arXivSub Start free trial

ACL 2026 Papers with Code β€” Page 5

Annual Meeting of the Association for Computational Linguistics Β· 557 papers

PubMed Reasoner: Dynamic Reasoning-based Retrieval for Evidence-Grounded Biomedical Question Answering

Yiqing Zhang (PayPal), Fabricio Murai (Worcester Polytechnic Institute)

CodeRetrievalExplainability and InterpretabilityComputational EfficiencyDrug DiscoveryTransformerLarge Language ModelPrompt EngineeringTextBiomedical DataRetrieval-Augmented GenerationChain-of-Thought

🎯 What it does: Constructed PubMed Reasoner, a three-stage biomedical question-answering agent that includes self-critical query optimization, batch reflective retrieval with early stopping, and evidence-based answer generation.

Putting HUMANS first: Efficient LAM Evaluation with Human Preference Alignment

Woody Haosheng Gan (University Of Southern California), Diyi Yang (Stanford University)

CodeRecommendation SystemComputational EfficiencyData-Centric LearningContrastive LearningTextBenchmarkAudio

🎯 What it does: Systematically study subset selection for evaluating large audio models (LAM), constructing a minimal evaluation set aligned with human preferences called HUMANS, and publicly releasing the data and regression models.

QuDAR: Query-Wise Dual-Perspective Adaptive Retrieval

Joeun Kim (Korea Advanced Institute of Science and Technology), Jae-Gil Lee (Korea Advanced Institute of Science and Technology)

CodeRetrievalTransformerLarge Language ModelContrastive LearningTextBenchmarkRetrieval-Augmented Generation

🎯 What it does: Propose a Query-wise Dual-Perspective Adaptive Retrieval (QUDAR) framework that dynamically fuses four retrieval signals: sparse retrieval, dense retrieval, original query, and expanded query;

Query-Efficient Agentic Graph Extraction Attacks on GraphRAG Systems

Shuhua Yang (Pennsylvania State University), Suhang Wang (Pennsylvania State University)

CodeSafty and PrivacyAdversarial AttackGraph Neural NetworkTransformerLarge Language ModelAgentic AIPrompt EngineeringTextGraphAgriculture RelatedRetrieval-Augmented Generation

🎯 What it does: This paper investigates privacy leakage in the GraphRAG system under black-box query budget constraints, and proposes an efficient query-based attack framework called AGEA.

Question Tells You Where the Answer Is: Intention-aware Long-Context KV Cache Compression

Liang Zhao (Harbin Institute of Technology), Bing Qin (Harbin Institute of Technology)

CodeCompressionComputational EfficiencyTransformerLarge Language ModelText

🎯 What it does: Propose an intent-based KV cache compression method called IntentKV, which identifies and retains critical KV pairs for subsequent generation by leveraging the attention distribution differences of intent tokens.

R1-RE: Cross-Domain Relation Extraction with RLVR

Runpeng Dai (University of Maryland), Hongtu Zhu (University of Maryland)

CodeDomain AdaptationReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextGraphChain-of-Thought

🎯 What it does: Re-defining the relation extraction task as a reasoning process based on annotation guidelines, and achieving cross-domain reasoning by training a small LLM with RLVR (Reinforcement Learning with Verifiable Rewards).

Ranking Reasoning LLMs under Test-Time Scaling

Mohsen Hariri (Case Western Reserve University), Vipin Chaudhary (Case Western Reserve University)

CodeTransformerLarge Language ModelTextBenchmark

🎯 What it does: This paper studies the stability and convergence of ranking large language models (LLMs) under test-time scaling conditions, and systematically compares various statistical ranking methods.

Re^3: Relevance & Recency Retrieval for Mitigating Temporal Hallucination

Jiawei Cao (University Of Science And Technology Of China), Shijin Wang (University Of Science And Technology Of China)

CodeGenerationData SynthesisRetrievalTransformerLarge Language ModelContrastive LearningTextTabularTime SeriesBenchmarkRetrieval-Augmented Generation

🎯 What it does: Proposes the Re3 framework, which reduces temporal illusion in retrieval-augmented generation through time-aware dual correlation encoding and conflict-aware temporal filtering.

Ready Jurist One: Benchmarking Language Agents for Legal Intelligence in Dynamic Environments

Zheng Jia (Fudan University), Zhongyu Wei

CodeTransformerLarge Language ModelAgentic AIPrompt EngineeringMixture of ExpertsTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought

🎯 What it does: Constructed the first dynamic interactive legal environment J1-ENVS and a dual-indicator evaluation framework J1-EVAL to systematically evaluate LLM-driven legal intelligent agents.

Reasoning Gets Harder for LLMs Inside A Dialogue

Ivan KartÑč (Charles University), Ondrej Dusek (Charles University)

CodeTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought

🎯 What it does: Built a dynamic benchmark called BOULDER to evaluate the reasoning capabilities of large language models (LLMs) in task-oriented dialogue (TOD) environments, and compared the performance of reasoning tasks conducted in isolation versus those embedded within dialogues.

