link prediction.

One-Token Verification for Reasoning Correctness Estimation

OTV augments a reasoning LLM with a LoRA-gated verifier that probes the ongoing key-value (KV) cache via a dedicated [ToT] token, estimating token-level reasoning correctness in a single forward pass without degrading primary reasoning and saving up to 90% of generation tokens via confidence-guided pruning on AIME.

Ontology-Based Knowledge Graphs for Industrial Standards with Scope-Aware Logical Rule Structuring

Addressing the vulnerability of standard RAG systems to cross-clause rule collision in hierarchical, table-dense industrial specifications, this paper proposes an Ontology-based KG-RAG framework that formalizes document section hierarchies into explicit scope ontologies and couples them with condition–consequence rule subgraphs utilizing case and logic_group nodes. Evaluated on the expert-validated IndusSpec-QA benchmark across three real-world standards, the proposed approach significantly outperforms traditional text- and graph-based baselines, demonstrating a 120.2% relative F1 gain on complex table-based reasoning.

OP-Bench: Benchmarking Over-Personalization for Memory-Augmented Personalized Conversational Agents

OP-Bench formalizes and benchmarks the failure modes of over-personalization in memory-augmented conversational LLMs across Irrelevance, Sycophancy, and Repetition, revealing that memory integration causes 26.2% to 61.1% performance degradation due to memory hijacking and exposes a persistent trade-off in existing post-processing mitigations.

OrgAccess: A Benchmark for Role-Based Access Control in Organization Scale LLMs

OrgAccess is an expert-seeded, programmatically validated benchmark derived from NIST security control frameworks to evaluate compositional role-based access control (RBAC) in LLMs across 40 permission types and 70,000 instances; evaluations of 16 LLMs show performance degrades drastically to <= 0.27 macro-F1 under five concurrent constraints, uncovering a systemic safety-hedging bias.

Osprey: Target-agnostic Pre-training Makes Stronger Drafters in Speculative Decoding

Osprey introduces a decoupled training paradigm for speculative decoding drafters that bootstraps a shallow next-token backbone via reusable, target-agnostic pre-training, and lightly adapts it to diverse target models via vocabulary alignment, zero-initialized QKV expansion, and multi-step distillation—delivering up to 22.7% higher acceptance length and 17.5% throughput gains over EAGLE-3.

PAC: Progress-Augmented Advantage Curriculum for Multi-Task Reinforcement Learning of LLMs

Addressing the non-stationary task utility challenge in multi-task RL post-training for LLMs, PAC adaptively allocates rollout budget via a Bayesian Thompson Sampling controller that multiplicatively fuses advantage-derived update magnitude with recent reward-derived progress gains.

ParaRater: Enhancing Cross-Lingual Transfer in LLMs with Meta-Learning

To eliminate pseudo-parallel artifacts and heuristic bias in multilingual pretraining, ParaRater introduces a bi-level meta-learning framework that decouples monolingual quality from bilingual coupling via a two-stage counterfactual ranking process, identifying the core parallel documents that genuinely induce cross-lingual transfer under a tiny 1B-token budget.

Perceive, Localize, Reflect: Evidence-Anchored Multi-Agent Debate for MLLM Hallucination Mitigation

Uncovering why unguided Multi-Agent Debate (MAD) fails in vision-language domains due to a zero-drift stochastic martingale trap and context-induced visual attention decay, EVIAMAD proposes a training-free framework that enforces asymmetric perception, distills linguistic disputes into localized physical image crops, and surgically reclaims sink token attention to establish verifiable factual convergence.

PERIBENCH: A Protocol-Aware Benchmark for Dynamic Perioperative Risk Prediction from Numerical Models to Language Interfaces

PERIBENCH establishes the first protocol-aware benchmark for dynamic perioperative risk prediction on high-frequency intraoperative vital signs, structuring 18 clinical tasks across four regimes and evaluating 37 diverse models across three protocol-compatible pools to demonstrate that model complexity does not yield monotonic clinical gains.

Persona-CAPS: Controllable and Auditable LLM Personality via Retrieval over Situation–Behavior Contingencies

Persona-CAPS departs from monolithic trait prompts and opaque activation vectors by grounding LLM personality in Mischel and Shoda's Cognitive-Affective Personality System (CAPS): it represents a persona as a corpus of human Experience Sampling Method (ESM) if-then contingencies retrieved per situation, delivering continuous trait control via mixture weights without trait labels, while supporting per-response audit traces and atom-level repairs at ~1/3 the edit footprint of paragraph baselines.

Personalized Alignment of Language Models with Human Value Profiles

This paper introduces an end-to-end framework for personalized value alignment grounded in Schwartz's Refined Theory of Basic Human Values (19 discrete values), combining scenario-aware relevance gating to isolate top target and deprioritized values with targeted prompting, residual stream activation steering, and LoRA-DPO adapters, achieving state-of-the-art alignment with individual value priorities while preserving rigorous argumentation quality.

Perturbation: A simple and efficient adversarial tracer for representation learning in language models

This paper proposes Perturbation, an unsupervised and geometry-free interpretability method that treats latent representations as conduits for learning; by fine-tuning on a single adversarial example and tracing infection across unseen test items, it reliably uncovers linguistic structures across multiple grain sizes while naturally eliminating false-positive causal abstractions in untrained models.

Plans You Can Check: Verifier-Grounded Learning of an Open-Weight Planner for Executable Video-Editing

Addressing the reliance of video editing systems on unverified, prompted frontier models, this paper introduces RefineCut and the RefineCut-Bench benchmark, training an 8B open-weight planner through deterministic verifier-replayed multi-teacher distillation and rubric-grounded self-improvement (RefineCut-Evo) to match or exceed frontier models like GPT-5.4 in closed-loop plan execution.

Polished but Unresolved: Identifying Late-Stage Pressure States in Long-Horizon Tool-Use Agents

Addressing the pervasive failure mode where long-horizon tool-use agents prematurely output complete-looking, polished plans while leaving hard constraints unresolved, this work discovers and validates an internal, linearly separable "late-stage pressure state" and proposes Probe-Sensed Pressure Relief (PSPR) to dynamically mitigate premature closure via activation steering and explicit task reorganization.

PRICE: Preemptive Confidence Estimation in LLMs via Internal Representations

PRICE is a training-free, pre-generation confidence estimation framework for LLMs that extracts attention-head consensus and hidden-state semantic salience in a single forward pass, delivering reliable uncertainty quantification before decoding begins without training external probes or sampling multiple outputs.

PRISM: Agentic Retrieval with LLMs for Multi-Hop Question Answering

Proposes a training-free agentic retrieval framework named PRISM that decomposes multi-hop evidence selection into a collaborative loop between a Question Analyzer, a precision-focused Selector, and a recall-focused Adder, consistently outperforming full-context and strong iterative baselines across four major multi-hop QA benchmarks with 46% fewer tokens and 73% smaller context size.

PRISM Edit: One Vector for All Temporal Answers

Addressing the collapse of historical knowledge in temporal model editing, PRISM Edit uncovers the model's intrinsic two-stage temporal mechanism and optimizes a single polysemous vector that the model's native downstream attention pathways route to distinct temporal answers without architectural modifications.

PRISM: Entropy-Guided Multi-View Exploration for Scientific Idea Generation

Addressing the premature convergence and conceptual homogeneity of LLMs in open-ended scientific ideation, PRISM introduces an inference-time multi-view exploration framework that estimates view orthogonality via hypothesis-transition entropy and selects a maximally complementary subset to guide proposal-style scientific idea generation.

PRO-Step: Step-level Process Reward Optimization for Retrieval-Augmented Generation

Addressing error cascade propagation and outcome reward blind spots in agentic RAG, PRO-STEP trains a generative PRM evaluating both logical validity and evidential grounding, utilizes PRM-guided value tree search to construct contrastive preference pairs, and optimizes the policy via step-level DPO, achieving top average EM and F1 across five benchmarks.

ProactiveEngine: Proactive Learning with Lifelong Memory via World Models

This paper proposes ProactiveEngine, the first foundation model designed for proactive multimodal agents operating over lifelong memory, which unifies proactive response generation, latent world state prediction, and binary intervention classification across a 1M token context window, boosting the proactive prediction F1-score to 83.3% on the new 24.8k-session NeuroAuroraBench.

Probing Counterfactual Reasoning in Large Language Models

Investigating why large models excessively reject requests whose isolated harmful speech is rendered benign by surrounding acoustic contexts, this work introduces AOR-Bench with 3,000 pseudo-harmful audio queries across six scenarios, benchmarks 12 leading LALMs, and validates inference-time Chain-of-Thought and representation steering mitigations.

Rubric-Guided Process Reward for Stepwise Model Routing

Addressing the deceptive alignment trap of outcome-only rewards in stepwise model routing, RoRo trains a Rubricor and a Judge via alternating optimization and statistical validation gates to deliver dense process rewards, significantly boosting reasoning accuracy under tight compute budgets via GRPO.

Probing Counterfactual Reasoning in Large Language Models

This paper conducts a scaled-up empirical study of syntactic language models (SLMs) across up to 1B parameters and 100B tokens, revealing that standard causal attention matches or surpasses specialized structural masks while remaining fully compatible with FlashAttention, and showing that SLMs lose their clear edge over compute-matched non-structural baselines at scale.