Reasoning Hijacking: The Fragility of Reasoning Alignment in Large Language Models

Yuansen Liu (National University of Singapore), Anthony Kum Hoe Tung (National University of Singapore)

CodeClassificationAdversarial AttackTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought

🎯 What it does: Proposes the 'reasoning hijacking' attack, which deceives the LLM's reasoning process by injecting fake decision criteria without altering the task objective.

Reasoning Structure Matters for Safety Alignment of Reasoning Models

Yeonjun In (Korea Advanced Institute of Science and Technology), Chanyoung Park (Korea Advanced Institute of Science and Technology)

CodeSafty and PrivacyExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought

🎯 What it does: Propose ALTTRAIN by rewriting the inference structure of large reasoning models to achieve safe alignment

Reasoning While Asking: Transforming Reasoning Large Language Models from Passive Solvers to Proactive Inquirers

Xin Chen (Nanjing University), Shujian Huang (Nanjing University)

CodeExplainability and InterpretabilityComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought

🎯 What it does: Propose the Proactive Interactive Reasoning (PIR) framework, enabling large language models to actively ask users questions during the reasoning process and integrate feedback, overcoming the problem of blind self-thinking.

ReCode: Reinforcing Code Generation with Reasoning-Process Rewards

Lishui Fan (Zhejiang University), Zhongxin Liu (Zhejiang University)

CodeOptimizationAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringContrastive LearningTextBenchmarkChain-of-Thought

🎯 What it does: Proposes the ReCode framework, which combines reinforcement learning to supervise and optimize the reasoning quality during the code generation process.

Rectifying the Emotional Flow: Aligning Priors and Dynamic Guidance for High-Arousal Text-to-Speech

Fangming Feng (Zhejiang University), Tao Jin (Zhejiang University)

CodeGenerationTransformerDiffusion modelFlow-based ModelRectified FlowTextAudio

🎯 What it does: A model-free retraining framework is proposed by introducing an emotion correction noise prior (ERNP) and likelihood inverse guided (LIG) dynamic guidance during the inference stage, aiming to address the stability and expressiveness challenges in high-intensity emotion text-to-speech.

ReEfBench: Quantifying the Reasoning Efficiency of LLMs

Zhizhang Fu (Westlake University), Yue Zhang (Westlake University)

CodeExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought

🎯 What it does: Evaluated the reasoning efficiency of LLMs and constructed a neuro-symbolic evaluation framework based on First-Order Logic called ReEfBench.

Reframing Responsibility: Framing-Aware Event Causality Identification

Jin Zhao (Brandeis University), Nianwen Xue (Brandeis University)

CodeRecognitionTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextChain-of-Thought

🎯 What it does: Proposed and implemented the Frame-based Event Causality Identification (FRECI) task, constructed a multilingual framed causality dataset, and designed a joint neural model to extract information such as responsibility, source, and modality.

REG: Retrieval via Emotion Similarity for Guiding Empathetic Dialogue Generation

Xu Wang (Tianjin University), Yuexian Hou (China Mobile Communication Group Tianjin Co., Ltd.)

CodeGenerationRetrievalKnowledge DistillationTransformerContrastive LearningTextRetrieval-Augmented Generation

🎯 What it does: Propose the REG framework, which achieves more empathetic dialogue generation by guiding retrieval with emotional attributes and combining it with a generator.

Reinforced Efficient Reasoning via Semantically Diverse Exploration

Ziqi Zhao (Shandong University), Xin Xin (Shandong University)

CodeExplainability and InterpretabilityComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringMixture of ExpertsTextBenchmarkChain-of-Thought

🎯 What it does: Proposes the ROSE framework, which improves the accuracy and reasoning efficiency of LLMs in mathematical reasoning tasks through semantic diversification exploration and length-aware advantage estimation.

Reinforcement Learning for Self-Improving Agent with Skill Library

Jiongxiao Wang (University of Wisconsin Madison), Lin Lee Cheong (Aws Agentic Ai)

CodeAutonomous DrivingOptimizationAI Code AssistantTransformerSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextSequentialRetrieval-Augmented GenerationChain-of-Thought

🎯 What it does: This paper proposes a self-improving agent framework called SAGE based on reinforcement learning, which utilizes a skill library to achieve knowledge transfer and reuse;

Render-of-Thought: Rendering Textual Chain-of-Thought as Images for Visual Latent Reasoning

Yifan Wang (Tencent), Yang Tang (Tencent)

CodeExplainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodalityBenchmarkChain-of-Thought

🎯 What it does: Render the textual steps in the Chain-of-Thought (CoT) reasoning process into images and perform reasoning in the visual space, thereby achieving compression and visualization of the thought chain.

ReportLogic: Evaluating Logical Quality in Deep Research Reports

Jujia Zhao (Leiden University), Zhaochun Ren (Leiden University)

CodeKnowledge DistillationTransformerReinforcement LearningPrompt EngineeringTextBenchmark

🎯 What it does: Propose the ReportLogic benchmark to evaluate the logical quality of deep research reports and construct human-annotated data

Representation Interventions Enable Lifelong Knowledge Memory Control in LLMs

Xuyuan Liu (Dartmouth College), Zhengzhang Chen (NEC Laboratories America)

CodeComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringContrastive LearningText

🎯 What it does: Achieving lifelong knowledge updating and control in large language models through controllable interventions in the model's hidden layer space.