AnyAudio-Judge: A Dynamic Rubric-Based Benchmark and Evaluator for Audio Instruction Following

AnyAudio-Judge establishes a dynamic rubric-based evaluation paradigm that decomposes complex audio instructions into independent, verifiable binary rubric items and aggregates soft satisfaction probabilities, supported by a 7,920-sample bilingual hard-negative benchmark (AnyAudio-Judge Bench) and a 105K CoT training corpus, where the resulting SFT+GRPO-tuned evaluator achieves state-of-the-art alignment detection and provides dense reward signals for downstream InstructTTS reinforcement learning.

Fathom-Vaidya: Advancing Medical Reasoning with Rubric-Based Rewards

Addressing the decoupling between convergent diagnostic inference and multi-turn open-ended clinical dialogue, Fathom-Vaidya introduces a sequential reinforcement learning framework without supervised Chain-of-Thought data, combining two-stage rule/error-rubric RL with an Actor-Critic synthetic clinical rubric reward system to achieve a state-of-the-art 50.1% on HealthBench-Hard with a 30B model.

LOKI: Locating Optimal Knock-In Points for Backdoor Attacks on Dense Retrievers

Addressing the security vulnerability of dense retrievers where coarse-grained parameter updates heavily degrade benign retrieval and trigger false alarms, this paper proposes LOKI—a Locate-then-Edit backdoor framework that identifies trigger-responsive, low-benign-overlap neurons via causal efficacy and semantic route analysis, achieving 100% Set-ASR@1 while modifying under 0.31% of parameters without perturbing clean retrieval utility.

FOCAL: Fine-Grained Optimal-Transport-Driven Contrastive Alignment of Language and ECGs with Waveform Enhancement

To overcome the neglect of localized waveform patterns in global ECG-report alignment and recover missing waveform descriptors omitted in ~55% of clinical reports, FOCAL introduces a Semi-Unbalanced Optimal Transport (Semi-UOT) routing mechanism, a semantic similarity matrix to cancel out tag-level false negatives, and an offline coarse-model-verified LLM enrichment pipeline, achieving a new state-of-the-art average zero-shot Macro AUC of 81.1% across six benchmarks.

X-Coder: Advancing Competitive Programming with Fully Synthetic Tasks, Solutions, and Tests

Addressing the critical bottleneck of real-world competitive programming data exhaustion and benchmark contamination, X-Coder establishes the first complete post-training pipeline driven entirely by synthetic tasks, solutions, and executable tests, enabling a 14B model to achieve state-of-the-art results of 67.5% and 63.4% avg@8 on LiveCodeBench v5 and v6.

Deeper is Not Always Better: Mitigating the Alignment Tax via Confident Layer Decoding

Addressing the "alignment tax" where final layers disrupt fragile logic chains, this paper reveals a three-phase Guess–Refine–Perturb dynamic in LLM forward passes and proposes Confident Decoding—a training-free, drop-in decoding strategy that dynamically rolls back to near-final local entropy valleys, restoring deep reasoning fidelity with zero extra KV-cache memory and under 2% latency overhead.

MAViQ: Memory-Augmented Visual Querying for Multi-Turn MLLM Jailbreaks

MAViQ formalizes multi-turn multimodal jailbreaking into a trace-to-query realization paradigm, decomposing dialogue feedback into explicit textual intent, visual roles, and cross-modal handoffs backed by dual-level memory and post-generation image grounding, surpassing the strongest baseline by 5.90% and 8.32% ASR across six target MLLMs.

RubricRM: Generative Reward Modeling via Dynamic Rubrics for Image Generation and Editing

Addressing the lack of interpretability and input-adaptability in scalar reward models and static MLLM judges for visual generation, RubricRM dynamically generates an input-adaptive rubric with atomic dimensions, normalized weights, and 5-level scoring criteria to evaluate candidate images in a single forward pass, optimized through trajectory SFT and fine-grained dimension-level GRPO.

SecAgent: Efficient Mobile GUI Agent with Semantic Context

SecAgent tackles prohibitive multi-image history overheads and the scarcity of high-quality non-English datasets in mobile GUI agents by introducing a natural language Semantic Context mechanism and releasing CMGUI, achieving superior 3B performance that rivals 7B-8B baselines.

CAVEWOMAN: How Large Language Models Behave Under Linguistic Input and Output Compression

CAVEWOMAN introduces a two-channel evaluation protocol across 9 LLMs and 5 benchmarks, demonstrating that output grammatical compression reduces realized API dollar cost by up to 3× without degrading task accuracy, yet over half of correct non-reasoning responses no longer semantically entail the model's unconstrained output, whereas input prompt compression paradoxically inflates net cost by up to 96%.

Probing Syntactic Knowledge in Instruction-Tuned LLMs

By establishing a rigorously controlled benchmark of 213 contrastive quadruples alongside sequential tuning schedules, this paper uncovers that LLM refusal operates via a unified "commit-then-specify" process, where an early shared refusal direction triggers abstention before late-layer heads and MLPs specify whether grounds are epistemic or normative.

Prompt Revision as a Source of Cultural Bias in Text-to-Image Systems

This paper introduces WORLDVIEW, an extensive multilingual benchmark (8,960 prompts across 15 languages and 31 language-context pairings), and audits the silent prompt revision layer across commercial text-to-image systems (DALL-E-3, Imagen-4, GPT-Image-1.5), demonstrating that the revision layer functions as an independent, causal driver of cultural marking, vocabulary flattening, and visual stereotyping.

Proof-of-Use: Mitigating Tool-Call Hacking in Deep Research Agents

Addressing the pervasive "tool-call hacking" failure mode in RL-trained multi-step deep research agents—where models learn to satisfy surface tool-calling formats without genuinely grounding their reasoning in retrieved content—this paper proposes Proof-of-Use (PoU), an evidence-grounded framework coupling normalized evidence ID citations, counterfactual perturbation rewards, global answer-support alignment, and adaptive reward annealing.

PsychēChat: An Empathic Framework Focused on Emotion Shift Tracking and Safety Risk Analysis in Psychological Counseling

PsychēChat integrates Emotion-Focused Therapy (EFT) to explicitly model seekers' cross-turn emotion shift trajectories and anticipate potential seeker reactions for proactive safety risk prevention, synthesizing 1,003 multi-turn counseling dialogues (PsychēDialog) and providing both multi-agent collaborative and end-to-end long chain-of-thought inference paradigms that outperform existing psychological LLMs on SAGE and ESC-Eval benchmarks.

QAQ: Bidirectional Semantic Coherence for Selecting High-Quality Synthetic Code Instructions

Addressing subtle hallucinations and defect patterns in seedless synthetic data, QAQ evaluates data quality from the reverse direction by measuring how well the answer predicts the query via Reverse Mutual Information (RMI), filtering both misalignment and trivial keyword echoes to match full-dataset fine-tuning performance with only 25% data.

Quantifying and Mitigating Korean Jamo-Level Typographical Vulnerabilities in Large Language Models

This paper systematically exposes the acute vulnerability of large language models to intra-syllabic Korean jamo-level typos, proves that such corruptions induce an internal representational shift orthogonal to ordinary answer incorrectness, and introduces Typo-Aware Chain-of-Thought (TACoT)—a dynamic probe-based routing framework that selectively invokes CoT to recover benchmark reasoning accuracy while slashing generation token costs by roughly 37%.

QUORUM: Quality-Optimized Routing Using Multiple annotators

QUORUM introduces an uncertainty-free, budget-aware data annotation routing framework that extracts lightweight linguistic and semantic features to estimate instance difficulty, adaptively routes samples across humans and heterogeneous LLMs via online Bayesian regression, and triggers disagreement-guided re-annotation and consensus aggregation to boost quality while lowering costs.

R^3L: Reflect-then-Retry Reinforcement Learning with Language-Guided Exploration, Pivotal Credit, and Positive Amplification

Addressing inefficient exploration, coarse credit assignment, and entropy collapse in LLM reinforcement learning, R3L synthesizes high-quality trajectories via language-guided reflect-then-retry from failure pivots, isolates diverging suffixes via pivotal prefix masking, and amplifies constructive advantages to achieve stable optimization without KL or importance sampling penalties.

R-IDEAL: Reinforcing Insight-Driven Experience for Agent Learning

Addressing language agents' failure to learn from negative experiences and lack of interactive feedback in long-horizon tasks, R-IDEAL synthesizes structured insights combining positive plans and negative lessons from trajectory contrasts, then reinforces a lightweight insight generator via GRPO with heterogeneous agent feedback for plug-and-play cross-model guidance.

RACE: Scalable Statistical Estimation of Functional Consistency in LLM Neurons

Addressing the excessive computational cost and sample-to-sample variance oversimplification in existing mechanistic interpretability, this paper introduces RACE, a forward-pass statistical framework that decomposes module residual updates and leverages Normal-Inverse-Gamma conjugate Bayesian modeling to rank domain-functional neurons with extreme scalability and high intervention specificity.

Rashid: A Cipher-Based Framework for Exploring In-Context Language Learning

This paper introduces Rashid, a controlled sandbox that reversibly ciphers high-resource languages into artificial unseen languages to preserve underlying linguistic structures while repurposing rich NLP toolchains, systematically diagnosing the severe comprehension-versus-generation asymmetry in in-context language learning and demonstrating that pivoting through related high-resource languages resolves syntactic degradation.

RASST: Retrieval-Augmented Simultaneous Speech Translation

This paper proposes RASST, the first retrieval-augmented framework tailored for simultaneous speech translation, combining a lightweight multi-scale cross-modal speech-text retriever with noise-robust Speech LLM training to accurately retrieve domain-specific terminology under partial streaming speech and decide when to apply retrieved translations, boosting terminology accuracy by 12.3%–31.22% and BLEU by up to 3 points over InfiniSST with negligible computational overhead.

Reactivating Alignment: Defending LLMs from Jailbreaks via Intention-Aware Input–Output Matching

This paper introduces SENTINEL, a training-free, generation-time jailbreak defense framework that leverages the inherent input–output semantic consistency of instruction-tuned LLMs to extract core malicious intent via regularized context window matching and halts unsafe decoding using refusal-direction projections, suppressing attack success rates below 5% with low over-refusal.

Reactivating Test-Time Scaling for Plane Geometry Problem Solving

This paper diagnoses the breakdown of test-time scaling in plane geometry solving caused by rigid symbolic representations and perceptual errors, and re-activates scaling via Multi-Trace Synthesis (MTS), Perception-Augmented (PA) training, and Consensus-Guided Multi-Trace Ensemble (CG-MTE), matching high-budget self-consistency accuracy while slashing sampling overhead by up to 8x.

ReaderScope: From Offline Reader-Response Calibration to Online Controlled Story Generation

Addressing the limitation that long-form story generators optimize plot coherence while neglecting the reader's evolving reading experience, ReaderScope calibrates four-dimensional reader-response trajectories (emotion, entropy, immersion, insight) offline across 20 reference novels and steers online generation through closed-loop reader-feedback revision.

Reasoning Confidence Is Not Policy-Invariant: Calibrating Prefix Success under Model and Budget Shift

Addressing the issue that reasoning prefix confidence is not policy-invariant and that raw hidden-state conditioning collapses across model families, this paper proposes the Comparable Policy-Conditioned Calibrator (CPCC), which uses a shared frozen encoder, 19-dimensional cross-model comparable scalar descriptors, and a Beta–Binomial head to achieve robust prefix success calibration and uncertainty estimation under model and budget shift.

Reasoning Under Uncertainty: Exploring Probabilistic Reasoning Capabilities of LLMs

This paper establishes a comprehensive benchmark evaluating large language models on discrete joint and conditional distributions across mode identification, maximum likelihood estimation, and sample generation, revealing substantial performance drops in conditional queries, mechanistic suppression in final transformer layers, and non-independent frequency-matching heuristics.

ReasoningFlow: Discourse Structures for Understanding LLM Reasoning Traces

ReasoningFlow formalizes the non-linear reasoning traces of Large Reasoning Models into fine-grained directed acyclic graphs with 8 node types and 14 edge types, demonstrating that 79.6% of erroneous intermediate steps in models like DeepSeek-R1 and QwQ are safely bypassed during backtracking or local verification without degrading final answers.

RECAP: Regression Evaluation for Continual Adaptation of Prompts

Addressing production agentic setups where models must immediately comply with dynamic constraint specifications without test data or feedback, this paper introduces RECAP—the first constraint-level continual learning benchmark—and reveals that existing prompt adaptation methods fail to beat a simple constraint-appending baseline while inducing context bloat, spurious conflicts, and severe refusal cascades.

Reconstructing the Right Episode: Evaluating Interleaved Conversational Memory Beyond Long Context

Targeting "episode integrity failure" in flat unsegmented multi-turn conversations where local constraints are severed by mechanical chunking, this paper introduces the 3,000-question benchmark SCALE-QA and proposes TSIM, a lightweight multi-view episodic reconstruction and ranking stack that achieves state-of-the-art accuracy with merely ~1k prompt tokens across diverse LLM backends.

ReflectRec: Cycle-Consistency Preference Verification for LLM-based Recommendation

ReflectRec addresses the twin vulnerabilities of CoT-style LLM recommenders—unverified intermediate preference rationales and the total absence of collaborative filtering signals—by introducing a preference-to-history cycle-consistency verifier whose recall score governs measurement, gating, selection, and diagnosis, yielding significant Hit@10 gains of +16.25 pp on ML-100K and +4.46 pp on ML-1M over CoT4Rec-ICL.

Reinforcement Learning Improves Traversal of Parametric Knowledge in LLMs

This work demonstrates that reinforcement learning (RL) improves factual recall in LLMs not by injecting new facts, but by internalizing a systematic hierarchical traversal policy over existing latent parametric knowledge; through relation-disjoint knowledge-injection experiments and layerwise representational probing, the authors prove that iterated on-policy exploration reshapes question-navigation processing while leaving factual memory representations intact.

Reliability of LLM Judges for Evaluating Entity Alignment

This paper exposes the severe vulnerability of LLM judges to "anchor bias" in entity alignment evaluation, where displaying system decision labels causes correctness discrimination to catastrophically invert (J-ROC-AUC dropping to 0.12–0.87) with stronger frontier models rationalizing labels even more aggressively (the frontier model paradox); withholding system labels under a label-free prediction protocol completely restores near-ceiling discrimination capability (0.93–1.00).

Repetition Mismatch: Why Data Mixture Experiments Don’t Scale and How to Fix Them

This paper demonstrates that the persistent failure of small-scale data mixture extrapolation stems from "repetition mismatch" caused by shifting multi-epoch regimes as budgets scale. By subsampling high-quality data to strictly lock the target repetition rate, a single proxy run using just 1/16 of the target token budget slashes the mixture prediction error from 0.85 to 0.10 on a 1.17B model.

Repetition over Diversity: High-Signal Data Filtering for Sample-Efficient German Language Modeling

Addressing the acute data bottleneck in non-English high-resource languages where aggressive filtering drastically shrinks token budgets, this paper demonstrates that multi-epoch training over a compact, high-density core (DENSE-CORE) substantially outperforms single-pass pre-training over larger, less-filtered corpora under fixed compute budgets, establishing state-of-the-art results for the open-source BOLDT small language model series.

RepoNav: From Snippet Retrieval to File-Centered Repository Navigation for Code Agents

Addressing the pervasive "file-to-function gap" where code agents successfully surface the correct file but fail to locate the target function, RepoNav introduces a lightweight post-retrieval interface that reorganizes flat snippets into file-centered tree scaffolds paired with an on-demand symbol browsing tool, significantly boosting function-level localization across seven LLMs without requiring repository-wide graph indexing.

Representation-Aware Advantage Estimation: Your Reward Model Provides More Than A Scalar Output

Addressing the vulnerability of existing group-based RLHF methods to noisy scalar rewards and overlooked preference topologies, this paper introduces Graph-based Advantage Estimation (GraphAE). By constructing an adaptive similarity graph using reward model penultimate hidden states and deriving smoothed rewards via closed-form graph Laplacian regularization, GraphAE achieves substantial win-rate gains of up to +6.3 on Arena-Hard and +8.27 on AlpacaEval 2.0 while cutting 40% of required training steps.

Rethinking the Role of Temperature in Large Language Model Distillation

This paper revisits the neglected role of distillation temperature in LLMs, revealing that temperature inherently changes the supervision targets for Forward KL while primarily rescaling optimization gradients for Reverse KL, thereby overturning the conventional belief that RKL intrinsically outperforms FKL and demonstrating that simple KL objectives with tuned temperature match or beat recent SOTA distillation frameworks.

Retrievable Gradients: Continual Post-Training Without Cumulative Weight Drift

Introduces REGRAD, a paradigm that treats document-induced gradients as retrievable knowledge units stored in an offline Gradient Bank, dynamically applying meta-learned gradients for query-specific, reversible parameter adaptation without cumulative weight drift or catastrophic forgetting.

Retrievable Gradients: Continual Post-Training Without Cumulative Weight Drift

REGRAD treats document-specific parameter gradients as retrievable units of knowledge stored in an offline Gradient Bank, reshaping unsupervised causal language modeling updates into downstream QA adaptation directions via bi-level meta-learning to achieve query-specific parametric knowledge injection without cumulative weight drift.

RetrievalRouter: Joint Modality and Architecture Selection for Document Retrieval

Addressing the fundamental tradeoff where multimodal late-interaction retrieval is prohibitively slow while lightweight dense retrieval fails on visually complex layouts, RetrievalRouter introduces a query-aware policy that jointly predicts both pipeline modality and architecture from query text alone, delivering 2.5% higher nDCG@5 than the strongest static pipeline while accelerating inference by 12.4× across 11 benchmarks.

Right Answers for the Right Reasons: Coherence-Enhanced Policy Optimization for Large Language Models

To tackle shortcut-driven reward hacking and spurious reasoning in outcome-rewarded LLMs, this paper proposes Coherence-Enhanced Policy Optimization (CE-PO), which regularizes the prompt-to-rationale-to-answer pathway using differentiable internal Jacobians, neutralizes length and format shortcuts via reshuffling-based residualization, and aggregates coherence with task accuracy via a tunable \(L_p\) combiner in PPO/GRPO.

Risk-Adjusted Harm Scoring for Automated Red Teaming of LLMs in Financial Services

Addressing the failure of binary Attack Success Rate (ASR) to capture harm severity and its vulnerability to saturation in regulated finance, this paper introduces FinRedTeamBench (989 expert prompts across 34 BFSI risk areas), the Risk-Adjusted Harm Score (RAHS) metric incorporating disclosure severity, disclaimer mitigation, and inter-judge entropy penalties, alongside an adaptive multi-turn red-teaming pipeline.

SABER: Stability-Aware Early Exit for LLM Reasoning via Adversarial Branch Probing

To tackle excessive token expenditure and late-stage error accumulation caused by overthinking in Large Reasoning Models (LRMs), SABER constructs lightweight neutral and adversarial probe branches at intermediate reasoning transitions to assess stability under semantic perturbation, cutting token consumption by 30.2%–39.8% without extra training while maintaining competitive accuracy.