ReRec: Reasoning-Augmented LLM-based Recommendation Assistant via Reinforcement Fine-tuning

Jiani Huang (Hong Kong Polytechnic University), Li Qing (Hong Kong Polytechnic University)

CodeRecommendation SystemReinforcement Learning from Human FeedbackGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningContrastive LearningTextGraphTabularRetrieval-Augmented GenerationChain-of-Thought

🎯 What it does: Propose a ReRec LLM recommendation assistant based on reinforcement learning fine-tuning, which can perform multi-step reasoning and provide recommendation results in single-round complex queries;

Response-G1: Explicit Scene Graph Modeling for Proactive Streaming Video Understanding

Ke Ma (Northwestern Polytechnical University), Yunhao Liu (Tsinghua University)

CodeRetrievalExplainability and InterpretabilityComputational EfficiencyTransformerPrompt EngineeringVision Language ModelVideoTextBenchmarkRetrieval-Augmented Generation

🎯 What it does: Propose the Response-G1 framework to achieve active streaming video understanding by explicitly modeling and retrieving scene graphs to determine when to answer.

REST: Stress Testing Large Reasoning Models by Asking Multiple Problems at Once

Zhuoshi Pan (Tsinghua University), Lijun Wu (Shanghai Artificial Intelligence Laboratory)

CodeExplainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought

🎯 What it does: Proposed the REST framework, which conducts parallel multi-question stress testing on large-scale reasoning models (LRM) by concatenating multiple questions into a single prompt;

Rethinking Data Mixing from the Perspective of Large Language Models

Yuanjian Xu (Hong Kong University of Science and Technology (Guangzhou)), Guang Zhang (Hong Kong University of Science and Technology (Guangzhou))

CodeDomain AdaptationComputational EfficiencyData-Centric LearningTransformerLarge Language ModelContrastive LearningText

🎯 What it does: Proposes DoGraph, a dynamic domain reweighting framework based on graph constraints, which dynamically adjusts domain sampling weights during the pre-training of large-scale language models according to the model's own gradients;

Rethinking Depression Prediction from a Fine-Grained Subscore Modeling Perspective via Multi-Task Learning

Zhenguang Wang (Northeastern University), Osmar Zaiane (University of Alberta)

CodeClassificationExplainability and InterpretabilityRepresentation LearningGraph Neural NetworkTransformerSupervised Fine-TuningContrastive LearningTextBiomedical DataElectronic Health Records

🎯 What it does: Propose the MTSP (Multi-Task Subscore Prediction) framework, which uses multi-task learning to simultaneously predict subscale scores of depression scales from clinical interview texts, achieving fine-grained depression assessment.

Rethinking Jailbreak Detection of Large Vision Language Models with Representational Contrastive Scoring

Peichun Hua (Washington University in St. Louis), Ning Zhang (Washington University in St. Louis)

CodeAnomaly DetectionSafty and PrivacyComputational EfficiencyRepresentation LearningTransformerVision Language ModelContrastive LearningMultimodalityBenchmark

🎯 What it does: Proposes a contrastive scoring framework (Representational Contrastive Scoring, RCS) that leverages internal representations of large vision-language models to detect multi-modal 'jailbreak' attacks, balancing generalization and efficiency;

Rethinking the Idiomaticity Decomposability Hypothesis: Evidence from Distributional Learning

Maggie Mi (University of Sheffield), Nafise Sadat Moosavi (University of Sheffield)

CodeExplainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelContrastive LearningText

🎯 What it does: Investigated the decomposability of idioms, re-examining the IDH hypothesis from a distributed learning perspective using contextualized language models.

Revisit What You See: Revealing Visual Semantics in Vision Tokens to Guide LVLM Decoding

Beomsik Cho (Yonsei University), Jaehyung Kim (Yonsei University)

CodeExplainability and InterpretabilityComputational EfficiencyTransformerPrompt EngineeringMixture of ExpertsVision Language ModelImageTextMultimodalityBenchmark

🎯 What it does: This paper proposes a decoding strategy called ReVisiT that does not require additional training, leveraging semantic information from visual tokens to guide large vision-language models in generating more accurate and less hallucinatory text.

Revisiting Audio-language Pretraining for Learning General-purpose Audio Representation

Wei-Cheng Tseng (University of Texas at Austin), Dong Yu (Tencent AI Lab Seattle)

CodeClassificationRecognitionRetrievalRepresentation LearningTransformerVision Language ModelContrastive LearningTextMultimodalityAudio

🎯 What it does: This paper constructs CaptionStew by aggregating 10.7M multi-source audio-text pairs, and systematically compares the performance of contrastive learning and captioning as two pre-training objectives on speech, music, and environmental sound tasks.

Revisiting Evaluation of Question Answering Systems in Low-Resource Indic Languages: Bridging Human and Metric Alignment

Anuj Kumar (Indian Institute of Technology Jammu), Virendra Singh (Indian Institute of Technology Bombay)

CodeRecommendation SystemData-Centric LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextMultimodalityBenchmarkRetrieval-Augmented Generation

🎯 What it does: This paper proposes and verifies a multi-dimensional evaluation metric called LRM²QAS for evaluating question-answering systems in low-resource Indian languages.