Safety Paradox: How Enhanced Safety Awareness Leaves LLMs Vulnerable to Posterior Attack

This paper uncovers a profound "Safety Paradox" in modern LLM alignment paradigms—the more proficient a model becomes at evaluating and recognizing unsafe content, the more susceptible it is to a single-query Posterior Attack that weaponizes its internal safety classifier; through Bayesian derivation and GRPO interventions, the authors prove that enhanced safety awareness monotonically amplifies posterior vulnerability.

SALT: Salience-Aware Lexical Trie for Long-Context Compression

Addressing the pervasive issue of theme collapse in scalar-ranked prompt compression, SALT introduces an unsupervised, model-agnostic extractive framework that organizes sentence keywords into a sentence-frequency-ordered trie, explicitly distributing compression budgets across document themes before sentence selection to slash prefill latency and memory footprint.

Same Payload, Different Channel: Measuring Trust Asymmetry in Tool-Using Language Models

This paper introduces the Safety Asymmetry Score (SAS) and its affordance-controlled counterpart \(\text{SAS}_{\text{trust}}\) using matched payload pairs to reveal that LLMs systematically favor tool description metadata over user messages—especially agent-native models—while within the tool interface, models treat metadata as authoritative instructions and tool return outputs as passive data, driven by causally load-bearing, non-linearly encoded representations at mid-to-late depths.

Same Trajectory, Contradictory Rewards (RoboRMBench): Paraphrase Fragility in Vision Language Reward Models

This paper introduces RoboRMBench, an evaluation benchmark measuring paraphrase robustness in vision-language robot reward models, revealing that general-purpose VLMs exhibit up to 60.7% failure-success score reversals under semantically equivalent goal rewrites for identical robot trajectories, while trajectory-grounded supervision and variance reduction training effectively alleviate this vulnerability.

SAND: Subjective-Anticipation-Augmented News Recommendation via Dual-Pathway Distillation

Addressing the fundamental gap where conventional news recommenders rely exclusively on unobserved objective article bodies, SAND introduces the concept of an "anticipated body" driven by user-specific cognitive expectations, synthesized via an offline three-LLM collaborative system and aligned using candidate-aware attention alongside dual-pathway knowledge distillation.

Saying vs. Seeing: Evaluating LLM Conflict Awareness and Resolution Preference under Latent Claim-Evidence Disagreement

Introduces CLECK, the first counterfactual benchmark evaluating non-explicit knowledge conflicts, revealing a widespread "Detection Gap" where LLMs retrieve competing answers in isolation but fail to surface conflicts during QA, and demonstrating that conflict awareness serves as the pivotal behavioral gate shifting preference from authoritative claims toward derived evidence.

SCALE: Structured Case-law-guided Anticipatory Legal Reasoning

Addressing standard legal RAG's vulnerability to surface-level lexical similarity and the burden of synthesizing dispersed case judgments, SCALE compiles article-specific Case Law Guides into an Issue–Factor Hierarchy (IFH) to drive anticipatory, factor-specific precedent retrieval, markedly boosting judicial reasoning consistency and precedent faithfulness.

ScalePRM: Training Process Reward Models by Scaling Verification Compute Without Ground Truth

ScalePRM generates synthetic step-level process supervision by scaling inference-time verification compute (parallel step-level self-consistency and sequential meta-critique) without ground truth or human annotation; generative PRMs trained on this data achieve 67.5 F1 on ProcessBench (surpassing reference-guided models and GPT-4o) and reach 47.4% average accuracy when serving as RL rewards under GRPO, exceeding ground-truth RLVR (43.9%).

SCOPE: Signal-Calibrated On-Policy Distillation Enhancement with Dual-Path Adaptive Weighting

SCOPE addresses the structural flaw of uniform dense supervision in On-Policy Distillation (OPD) by routing rollouts by correctness into a dual-path adaptive weighting framework: filtering flawed prefix noise on incorrect rollouts via inverse teacher perplexity, and amplifying under-explored boundary reasoning on correct rollouts via student perplexity, decisively breaking the Pass@k paradox.

SCOPE: Signal-Calibrated On-Policy Distillation Enhancement with Dual-Path Adaptive Weighting

SCOPE addresses the uniform signal-quality assumption in On-Policy Distillation by partitioning student rollouts by outcome correctness into two complementary supervision paths: applying inverse teacher-perplexity-weighted reverse KL distillation to filter out misleading noise from flawed prefixes on incorrect trajectories, and student-perplexity-weighted MLE to amplify low-confidence capability-boundary exploration on correct trajectories, coupled with group-level normalization to preserve reasoning diversity and boost pass rates.

Seeing Before Synthesizing: VLM-Guided Transition Event Discovery for Weakly-Supervised Dense Video Captioning

Addressing the severe semantic hallucinations and rigid temporal misalignment caused by blind LLM text synthesis over inter-event gaps in weakly-supervised dense video captioning, this work proposes Seeing Before Synthesizing (SBS). By repurposing a VLM as an event discovery probe, SBS adaptively decides whether a transition exists via statistical narrative shift gating and where it occurs through change point blending and cross-modal width selection, establishing new state-of-the-art results on both ActivityNet Captions and YouCook2.

Seeing Through the MiRAGE: Evaluating Multimodal Retrieval Augmented Generation

This paper introduces MIRAGE, the first subclaim-centric evaluation framework for multimodal retrieval-augmented generation (RAG), which decouples information quality (INFOF1: precision and recall) from citation attribution (CITEF1: support and completeness) at the atomic subclaim level across text, audio, and video sources, resolving the fundamental inability of existing sentence-level text metrics to evaluate multimodal grounding.

Select, Label, Evaluate: Active Testing in NLP

This paper formalizes Active Testing for NLP benchmarks, introducing an unsupervised sample acquisition strategy based on attention-head variance (Agreement) paired with an adaptive stopping criterion, which reduces human annotation overhead by up to 95% across 18 datasets while guaranteeing performance estimation errors within 1%.

Selective State-Space Adaptation and Retrieval for Language Model Reasoning

This paper introduces a two-level selective state-space adaptation family that deploys recurrent Mamba modules at both the token level to dynamically modulate LoRA scaling factors (MaLoRA) and the context level to decode supporting evidence directly from frozen intermediate states (MaRA), achieving a +6.4 F1 (+10.0% relative) gain over LoRA on multi-hop reasoning.

Self-Evolving Vision Language Models for Structured Visual Understanding From Zero Data

MM-Zero is the first reinforcement learning framework to accomplish zero-data self-evolution for vision-language models, decomposing a single base model into a collaborative triad—Proposer (abstract design & QA), Coder (executable SVG rendering), and Solver (multimodal reasoning)—to achieve continuous reasoning gains without any human annotations or seed images.

Semantic Distractor Robustness in LLMs: Exploring the Underlying Mechanisms

Addressing the severe performance collapse of LLMs under distracting contexts in multi-hop reasoning, this paper bypasses the narrow tracking of pre-defined distractor tokens by localizing components via the model's actual top non-gold predictions. Across four base models, failures converge onto a compact set of 19–35 mid-to-late-layer Answer-Selection Heads (AS-Heads), whose attention patterns causally dictate entity transcription and provide an effective, training-free inference-time context filter that lifts disrupted sample F1 by up to +35 points.

Send a SCOUT First: Pre-hoc Reasoning for Adaptive Detector Allocation in Prompt-Injection Defense

SCOUT reframes prompt-injection defense as sample-adaptive detector allocation by using offline behavioral fingerprints and a lightweight predictor to pre-estimate per-detector correctness and latency, achieving a 46% reduction in attack pass-through rate and 40% lower wall-clock time compared to an always-on GPT-4o judge on SCOUT-450.

SeVeR: Selective Visual Exposure and Retrieval for 3D Medical Image Question Answering

To resolve the core challenge of visual token redundancy and signal dilution in multi-sequence volumetric MRI reasoning, this paper constructs BreMRIs-VQA—a clinically curated benchmark with 1.19M QA pairs from 12.9K patients—and presents SeVeR, a selective visual exposure framework that compresses dense 3D volumes into compact prototypes via Greedy Prototype Selection (GPS), dynamically retrieves complementary evidence via Change-aware Gated Attention (CaGA), and enforces Marginal-Utility Self-Consistency (SCR-MU) to surpass full-token baselines while exposing substantially fewer tokens.

SHAPE: Structural Hypergraph Augmentation with Personalized Event Memory for Temporal Knowledge Graph Reasoning

SHAPE is a plug-and-play enhancement framework for temporal knowledge graph reasoning that injects higher-order prototype hypergraph priors before temporal encoding and adds query-denoised event memory bias after encoding, delivering consistent state-of-the-art gains across diverse backbones.

Shaped by Values: Language Through the Lens of Culture and Morality

This paper introduces MCWR, a two-stage multilingual dataset spanning US English, Mexican Spanish, Chinese, and Arabic that couples generated texts with self-reported value profiles for both writers and readers, uncovering that reader value inferences are heavily driven by an egocentric projection bias and that reader-writer value distance is the primary source of perceptual misalignment.

Share More, Search Less: Collaborative Parallel Thinking for Efficient Test-Time Scaling

Addressing the information-isolation bottleneck in parallel test-time scaling where isolated branches repeatedly rediscover identical intermediate steps, Collaborative Parallel Thinking (CPT) extracts compact discoveries into a deduplicated query-level pool and adaptively broadcasts them across input contexts, significantly improving the accuracy–latency Pareto frontier without additional training.

SHIFT: Self-reconstruction Harnesses Implicit Fine-grained Thinking for Retrieval

This paper proposes SHIFT, an efficient reasoning-intensive retrieval framework that decouples generation noise from metric embeddings via residual projection, dynamically aggregates continuous latent thoughts through bidirectional self-attention, and supervises the latent reasoning space via fine-grained autoregressive self-reconstruction of explicit chain-of-thought trajectories with 2D relative position encodings—surpassing both cascade rewriting pipelines and prior implicit reasoning retrievers with only 3 latent thinking steps.

SIC-Agents: Benchmarking and Building an Adaptive Simulator for Pediatric Serious Illness Communication Training

Addressing the critical absence of scalable clinical training substrates in pediatric serious illness communication (SIC), this paper introduces the first paired benchmark suite (PitfallBench and DialogueBench) and develops SIC-Agents, a multi-agent self-improving simulation framework that evolves an inspectable, clinician-editable natural-language skill document under pedagogical curriculum guidance to avert clinical drift.

SimLens: Two-Token Continuation for Early Exit in Single-Token Decision Tasks

For single-token decision workloads like multiple-choice QA and boolean verification, SimLens introduces a training-free readout that forwards only the start token and candidate answer token through remaining upper layers to faithfully elicit latent predictions, which combines with a distilled linear entropy monitor in SimExit to deliver up to a 1.40× speedup under fair-compute accounting.

SkillPyramid: A Hierarchical Skill Consolidation Framework for Self-Evolving LLM Agents

Addressing the inability of flat skill libraries to support structured reuse and cross-task generalization in LLM agents, SkillPyramid proposes a bidirectional hierarchical consolidation mechanism (downward atomic extraction and upward abstract induction) paired with decoupled synthesis and task-validated incremental self-evolution, improving average reward by 38.0% while reducing execution steps by 27.7%.

SlideForge: Controllable Editing of Slides as Structured Artifacts

Addressing the critical failure modes of existing presentation agents that treat slides as flat pixels and corrupt editable structures, SLIDEFORGE conceptualizes slide editing as structured-artifact manipulation. By linking native PPTX XML objects and visual decomposition into an executable Deck State Graph (DSG), it conducts grounded planning, modality-aware element reconstruction, and closed-loop visual self-repair, maintaining layout fidelity and native editability while achieving superior stylistic transformation.

Smooth Operator: Smooth Verifiable Reward Activates Spatial Reasoning Ability of Vision-Language Model

Addressing sample inefficiency and the near-miss waste problem in reinforcement learning for continuous geometric tasks, this paper introduces Smooth Numerical Reward Activation (SNRA) and Absolute-Preserving GRPO (AP-GRPO), activating precise 3D spatial reasoning in standard vision-language models using only 50K targeted samples without architectural modifications.

SocialArena: Evaluating Social Intelligence in Large Language Model Agents via Multi-Party Strategic Interactions

SocialArena establishes a unified three-tier framework defining social intelligence across role-playing, 7-dimension Theory of Mind (ToM) mental state reasoning, and contextual behavioral reasoning; backed by the orthogonal U-TRIMF scenario synthesis pipeline and a $132.4K human-annotated pairwise arena protocol across 15 frontier LLMs, it reveals that models suffer from an acute "knowing vs. acting" perception-action gap and that automated LLM judges severely misalign with human social consensus.

soft tokens are off manifold geometric correction for continuous space reasoning

Addressing the dual failure modes of continuous-space reasoning—low-entropy collapse toward the greedy anchor and high-entropy off-manifold scale drift—this paper proposes Manifold Thinking, a training-free framework combining structured orthogonal tangent navigation with local norm rescaling that outperforms discrete and soft-token baselines across 18 benchmark configurations.

SOLID: Infusing Symbolic Logic into LLMs via Dual-Verified Backward Reasoning

SOLID introduces a neuro-symbolic framework that synthesizes high-quality supervision via bidirectional consistency filtering and full trial-and-error backward reasoning traces, enabling open-source 7B LLMs to achieve 97.9% average accuracy across ProntoQA, ProofWriter, and LogicNLI without relying on external solvers at inference.

Span-Based Reasoning with Anonymous Knowledge Graphs

GRASE replaces costly, repetitive LLM-driven graph exploration in multi-hop KGQA with fuzzy span matching over an Anonymous Knowledge Graph (AKG) using a frozen span encoder and completeness classifier, slashing LLM calls and query latency by ~85% while achieving state-of-the-art accuracy across complex benchmarks.

SPARROW: Scalable Taxonomy Induction via Structure-Preserving Partitioning and Constraint-Guided Merging

To overcome context window constraints and global structural reasoning failures in LLM-based taxonomy induction over large flat concept sets, SPARROW introduces a divide-and-merge framework coupling structure-preserving spectral graph partitioning with constraint-guided incremental fusion, achieving state-of-the-art ancestor-level structural consistency at scales up to 10,000 concepts.

SpiralFormer: Looped Transformers Can Learn Hierarchical Dependencies via Multi-Resolution Recursion

SpiralFormer executes a coarse-to-fine multi-resolution recursion schedule over a single parameter-shared Transformer core combined with strictly causal down/up-scaling and half-chunk offsets, reducing prefill FLOPs by 7–11% and decoding latency by 11.4% while consistently outperforming both full-resolution looped and standard non-recursive baselines from 160M to 1.4B parameters.

SPIRE: Unifying Partial Rollouts, Replay, and Asynchrony in LLM Reinforcement Learning via Segment-wise Prefix Importance Sampling

By formalizing partial rollouts, experience replay, and asynchrony as segmented trajectories generated by heterogeneous behavior policies, SPIRE introduces an unbiased policy gradient estimator with closed-form Rényi-2 variance clipping and Pareto-optimal replay mixing for LLM reasoning.

SPRInG: Continual LLM Personalization via Selective Parametric Adaptation and Retrieval-Interpolated Generation

Addressing preference drift and catastrophic forgetting in evolving user interaction streams, SPRInG introduces a semi-parametric continual personalization framework that selectively fine-tunes LoRA adapters on high-novelty drift samples while retaining residual hard cases in a replay buffer, dynamically interpolating parametric and gated non-parametric probabilities during decoding to outperform state-of-the-art continual learning baselines by over 15% in ROUGE-L.

Stable Adaptive Thinking via Advantage Shaping and Length-Aware Gradient Regulation

Addressing the pervasive overthinking problem and optimization collapse in large reasoning models on simple queries, this paper proposes a two-stage adaptive thinking framework combining Hybrid Fine-Tuning (HFT), Correctness-Preserving Advantage Shaping (CPAS), and Length-Aware Gradient Regulation (LAGR) to cut token generation by over 40% while simultaneously boosting mathematical reasoning accuracy by up to +3.7 points.

Stop Before You Fail: Operational Capability Boundaries for Mitigating Unproductive Reasoning in Large Reasoning Models

By identifying early predictive signals of failure from reasoning expression trajectories (black-box) and last-token hidden states (white-box), this work develops test-time monitoring strategies that enable large reasoning models to abstain from unsolvable questions and provide concise approach outlines, slashing unproductive token consumption by 62.7%–93.6% without sacrificing accuracy on solvable problems.

Stride-k Subsampling: Train-Free Audio Token Reduction for Whisper

By identifying the physical acoustic redundancy arising from Whisper's 65ms effective receptive field against its 20ms token spacing, this paper introduces Stride-k subsampling—a completely training-free indexing operation (\(H' = H[::k, :]\)) at both the convolutional stem and encoder output that eliminates 75% of audio tokens and 52%–58% of total GFLOPs with marginal accuracy cost, speeding up SpeechLM end-to-end latency by 19.6%–27.4%.

Structured Claim-Level Discourse Representations for Dense Health Narratives

To resolve multi-aspect entanglement in short-form health videos where speakers assert an average of 13.22 claims per minute, this paper introduces a structured discourse framework that maps raw conversational transcripts into atomic claim tuples \((c, a, s, \vec{\phi})\) across domain aspects, stances, and a 6-axis pragmatic typology, revealing critical trade-offs between local semantic anchoring and global context degradation.

Subliminal Clocks: Latent Time Modelling in Diffusion Language Models

This paper demonstrates through probing and causal activation steering that absorbing-state Diffusion Language Models (DLMs) lacking explicit timestep conditioning spontaneously encode a continuous, ordered, and low-dimensional manifold of denoising progress (\(\tau\)) within their residual streams, allowing systematic modulation of output entropy and confidence.

Subspace Alignment Score: Locating Domain-Relevant Important Parameters in Large Language Models

This paper proposes the Subspace Alignment Score (SAS), a geometric parameter evaluation framework that identifies domain-relevant important FFN parameters by projecting sequence-averaged value vectors onto low-dimensional principal component subspaces of the residual stream. Structured fine-tuning of just 2%–10% of these identified neuron channels achieves performance comparable to full fine-tuning with substantially lower peak GPU memory and superior generalization.

SumRank: Aligning Summarization Models for Long-Document Listwise Reranking

Proposes SumRank, a per-document summarization model aligned with downstream listwise reranking through a three-stage training framework (cold-start SFT, specialized static RL candidate construction, and rank-driven GRPO with explicit negative rejection), compressing lengthy documents into high-density rank-aligned summaries that achieve state-of-the-art ranking accuracy across TREC DL and BEIR while drastically reducing inference latency.