Revitalizing Black-Box Interpretability: Actionable Interpretability for LLMs via Proxy Models

Junhao Liu (Peking University), Xin Zhang (Peking University)

CodeOptimizationExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringContrastive LearningText

🎯 What it does: Propose a Screen-and-Apply framework based on a proxy model, which generates high-fidelity post-hoc explanations on large models using cost-effective proxy LLMs, and verifies its operability in practical optimization tasks such as prompt compression and detoxification examples.

RL-PLUS: Countering Capability Boundary Collapse of LLMs in Reinforcement Learning with Hybrid-policy Optimization

Yihong Dong (Peking University), Ge Li (Alibaba Group)

CodeOptimizationAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringText

🎯 What it does: Propose RL-PLUS, a hybrid strategy optimization framework that integrates internal exploration with external data, aimed at enhancing the reasoning capabilities of large language models (LLMs) in reinforcement learning with human feedback (RLVR), and addressing the capability boundary collapse problem.

Robust Membership Inference for Large Language Models under Adversarial Generative Corruption

Yuanhong Huang (Beijing University of Posts and Telecommunications), Tao Qi (Beijing University of Posts and Telecommunications)

CodeSafty and PrivacyExplainability and InterpretabilityAdversarial AttackTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsTextRetrieval-Augmented Generation

🎯 What it does: Studied the interference caused by AI-generated text on membership inference attacks (MIA) in large language models, and proposed a Mixture-of-Experts framework called MoMIA to enhance membership inference robustness under adversarial generated text.

ROSE: An Intent-Centered Evaluation Metric for NL2SQL

Wenqi Pei (Hong Kong University of Science and Technology), Yuyu Luo (Hong Kong University of Science and Technology)

CodeAI Code AssistantTransformerLarge Language ModelPrompt EngineeringGenerative Adversarial NetworkTextTabularBenchmark

🎯 What it does: This paper proposes ROSE, an intent-oriented NL2SQL evaluation metric, which determines the semantic correctness of predicted SQL through an adversarial Prover-Refuter cascade.

RouteMoA: Dynamic Routing without Pre-Inference Boosts Efficient Mixture-of-Agents

Jize Wang (Shanghai Jiao Tong University), Dacheng Tao

CodeComputational EfficiencyAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringMixture of ExpertsText

🎯 What it does: The paper proposes RouteMoA, an efficient hybrid agent framework that selects high-potential LLMs and performs multi-round collaborative reasoning without executing full inference through dynamic routing.

RSDA: Restoring Stale Data Affinity via Dynamic Renovation Strategy for Mitigating Data Scarcity

Yidan Liang (Zhejiang Normal University), Jiajie Xu (Southeast University)

CodeExplainability and InterpretabilityComputational EfficiencyRepresentation LearningData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark

🎯 What it does: Propose the RSDA framework, which quantifies the reconstruction value of samples through potential entropy and dynamically selects component-level renovation strategies to enhance the adaptability of scarce data;

RST-Guarder: Enhancing Long-Context Robustness for Safeguards via RST Parsing and Probabilistic Inference

Xu Zhang (Peking University), Xiaojun Wan (Peking University)

CodeSafty and PrivacyExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought

🎯 What it does: Propose a method called RST-Guarder for enhancing the safety detection of long texts during inference. It utilizes Rhetorical Structure Theory (RST) to parse and construct discourse hierarchy structures, and performs hierarchical probabilistic reasoning on this basis, thereby improving the detection accuracy of existing Guardrail models on long texts.

SAC: Neural Speech Codec with Semantic-Acoustic Dual-Stream Quantization

Wenxi Chen (Shanghai Jiao Tong University), Xie Chen (Shanghai Jiao Tong University)

CodeCompressionTransformerSupervised Fine-TuningAuto EncoderGenerative Adversarial NetworkContrastive LearningAudio

🎯 What it does: Proposed a semantic-acoustic dual-stream neural speech codec (SAC), which decouples and separately optimizes semantic content and acoustic details by using a pre-trained semantic tokenizer and an acoustic quantization module respectively, ultimately achieving high-quality reconstruction and semantic expression;

SAD: A Large-Scale Strategic Argumentative Dialogue Dataset

YongKang Liu (Northeastern University), Hinrich Schuetze (LMU Munich)

CodeGenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextSequentialRetrieval-Augmented Generation

🎯 What it does: Constructed and publicly released a large multi-turn argumentative dialogue dataset called SAD, and defined a strategy-controlled argument generation task on this dataset.

Safety-Utility Conflicts Are Not Global: Surgical Alignment via Head-Level Diagnosis

Wang Cai (Peking University), Yunfang Wu (Peking University)

CodeOptimizationSafty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText

🎯 What it does: Propose the CAST framework, which addresses the conflict between the safety and general capabilities of large language models by performing conflict diagnosis and sparse fine-tuning at the attention head level;

SAFO: Stable Adaptive Fairness Optimization for LLM-Based Social Survey Simulation

Chenxi Lin (Zhejiang University), Yiquan Wu (Alibaba Group)

CodeOptimizationFederated LearningExplainability and InterpretabilityAdversarial AttackData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTabularBenchmark

🎯 What it does: Propose a dynamic fairness and stability optimization framework named SAFO for training large language models in social survey simulations;

SAGE: Synergistic Adaptive Gating of Experts for Hateful Video Detection

Jie Huang (State Key Laboratory of Complex System Modeling and Simulation Technology), Qing Wang (State Key Laboratory of Complex System Modeling and Simulation Technology)

CodeClassificationAnomaly DetectionTransformerLarge Language ModelMixture of ExpertsContrastive LearningVideoTextMultimodalityAudio

🎯 What it does: Propose the SAGE framework for detecting hate videos, adopting a decoupled expert and instance-level decision arbitration approach, rather than traditional feature fusion;

SAHM: A Benchmark for Arabic Financial and Shari’ah-Compliant Reasoning

Rania Elbadry (Mohamed bin Zayed University of Artificial Intelligence), Zhuohan Xie (Mohamed bin Zayed University of Artificial Intelligence)

CodeDomain AdaptationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmarkFinance RelatedRetrieval-Augmented Generation

🎯 What it does: Constructed and publicly released the first multi-task benchmark for Arabic financial text, SAHM, covering seven tasks (AAOIFI standard QA, fatwa QA, accounting and business multiple-choice questions, financial sentiment analysis, extractive summarization, and event-cause reasoning), and evaluated 20 LLMs, proposing that domain adaptation can significantly enhance Arabic financial reasoning capabilities.

SCAN: Structured Capability Assessment and Navigation for LLMs

Zongqi Wang (Tsinghua University), Yujiu Yang (Tsinghua University)

CodeClassificationExplainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsTextBenchmarkRetrieval-Augmented Generation

🎯 What it does: Established the SCAN framework for structured, fine-grained capability evaluation and navigation of LLMs.

Schoenfeld’s Anatomy of Mathematical Reasoning by Language Models

Ming Li (University of Maryland College Park), Tianyi Zhou (University of Maryland College Park)

CodeExplainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextReview/Survey PaperChain-of-Thought

🎯 What it does: Propose the ThinkARM framework, which divides the reasoning process of LLMs into eight categories (Read, Analyze, Plan, Implement, Explore, Verify, Monitor, Answer) based on Schoenfeld's Episode Theory, achieving sentence-level automatic annotation;

SciCoQA: Quality Assurance for Scientific Paper–Code Alignment

Tim BaumgΓ€rtner (TU Darmstadt), Iryna Gurevych (TU Darmstadt)

CodeLarge Language ModelTextMultimodalityBenchmarkPhysics Related

🎯 What it does: Designed and released the SCICOQA dataset to evaluate the performance of large language models in detecting inconsistencies between scientific papers and code.

SCOPE: Boosting LLM Efficiency with Scoped Position Encoding

Qingguo Qi (Zhejiang University), Zhao Li (Zhejiang Lab)

CodeComputational EfficiencyTransformerLarge Language ModelText

🎯 What it does: Proposes SCOPE, a framework that achieves implicit position encoding through exponentially expanded attentionθ§†ι‡Ž.

SED-SFT: Selectively Encouraging Diversity in Supervised Fine-Tuning

Yijie Chen (Tencent Inc), Fandong Meng (Tencent Inc)

CodeOptimizationData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmark

🎯 What it does: This paper proposes a method called SED‑SFT, which encourages generation diversity during the supervised fine-tuning (SFT) stage through selective entropy regularization.

See the Forest for the Trees: Loosely Speculative Decoding via Visual-Semantic Guidance for Efficient Inference of Video LLMs

Yicheng Ji (State Key Laboratory of Blockchain and Data Security, Zhejiang University), Huan Li (State Key Laboratory of Blockchain and Data Security, Zhejiang University)

CodeComputational EfficiencyRepresentation LearningTransformerLarge Language ModelVision Language ModelVideoText

🎯 What it does: Proposes a loose inference framework called LVSPEC based on visual semantic guidance, which significantly accelerates the autoregressive inference of video LLMs without training a draft model;

See2Refine: Vision-Language Feedback Improves LLM-Based eHMI Action Designers

Ding Xia, Takeo Igarashi (University of Tokyo)

CodeAutonomous DrivingOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringVision Language ModelImageVideoTextMultimodality

🎯 What it does: Built a closed-loop system called SEE2REFINE, which uses the perceptual evaluation of a vision-language model (VLM) as automated, human-free feedback to iteratively improve the action design of the external human-machine interface (eHMI) generated by the large language model (LLM).