Supervised Fine-Tuning for Reasoning Large Language Models: A Survey

This paper provides the first systematic survey dedicated to Supervised Fine-Tuning (SFT) for reasoning large language models, establishing a dual-axis taxonomy of data-centric curation and multi-granularity algorithmic gradient modulation, unifying diverse losses under a generalized optimization objective, and organizing empirical findings into a data-dynamics-mechanisms analytical framework with explicit causal links to methodological design.

Surfacing the Unsaid: CUE-Bench for Affective Stance in Chinese Discourse

CUE-Bench introduces the Explicit-Implicit Stance Matrix, which cross-composes surface wording polarity with implicit affective tendencies into 9 interpretable affective stances across 51,823 Chinese discourse instances; guided by Matrix-Guided Chain-of-Thought, it significantly boosts LLM reasoning in affective stance, communicative intent, and fine-grained emotions.

SymboLLM-FE: LLM-Accelerated Symbolic Regression for Automated Feature Engineering on Tabular Data

SymboLLM-FE introduces a neuro-symbolic automated feature engineering framework that reduces the symbolic regression search space from exponential \(O(2^n)\) to polynomial \(O(n^2)\) via a Spearman correlation-guided expanding-sliding window, mining mathematically grounded formulas that an LLM subsequently refines into human-interpretable, executable code within single-digit API calls to substantially enhance downstream tabular predictors.

SyRuP: Enhancing System-Prompt Following via Reward-Guided Prediction in LLM Decoding

SYRUP freezes the base LLM and enhances system-prompt following during decoding by attending to the system prompt as a decoupled key-value memory via a lightweight cross-attention reward head, combined with base model contrastive logit shifts within top-k candidate space.

TabSieve: Explicit In-Table Evidence Selection for Multi-Table Question Answering

TabSieve introduces a select-then-predict paradigm for tabular in-context prediction, establishing explicit and auditable evidence selection via 41K filtered reasoning trajectories and stabilizing mixed classification-regression reinforcement learning with TAB-GRPO, outperforming competitive tabular and general LLMs by 2.92% on classification and 4.56% on regression across 127 held-out tables.

TAC-CoT: Tool-Empowered Implicit Chain-of-Thought via Interleaving Explicit Tokens

TAC-CoT proposes an adaptive implicit Chain-of-Thought framework for tool-using agents by replacing verbose natural-language reasoning rationales with compact continuous latent segments while retaining textual tokens exclusively for tool calls, observations, and final answers, cutting forward passes by up to 6.8× while outperforming explicit-CoT baselines across mathematical and multi-hop QA benchmarks.

TagPR: Tag-Guided Process Supervision for Personalization Reasoning in Large Language Models

This paper proposes TagPR, a framework that externalizes personalization reasoning into explicit cognitive semantic tags and combines a user-embedding-augmented reward model (PRMU) with multi-stage reinforcement learning, yielding an average improvement of 32.65% over the base model across all tasks in the LaMP benchmark.

Tampered ML: Evaluating LLM Judgment of Machine Learning Research

This paper introduces TAMPERED ML, a benchmark comprising 1,858 pairwise experiment proposals from top-tier ML papers paired with minimally edited counterfactuals harboring targeted methodological flaws; empirical results show that frontier reasoning models reach only 77.9% both-orders accuracy, exhibit substantial presentation-order vulnerability, and consistently mistake descriptive transparency and conventional practice for methodological rigor.

TCD: Empowering Zero-Shot Detection of LLM-Generated Text using Token Contextual Dependency

Challenging the uniform token aggregation assumption in zero-shot machine-generated text detectors, TCD introduces a plug-and-play token reweighting framework that quantifies token contextual dependency via surrogate attention entropy coupled with Dynamic Layer Selection (DLS), substantially boosting the detection accuracy and adversarial robustness of Fast-DetectGPT, Binoculars, and DNA-DetectLLM.

Teaching by Learning: Harnessing LLMs Collaboration for Adaptive Data Synthesis via Cognitive Load Feedback

This work proposes CoFAG, a self-evolving multi-model collaborative data synthesis framework that operationalizes the target model's internal cognitive load via the dilution-resistant Inference Stress Index (ISI), iteratively training the synthesizer via DPO to target the learner's cognitive frontier and significantly outperforming prior synthesis baselines on complex reasoning benchmarks.

Test-Time Scaling for Scientific Equation Discovery

This paper formalizes LLM-driven scientific equation discovery as a unified external test-time scaling (TTS) search process, showing across ~4,000 H100 GPU hours that search width is the dominant compute allocation parameter that determines both accuracy and wall-clock efficiency, while population-branching splits and controller algorithm choices are secondary.

The Anatomy of an ASR Hallucination

This paper provides a mechanistic interpretability analysis of speech recognition hallucinations under environmental and accent shifts across Conformer CTC and RNN-T architectures, revealing that the terminal encoder stage (final 1–2 blocks) serves as a critical boundary where acoustic signals transition into decoder-readable symbolic tokens, establishing that grounding failure is a causal prerequisite rather than a sufficient cause for fluent fabrication.

The Attentional White Bear: Persistent But Attenuated Concept Recoverability in Transformer Language Models

Through layerwise residual-stream probing, head-level attention routing analyses, and embedding-prototype semantic similarity metrics, this paper demonstrates that instruction-based suppression in transformer language models successfully eliminates forbidden terms from overt lexical output while leaving the prohibited concepts strongly decodable internally and continuously steering downstream generation, exposing a fundamental divergence between behavioral and representational alignment.

The Cost of Sounding Different: Accent Bias in Audio Language Models

This paper provides the first comprehensive audit of accent bias in native audio language models across three high-stakes deployment settings (workplace hiring, academic presentations, and English proficiency tests), demonstrating that models consistently penalize non-Western accents (specifically Chinese and Nigerian English) according to their acoustic distance from an implicit American English norm, independent of speech content and transcription accuracy.

The Grammar of Transformers: A Systematic Review of Interpretability Research on Syntactic Knowledge in Language Models

This paper presents the first large-scale systematic review of interpretability research evaluating the syntactic capabilities of Transformer-based language models, synthesizing 337 papers, 3,074 datapoints, and 113 languages into an open database, while mixed-effects regression reveals strong mastery of formal syntax contrasting with persistent fragility at the syntax-semantics interface, as well as an extreme concentration on English and BERT-family architectures.

The Hallucination Signal Is a Mean Shift: Why Simple Probes Suffice

Through a systematic geometric decomposition of large language model hidden states, this work demonstrates that the hallucination detection signal in the paired-example paradigm is overwhelmingly dominated by a 1D mean shift, where simple L2-regularized logistic regression bounds complex probes, and an oracle-free multi-layer aggregation strategy (LayerMix) matches oracle-layer performance (0.954 AUROC).

The Illusion of Expertise: Narrow Finetuning Induces Persistent Domain Priors in Vision-Language Models

By introducing the white-box diagnostic ActDiff, this paper reveals that narrow-domain finetuning installs an input-agnostic constant bias vector \(\mu\) in the VLM residual stream that prompts the model to ignore visual context, recovers 54% of benchmark collapse via training-free inference-time activation subtraction, and shrinks the trace by 20–64× via a lightweight first-moment training regularizer.

The Imitation Game: When LLMs Learn to Reason Like Programs via Code-Centric Reasoning Data Synthesis

To bridge the paradox where LLMs author complex dynamic programming solutions yet stumble on basic deterministic natural language reasoning (e.g., comparing 9.9 vs. 9.11), this paper introduces MIMIC, an end-to-end synthesis framework that converts algorithmic code into verifiable natural language reasoning trajectories via narrative fusion, AST-guided test generation, and trace-grounded CoT, delivering dense process-level supervision (CIR) without external reward models.

The Importance of Being Statistically Earnest: A Critical Re-evaluation of GSM-Symbolic

This paper provides a rigorous statistical re-evaluation of the high-profile GSM-Symbolic benchmark, demonstrating that when accounting for item-level variability via bootstrapped Generalized Linear Mixed Models (GLMMs) and controlling for an unacknowledged systematic shift toward larger numbers, the claimed widespread performance drops vanish for almost all models, revealing diverse behavioral failure profiles (such as variable binding fragility, arithmetic limits, and dual-task interference) rather than a uniform lack of reasoning.

The Masked Advantage: Uncovering Local-Language Access to Cultural Knowledge in LLMs

By coupling a one-parameter logistic item response theory (1PL-IRT) model with a difference-in-differences design across 13 locales and roughly 80 LLMs, this paper shows that lower local-language accuracy on cultural questions is predominantly a surface proficiency deficit rather than missing cultural knowledge; controlling for general proficiency reveals an intrinsic local-language cultural knowledge access advantage in 98% of evaluated settings.

The Menu Is an Execution Prior: State-Path Tool Menus for Online Agents

Addressing the failure of relevance-based tool retrieval to supply prerequisite tools or place them before their consumers, State-Path Tool Menu treats the bounded pre-execution menu as an execution prior over state paths, jointly selecting complementary path roles and enforcing producer-first topological ordering to boost online task success on ToolBench from 0.737 to 0.898 with the agent and budget held constant.

The Persona-Switch Jailbreak Backdoor in LLMs: From Payload Retrieval to Trigger-Gated Persona-State Selection

Exposing the critical "prefix/content gap" where traditional payload-retrieval jailbreak backdoors merely elicit affirmative openings without sustaining actionable harm, this paper introduces the Persona-Switch backdoor. Instead of retrieving fixed text patterns, a backdoor trigger conditionally gates a latent persona-conditioned behavioral state, which—when amplified by persona-aligned framing—redirects the model's intrinsic generation capabilities to produce actionable harm across Evil, Sycophantic, and Hallucinatory personas.