Seeing the Whole Elephant: A Benchmark for Failure Attribution in LLM-based Multi-Agent Systems

Mengzhuo Chen (State Key Laboratory of Complex System Modeling and Simulation Technology), Qing Wang (State Key Laboratory of Complex System Modeling and Simulation Technology)

CodeExplainability and InterpretabilityTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmark

🎯 What it does: Designed and released TraceElephant, a failure attribution benchmark for large language model (LLM)-based multi-agent systems (MAS), which collects complete executable execution trajectories and reproducible environments; and systematically evaluated multiple attribution methods on this benchmark;

Segment, Embed, and Align: A Universal Recipe for Aligning Subtitles to Signing

Zifan Jiang (University of Oxford), Andrew Zisserman (University of Zurich)

CodeRecognitionSegmentationRetrievalRecurrent Neural NetworkTransformerVision Language ModelContrastive LearningOptical FlowVideoTextMultimodality

🎯 What it does: Propose the SEA method, which automatically aligns captions with continuous sign language videos through three steps: segmentation, embedding, and alignment.

Selective Knowledge Distillation: Fusing LLM Semantic Strengths with DNN Efficiency for Binary Code Similarity Detection

Shize Zhou (Zhejiang University), Wenhai Wang (Hangzhou Dianzi University)

CodeRetrievalComputational EfficiencyKnowledge DistillationConvolutional Neural NetworkTransformerLarge Language ModelContrastive LearningText

🎯 What it does: Proposed the BinSKD framework, which transfers the high-level semantic knowledge of large language models to lightweight deep neural networks through selective distillation, thereby improving the accuracy and robustness of binary code similarity detection (BCSD).

Self-Guided Alignment: Adaptive Preference Sensing for Multi-Objective Generation

Ning Wang (Jiangnan University), Haojie Zhou (Jiangnan University)

CodeGenerationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBenchmark

🎯 What it does: This paper proposes the Self-Guided Alignment (SGA) framework, which unifies preference space learning and conditional generation through a dual-head structure, enabling adaptive preference perception and self-guided generation during inference without requiring manual input.

Self-Reflective Generation at Test Time

Jian Mu (Hong Kong University of Science and Technology (Guangzhou)), Yao Shu (Hong Kong University of Science and Technology (Guangzhou))

CodeGenerationExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextSequentialRetrieval-Augmented GenerationChain-of-Thought

🎯 What it does: Propose a generation framework called SRGen that self-reflects during testing, dynamically detecting high-uncertainty tokens and instantly optimizing correction vectors to improve the reasoning reliability of large language models.

SGPVT: Self-Generated Proximal Visual Tokens for Mitigating Proximal Collateral Damage in MLLM Unlearning

Jiaqi Li (Southeast University), Guilin Qi (Southeast University)

CodeSafty and PrivacyExplainability and InterpretabilityKnowledge DistillationPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodalityBenchmark

🎯 What it does: This paper proposes a novel MLLM forgetting mechanism that utilizes self-generated approximate visual tokens (SGPVT) to forget target concepts while minimizing damage to related concepts.

SHARP: Self-adaptive Harmful Category-aware Prompt Generation for Black-box Jailbreaking

Yingjie Xue (Wuhan University), Fei Li (Wuhan University)

CodeSafty and PrivacyAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextRetrieval-Augmented Generation

🎯 What it does: Developed SHARP, the first category-aware black-box jailbreak framework, capable of adaptively generating prompts that bypass security mechanisms based on the semantic categories of harmful questions.

Sherry: Hardware-Efficient 1.25-Bit Ternary Quantization via Fine-grained Sparsification

Hong Huang (City University of Hong Kong), Dapeng Wu (City University of Hong Kong)

CodeComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelText

🎯 What it does: Proposed the Sherry framework, achieving 1.25-bit hardware-friendly ternary quantization, utilizing a 3:4 sparse structure to compress four weights into five bits, and addressing the weight trap in sparse ternary training through the Annealing Residual Synapse (Arenas) mechanism;

Shuttle Between Symbolic Instructions and Neural Parameters of Large Language Models

Wangtao Sun (Key Laboratory of Cognition and Decision Intelligence for Complex Systems, Institute of Automation, Chinese Academy of Sciences), Kang Liu (Key Laboratory of Cognition and Decision Intelligence for Complex Systems, Institute of Automation, Chinese Academy of Sciences)

CodeRepresentation LearningData-Centric LearningAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringAuto EncoderText

🎯 What it does: Proposed the SHIP framework, achieving bidirectional mapping between symbolic instructions and LLM parameters, and verified its effectiveness in tasks such as reasoning and induction.

SILO-BENCH: A Scalable Environment for Evaluating Distributed Coordination in Multi-Agent LLM Systems

Yuzhe Zhang (Beijing University of Technology), Wenyuan Jiang (ETH ZΓΌrich)

CodeTransformerLarge Language ModelAgentic AITextBenchmark

🎯 What it does: Proposed a role-agnostic, scalable multi-agent LLM evaluation benchmark called SILO-BENCH, designed to measure agents' distributed coordination capabilities under information silos.

SLoRA: Balancing Plasticity and Forgetting in Large Language Models for Continual Learning

Lina Yang (Shanghai Jiao Tong University), Yu Wang (Shanghai Jiao Tong University)

CodeComputational EfficiencyRepresentation LearningTransformerLarge Language ModelTextBenchmark

🎯 What it does: This paper studies the problem of catastrophic forgetting in large language models during continual learning, and proposes an SLoRA framework based on subspace denoising, which automatically removes noisy components by leveraging the subspace similarity in LoRA low-rank updates, thereby mitigating forgetting.

SLR: Automated Synthesis for Scalable Logical Reasoning

Lukas Helff (TU Darmstadt), Kristian Kersting (TU Darmstadt)

CodeData SynthesisExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBenchmarkChain-of-Thought

🎯 What it does: Proposed the SLR (Scalable Logical Reasoning) framework, which automatically generates verifiable inductive logical reasoning tasks, constructs a 20-layer progressive SLR-BENCH benchmark, and uses this framework to train and evaluate LLMs.

SOAR: Supervision from Observation for Agentic Reinforcement Learning

Meng Li (Renmin University of China), Zang Li (Tencent)

CodeOptimizationTransformerReinforcement LearningAgentic AIPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought

🎯 What it does: Propose the SOAR method, which in Agentic Reinforcement Learning treats environmental observations as learning signals, assigns positive advantages to observation tokens, and uses the negative entropy of the previous action as weights, encouraging the agent to consider the results of actions during learning;

SOS-LoRA: Static Orthogonal-Subspace Low-Rank Adaptation with Fixed Multi-Scale Scaling

Yupeng Chang (Jilin University), Yi Chang (Jilin University)

CodeComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningMixture of ExpertsTextBenchmark

🎯 What it does: Propose a novel parameter-efficient fine-tuning method called SOS-LoRA based on LoRA, which splits the total rank into multiple static experts and introduces multi-scale scaling and cross-expert orthogonalization.

SPARK: Strategic Policy-Aware Exploration via Dynamic Branching for Long-Horizon Agentic Learning

Jinyang Wu (Tsinghua University), Jianhua Tao (Tsinghua University)

CodeOptimizationRobotic IntelligenceTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIPrompt EngineeringWorld ModelText

🎯 What it does: Propose the SPARK framework, which realizes dynamic branch exploration based on key decision points, utilizing the LLM's own <explore> signal to adaptively allocate exploration budget in long-horizon tasks;

Sparrow: Text-Anchored Window Attention with Visual-Semantic Glimpsing for Speculative Decoding in Video LLMs

Libo Zhang (National University of Defense Technology), Dongsheng Li (National University of Defense Technology)

CodeComputational EfficiencyRepresentation LearningTransformerLarge Language ModelMixture of ExpertsVision Language ModelDiffusion modelVideoTextMultimodality

🎯 What it does: Propose the Sparrow framework, which offloads visual computation to the target model through visual semantic internalization and hidden state reuse, achieving lossless acceleration for long video inference.

Sparse-RL: Breaking the Memory Wall in LLM Reinforcement Learning via Stable Sparse Rollouts

Sijia Luo (Renmin University of China), Jing Zhang (Ant Group)

CodeTransformerLarge Language ModelReinforcement LearningText

🎯 What it does: Studying how to use KV cache compression for sparse replay in large language model reinforcement learning to eliminate memory bottlenecks

SSG: Logit-Balanced Vocabulary Partitioning for LLM Watermarking

Chenxi Gu (Monash University), John C. Grundy (Monash University)

CodeGenerationSafty and PrivacyData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation

🎯 What it does: This paper proposes an LLM watermarking method called SSG based on vocabulary partitioning, aimed at improving the detectability of watermarks in low-entropy tasks.

Stable Language Guidance for Vision–Language–Action Models

Zhihao Zhan (Sun Yat-sen University), Guangrun Wang (Sun Yat-sen University)

CodeRobotic IntelligenceTransformerLarge Language ModelPrompt EngineeringVision-Language-Action ModelDiffusion modelImageTextMultimodality

🎯 What it does: Propose the Residual Semantic Steering (RSS) framework, which enhances the robustness of VLA models to instructions by decoupling visual priors from linguistic semantics.

StealthGraph: Exposing Domain-Specific Risks in LLMs through Knowledge-Graph-Guided Harmful Prompt Generation

Huawei Zheng (Zhejiang University), Dazhen Deng (Zhejiang University)

CodeSafty and PrivacyAdversarial AttackTransformerLarge Language ModelPrompt EngineeringTextGraphFinance RelatedRetrieval-Augmented Generation

🎯 What it does: Propose the StealthGraph framework, using knowledge graphs to guide LLMs in generating domain-specific stealthy harmful prompts

STReasoner: Empowering LLMs for Spatio-Temporal Reasoning in Time Series via Spatial-Aware Reinforcement Learning

Juntong Ni (Emory University), Wei Jin (Emory University)

CodeData SynthesisTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextMultimodalityTime SeriesBenchmarkChain-of-ThoughtStochastic Differential Equation

🎯 What it does: Proposes STReasoner, a multi-modal large language model designed for spatial-temporal reasoning in time series, and constructs the ST-Bench evaluation benchmark and an SDE-based multi-agent data synthesis pipeline.

STRIDE-ED: A Strategy-Grounded Stepwise Reasoning Framework for Empathetic Dialogue Systems

Hongru Ji (Northwestern Polytechnical University), Chao Gao (Northwestern Polytechnical University)

CodeGenerationExplainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextChain-of-Thought

🎯 What it does: Proposed the STRIDE-ED framework, which enhances empathetic dialogue generation quality by constructing a comprehensive emotional strategy system, step-by-step reasoning, and strategy-aware data processing.