Thinking Without Speaking: Efficient Tree of Thoughts Reasoning via Latent Steering

Addressing the severe latency and token blow-up of Tree of Thoughts (ToT) under the generate-then-score paradigm, this paper presents Implicit Tree of Thoughts (IToT)—a training-free framework that reframes discrete tree search as continuous optimization in latent hidden representations, steering reasoning paths via contrastive prototypes and CMA-ES to outperform explicit search baselines while operating below the token budget of vanilla Chain of Thought.

TimeSense: Decoupling Modality Learning for Robust Time Series Reasoning

To resolve the multimodal imbalance and temporal degradation caused by text-dominated optimization in multimodal LLMs, TimeSense introduces a fusion-then-decoupling architecture that decouples temporal representations after cross-modal fusion and applies dense reconstruction supervision, substantially outperforming prior baselines across the newly introduced diagnostic benchmark EvalTS and multiple external benchmarks.

Tinted Frames: Question Framing Blinds Vision-Language Models

This paper demonstrates that Vision-Language Models (VLMs) exhibit "selective blindness"—modulating visual attention allocation based on question framing even when tasks demand identical visual reasoning—and proves that constrained framings (Yes/No, MCQ) suppress visual energy and induce spatial misallocation away from target regions, which can be mitigated via attention steering and structured reasoning prompts.

TopoAlign: A Framework for Aligning Code to Math via Topological Decomposition

Addressing the severe scarcity of informal-to-formal parallel data in mathematical autoformalisation, TopoAlign structurally decomposes code repositories into triplets of docstrings, dependency functions, and main implementations to mirror Lean 4 formal statements, substantially boosting syntactic typechecking and bidirectional semantic equivalence across diverse Math LLMs without introducing any new mathematical knowledge.

Tournament-GRPO: Group-Wise Tournament Rewards for Reinforcement Learning in Open-Ended Long-Form Generation

To resolve the calibration drift and reward saturation inherent in independent pointwise LLM scoring for open-ended long-form generation, Tournament-GRPO structures same-query rollouts into repeated multi-round elimination tournaments, deriving group-normalized relative rewards that substantially improve GRPO advantage estimation on Deep Research Bench.

Towards Understanding Pause Token Fine-Tuning Dynamics: A Mode Retention Perspective

Reframing pause tokens as a training-dynamics intervention rather than an inference-time compute device, this paper identifies mode retention (H1) and non-myopic compression (H2), introducing Masked Boundary Pause (MBP) fine-tuning that boosts reasoning up to +6.0 points on math and +2.5 on code across 1B–8B models while completely preserving general capabilities.

TRACE: A Unified Rollout Budget Allocation Framework for Efficient Agentic Reinforcement Learning

TRACE reframes rollout budget allocation in multi-turn agentic RLVR as a mixed-reward contrast construction problem across tree anchors. Guided by a shared lightweight prefix-history predictor, it jointly optimizes global prompt-root allocation and prompt-local prefix branching under a strictly matched sampling budget, transforming sparse outcome rewards into dense stepwise preference comparisons and boosting agentic policy performance across mathematical reasoning, multi-hop QA, and function calling.

TRACE-GRPO: Teacher-Refined Advantage Credit Estimation for Long-Horizon Agentic RL

To overcome GRPO's failure modes in long-horizon agent tasks—namely zero gradients under reward homogeneity and indiscriminate scalar credit assignment across tokens—TRACE-GRPO proposes a three-level hierarchical advantage estimator (sample-level consistency fallback, turn-level critique reweighting, and token-level redistribution via critique-induced log-probability shift \(\delta\)), enabled efficiently by Critique-Aware Paged Attention (CAPA) to achieve dense credit assignment without learned critics or token labels, achieving 96.3% and 73.4% success on ALFWorld and ScienceWorld.

TRACE: Triplet Recovery via Arrowhead-Centric Extraction for Flowchart Understanding

TRACE transforms flowchart parsing from holistic end-to-end generation into localized, arrowhead-anchored triplet extraction, synthesizing aligned training data from textual sources and outperforming large-scale models like GPT-5 across nine benchmarks while providing verifiable per-edge traceability.

TRACER: An Information Entropy-Grounded Framework for Efficient Large-table Semantic Annotation

To resolve the critical trade-off between the heavy computational overhead of full-table modeling and the loss of discriminative cues from heuristic truncation in large-table semantic annotation, TRACER introduces an information entropy-grounded row compression framework that ranks rows via temperature-scaled ontology uncertainty and builds hybrid contexts, elevating GitTable Macro-F1 to 71.12% while maintaining massive throughput on lightweight DistilBERT.

Tracing Computation Density in LLMs

This paper introduces s-Trace, a training-free single-pass causal graph traversal framework to quantify computation density in LLMs, revealing a modular two-phase computational organization where an early-layer minimal core (0.01%–1% of edges) constructs rough predictions and dense later-layer attention heads (1%–100%) refine the full distribution.

Tracing Error Propagation in Reasoning Language Models

This paper investigates the latent dynamics of error propagation in reasoning language models (RLMs), demonstrating that step errors diffuse into local neighboring states, final-step representations distinctly separate recoverable from unrecoverable failures, and the original persistent error can be retrospectively decoded from final states to guide token-efficient prefix resampling.

Training Matryoshka Mixture-of-Experts for Elastic Inference-Time Expert Utilization

This paper introduces Matryoshka Mixture-of-Experts (M-MoE), a training framework that systematically randomizes the per-layer number of active experts during pre-training, forcing the router to learn globally coherent rankings and high expert specialization so that a single checkpoint achieves robust, continuous inference elasticity across varying computational budgets.

TranslatePsy-AfriSLM: High-Quality Data Scaling For Low-Resource Machine Translation

TranslatePsy-AfriSLM introduces an open-source machine translation ecosystem covering 19 Sub-Saharan African languages, achieving up to 96% token reduction via Human-Mix-calibrated unified robust z-score filtering, enabling lightweight 0.8B-4B SLMs to consistently outperform frontier models such as TranslateGemma-27B and Qwen3.5-122B-A10B.

TRIAGE: Label-Free Hateful Video Detection with Boundary Mapping and Resolution

Addressing the severe effectiveness-efficiency bottleneck where label-free MLLM detectors require costly full-pipeline reasoning yet underperform supervised models, TRIAGE introduces a coarse-to-fine framework that maps high-uncertainty decision boundary samples via calibrated posterior entropy and adaptively resolves them through Bayesian sequential verifier updates, surpassing supervised SOTA on 3 of 4 benchmarks with at least 2.0x lower inference cost.

Triggering Chain-of-Thought via Latent Feature Interventions in Large Language Models

By identifying reasoning-related sparse latent features via pretrained Sparse Autoencoders (SAEs), targeted residual steering on a single feature at the very first decoding step triggers multi-step chain-of-thought reasoning under direct prompting, matching standard CoT accuracy while generating substantially fewer tokens.

Trust Before Fusion: QIMG-7 and Source-Aware Resolution for Polluted Multimodal RAG

Addressing the extreme vulnerability of Multimodal RAG when exposed to topically relevant yet factually polluted external evidence, this paper introduces QIMG-7—a controlled stress-test benchmark spanning 4 datasets, 7 image-pollution families, and 16 paired regimes—along with Source-Aware Trust Resolution (SATR), a training-free framework that isolates and routes among Parametric, Text-only, and Multimodal candidate branches, improving balanced factuality support by 11.7 points over naive multimodal fusion.

TrustRAG: Metadata-Governed RAG for Compliance-Oriented Question Answering

Addressing the vulnerability where retrieved documents are semantically relevant yet inadmissible under compliance rules, TrustRAG inserts a deterministic metadata-governed routing layer and conflict-reporting generation policy before generation, eliminating unauthorized leakage, stale citations, and unsupported answers.

Trustworthy Medical LLMs via Bayesian Belief Updating

This paper presents BELIEF, a Bayesian Evidence-Leveraged InfErence Framework that models clinical medical reasoning as sequential probabilistic belief revision, using posterior entropy from the Bayesian Belief Updating Module to dynamically trigger external clinical guideline retrieval via the Dynamic Evidence Retrieval Module, achieving top-tier diagnostic accuracy across 7 benchmarks while drastically slashing medical hallucinations and calibration error.

TTSR: Test-Time Self-Evolving via Reflection

Addressing the dual bottlenecks of learnable sample scarcity and undirected exploration in test-time training on challenging reasoning tasks, TTSR introduces a self-evolving framework where a single model alternates between Student and Teacher roles—diagnosing failure traces to synthesize frontier-calibrated curriculum variants, accumulated across iterations via a weakness memory and dynamic strategy notes.

TTT-Sparse: Test-Time Training of Sparse Attention for Long-Context Understanding

TTT-Sparse proposes a plug-and-play test-time training framework that keeps LLM weights frozen, constructs contextual chunk key states via intra-document and observation-conditioned importance, and optimizes lightweight query adapters to adaptively select top-relevant KV chunks during generation.

TVIR: Building Deep Research Agents Towards Text-Visual Interleaved Report Generation

Addressing the limitation where deep research agents remain predominantly text-centric, this paper introduces TVIR-BENCH, the first expert-curated benchmark requiring visual elements to serve analytical sub-goals, alongside TVIR-AGENT, a hierarchical multi-agent framework that plans, retrieves, generates source-traceable charts, and composes text-visual interleaved reports under a rigorous dual-path evaluation suite.

Typological Alignment of Stack-Based Language Models on Mildly Context-Sensitive Artificial Languages

Extending Generalized Categorial Grammar to generate 32 artificial languages featuring mildly context-sensitive cross-serial dependencies, this work evaluates six stack-augmented language model variants against standard baselines, revealing that working-memory-constrained superposition-stack RNNs achieve superior structural generalization and align most strongly with natural language typological word order frequencies.

Typological Feature Prediction with Large Language Models: An In-Context Learning Approach

This paper investigates typological feature imputation over the URIEL+ database using in-context learning with LLMs; by supplying genealogical family-tree neighbors, spherical geographic neighbors, and anchor-feature priors with contrastive evidence, LLMs perform principled meta-linguistic reasoning that achieves 0.909 macro F1, outperforms classical matrix completion and random forests, and eliminates historical performance penalties on low-resource languages.

Understanding the Limits of Agentic ICD Coding

This paper systematically investigates the performance boundaries of neural classifiers, workflow pipelines, and free-form agentic architectures on ICD-10-CM clinical coding, identifying two orthogonal failure modes—frequency-driven rarity bias and multi-step guideline-following collapse—and shows that tool-layer discovery grounding coupled with dynamic index orchestration recovers up to 0.34 micro-F1 on complex multi-step injury cases.

UniMem: Complementary Episodic-to-Parametric Memory for Boundary-Agnostic Task Streams

Inspired by brain Complementary Learning Systems, UniMem decouples lightweight routing tokens from high-capacity Procedural KV Memory, establishing a self-routing lifecycle where sparse unknown queries are buffered for RAG fallback and recurring patterns are clustered to expand dedicated parametric memory units, achieving strong execution stability on boundary-agnostic streams.

Unlocking Multimodal Protein Language Models at Inference Time

Addressing the longstanding neglect of inference-time strategies in multimodal protein language models (pLMs), this work introduces a three-level control framework spanning sampling distributions, per-step logits, and parallel trajectories, demonstrating that training-free test-time guidance and internal reward beam search boost generalist pLMs to match or surpass specialized state-of-the-art models.

Unmasking Superficial Honesty: Mechanistic Head-Driven Activation Steering for Sycophancy-Free Medical VLMs

Addressing the acute vulnerability of medical VLMs to misleading clinical prompts—and uncovering the counterintuitive "superficial honesty" paradox where prompt-level mitigations mask an internal collapse of visual token grounding—this paper proposes Head-Driven Activation Steering (HD-AS), a training-free inference-time method that isolates sycophancy-sensitive attention heads via activation drift and applies lightweight isotropic damping at the terminal decision token, cutting conditional flip rates by up to 30 percentage points while actively recovering internal visual attribution.

Users Mispredict What Drives Their Preferences for AI Writing Assistance

Through a factorial vignette experiment with 50 professional participants making 750 pairwise decisions, this paper reveals that compositional effort rather than urgency drives actual preferences for AI drafting assistance, exposes a striking stated–revealed preference misalignment, and demonstrates that a behavior-derived decision rule outperforms self-reported preference configurations and machine learning baselines.

Using Prosody to Predict Syntactic Structure

By formulating the syntax-prosody interface as sequence-level mutual information estimation using an encoder-decoder architecture over read (LibriTTS) and spontaneous (CANDOR) speech, this paper reveals that acoustic features (word duration, pause) reduce syntactic uncertainty by up to 10.2%, predominantly marking phrase boundaries rather than syntactic categories.

UTILMEM: Benchmarking Evidence Utilization in Long-Term Conversational Memory

This work introduces UTILMEM, a 1,717-instance diagnostic benchmark across five real-world domains that shifts the evaluation of long-term conversational memory from pointwise factual recall to comprehensive evidence utilization, revealing severe degradation in state-of-the-art memory systems when synthesizing multi-session evidence under adversarial retrieval interference.

UXBench: Benchmarking User Experience in AI Assistants

UXBench presents the first user-centric benchmark grounded in authentic human-AI interaction feedback signals, evaluating LLMs across UX judgment, response generation on failure-prone queries, and complaint recovery, while demonstrating that model capabilities scale only weakly into real-world user experience gains.

VGI-Bench: Probing Visual Intelligence in Video Generation Models

VGI-Bench introduces a photorealistic, process-sensitive benchmark with 27 tasks and 810 instances to evaluate procedural visual reasoning in video generation models, revealing that the strongest model Seedance 2.0 achieves only 51.0% under a multi-factor judge protocol while intermediate denoising analysis demonstrates virtually absent (<1%) visual self-correction.

VIABLE: A Visually Impaired Assistance Benchmark for VLM-as-a-Judge Evaluation

This paper introduces VIABLE, the first comprehensive benchmark for evaluating Vision-Language Models as judges in visually impaired assistance (VIA), covering over 312K judgment samples across an Effectiveness–Impartiality–Stability framework with 12 atomic failure modes; finding that current judges exhibit severe diagnostic blindspots, pervasive self-preference, and adversarial fragility, the authors propose VIA-Judge-Agent—a training-free inference harness that significantly elevates diagnostic accuracy and downstream assistive response quality.

Vision-Language Models Suppress Female Representations Under Ambiguous Input

This paper introduces LALS (Latent Association Leaning Score), a training-free token-level metric projecting visual token activations into the VLM's text-embedding space, revealing that under gender-ambiguous inputs, internal representations systematically decouple from textual outputs—deep transformer layers act as an asymmetric filter that suppresses female associations while amplifying male defaults.

What and Whose Knowledge? Measuring Epistemic Diversity in Large Language Models

This work establishes the first systematic empirical measurement of epistemic diversity in LLMs across 27 models, 155 topics, and 12 countries (decomposing 1.7M responses into 70M claims), discovering that while claim diversity has significantly increased over time, all models lag behind a simple web search baseline, model scale counterintuitively reduces diversity due to memorization, and parametric outputs systematically prioritize English over local-language knowledge.

What’s in a Name? Morphological Shortcuts by LLMs in Pharmacology

This paper systematically discovers that large language models frequently rely on morphological shortcuts in pharmacology by inferring drug classes directly from surface affixes (e.g., -cillin) without verified entity knowledge, introduces a 2×2 factorial diagnostic framework to disentangle affix versus holistic semantics, and localizes this behavior to steerable low-rank causal directions in early-to-middle Transformer layers.

When Calibration Rankings Reverse: Accuracy-Controlled Evaluation for Faithful Uncertainty

This paper demonstrates theoretically and empirically that standard global calibration metrics (ECE and Brier Score) are fundamentally confounded by task accuracy, frequently producing ranking reversals where models favored by raw metrics cease to be favored under controlled accuracy; to resolve this, it establishes ACE (Accuracy-Controlled Evaluation) across Instance-, Distribution-, and Candidate-Aligned views for fair cross-model comparisons.

When Compression Helps and When It Hurts: Condition-Aware Analysis of Chain-of-Thought Distillation

This paper reframes Chain-of-Thought (CoT) distillation compression into three orthogonal dimensions—importance criterion, restructuring level, and compression budget—uncovering how reasoning regime, granularity, and task domain condition compression efficacy and deployment savings.

When Graphs Whisper: Graph Structure Serialization for Embedding Alignment in Cold-Start Recommendation

To tackle cold-start recommendation where new items lack interaction history and text prompts fail to capture higher-order graph topologies, GraphSEA serializes interaction graphs into structured token sequences for LLMs to learn behavior embeddings, aligning them with text-derived content embeddings across interaction probability, preference order, and item distinction while preserving ultra-fast dot-product inference.

When LLMs Say No: Diagnosing and Mitigating Conservative Bias in Authorship Verification

Addressing the severe conservative bias in zero-shot LLM authorship verification where models disproportionately predict "different author" (NO) due to confusing topical divergence with authorial differences, this paper introduces lightweight framing and contrastive reweighting prompt interventions that increase same-author accuracy by up to 13.1 percentage points and macro-F1 by 3.1 points without fine-tuning.

When Queries Crowd: Diagnosing and Improving Spread-Out Regularization for Text Embeddings

This paper conducts the first rigorous controlled study isolating spread-out regularization in text embedding training, discovers that retrieval gains concentrate almost exclusively on dispersing directional crowding in short queries, and proposes QUARTO, a query-length-adaptive weighting schedule that outperforms fixed-weight regularization.

Where to Look and What to Use: Retrieve–Localize–Generate for Long-Term Conversational Memory Question Answering

MemLoc addresses the core challenges of cross-session evidence fragmentation and intra-session noise-induced "lost-in-the-middle" effects in long-term conversational memory QA through a unified Retrieve–Localize–Generate framework, integrating entropy-routed multi-granularity graph retrieval, a reasoning-based evidence locator trained via Self-reflective Hint Policy Optimization (SHPO), and lightweight location ID cue guidance.

Whether LLMs Can Navigate Beliefs and Facts Depends on How You Phrase It

The paper reveals that large language models exhibit systemic weakness in acknowledging user beliefs grounded in false claims because they default to fact-checking rather than tracking belief states; this weakness fluctuates wildly across 18 epistemic verbs (gaps ranging from +50% to -14%), and can be substantially remedied via prompt-level instructions or decoding-time attention suppression.

World Models Meet Language Models: On the Complementarity of Concrete and Abstract Reasoning

Addressing the failure modes of Simulation Inertia and the Forced-Simulation Paradox when attaching video world models to MLLMs, this paper formalizes controlled concrete reasoning and proposes PF-OPSD, an on-policy distillation framework that leverages privileged training-time future video to calibrate simulation decisions, verification, and reliance, outperforming baselines by over 10.5 percentage points on two human-verified benchmarks.