Subject-level Inference for Realistic Text Anonymization Evaluation

Myeong Seok Oh (Tscientific), Hansaem Kim (Yonsei University)

CodeSafty and PrivacyTransformerLarge Language ModelPrompt EngineeringTextBenchmark

🎯 What it does: Propose the SPIA benchmark, using individuals as the unit of evaluation, to measure the inferability of individual information after text anonymization.

SwissGov-RSD: A Human-annotated, Cross-lingual Benchmark for Token-level Recognition of Semantic Differences Between Related Documents

Michelle Wastl (University of Zurich), Rico Sennrich (University of Zurich)

CodeRecognitionTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringContrastive LearningTextMultimodalityBenchmark

🎯 What it does: Constructed and released SwissGov-RSD, a human-annotated cross-lingual document-level semantic difference identification dataset, and evaluated multiple models on this dataset and its synthetic derivative sets.

Systematicity between Forms and Meanings across Languages Supports Efficient Communication

Doreen Osmelak (Saarland University), Kate McCurdy (Saarland University)

CodeExplainability and InterpretabilityComputational EfficiencyRepresentation LearningRecurrent Neural NetworkContrastive LearningText

🎯 What it does: Systematically study the form-meaning mapping of verbs and pronouns across languages, evaluating their learnability and simplicity within an information efficiency framework.

TA-GRPO-d: Trajectory-Aware GRPO for Optimizing Denoising Trajectories in Diffusion LLMs

Gyunyeop Kim (Gachon University), Sangwoo Kang (Gachon University)

CodeOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningDiffusion modelText

🎯 What it does: This paper proposes a trajectory-aware strategy based on reinforcement learning to dynamically decide which masked positions in the decoding process of discrete diffusion large language models need to be unmasked, as well as how many positions should be unmasked at each step, thus achieving adaptive and low-latency decoding.

TabEmb: Joint Semantic-Structure Embedding for Table Annotation

Ehsan Hoseinzade, Anandharaju Durai Raju

CodeClassificationRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelContrastive LearningTabularBenchmark

🎯 What it does: Proposes the TabEmb framework, which uses frozen LLMs to generate column semantic embeddings, and then performs structured information passing on the column graph through graph neural networks, completing three table annotation tasks: column type, column property, and table type.

TAIGR: Towards Modeling Influencer Content on Social Media via Structured, Pragmatic Inference

Nishanth Sridhar Nakshatri (Purdue University), Dan Goldwasser (Purdue University)

CodeRecommendation SystemExplainability and InterpretabilityData-Centric LearningGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringVideoTextBiomedical DataRetrieval-Augmented Generation

🎯 What it does: Propose the TAIGR framework, which conducts a three-stage analysis of influencer content on social media: extracting core recommendations (takeaway), constructing an argumentation structure, and inferring credibility through probabilistic factor graphs;

Talking to a Know-It-All GPT or a Second-Guesser Claude? How Repair reveals distinct Multi-Turn Behavior in LLMs

Clara Lachenmaier (Bielefeld University), Sina Zarrieß (Bielefeld University)

CodeExplainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought

🎯 What it does: Studying how LLMs perform repair in multi-turn dialogues and whether they proactively or passively initiate repair when faced with unanswerable questions.

TalkLoRA: Communication-Aware Mixture of Low-Rank Adaptation for Large Language Models

Lin Mu (Anhui University), Yiwen Zhang (Anhui University)

CodeComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelMixture of ExpertsText

🎯 What it does: Propose a communication-aware LoRA Mixture-of-Experts framework called TalkLoRA, which uses a lightweight dialogue module to achieve information interaction between low-rank experts, thereby enhancing the parameter-efficient fine-tuning effect of large language models.

TAMAS: Benchmarking Adversarial Risks in Multi-Agent LLM Systems

Ishan Kavathekar (International Institute of Information Technology Hyderabad), Tanuja Ganu (Microsoft Research India)

CodeSafty and PrivacyAdversarial AttackTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought

🎯 What it does: Constructed a multi-agent LLM security evaluation benchmark named TAMAS, which includes 5 high-risk domains, 6 attack types, 300 attack instances, and 100 harmless tasks. It evaluated the performance of 10 mainstream LLMs and 3 multi-agent collaboration configurations (Autogen and CrewAI) on metrics such as ARIA, PNA, and ERS.

TaxPraBen: A Scalable Benchmark for Structured Evaluation of LLMs in Chinese Real-World Tax Practice

Gang Hu (Yunnan University), Kun Yue (Wuhan University)

CodeTransformerLarge Language ModelPrompt EngineeringTextBenchmarkFinance RelatedRetrieval-Augmented GenerationChain-of-Thought

🎯 What it does: Designed and released TaxPraBen β€” an end-to-end evaluation benchmark tailored for Chinese tax practice, integrating 14 self-built datasets (totaling 7.3k instances), covering 10 traditional application tasks and 3 real-world scenarios (tax risk prevention, tax audit analysis, tax planning), and conducting zero-shot and one-shot evaluations on 19 mainstream LLMs.

Team-Based Self-Play With Dual Adaptive Weighting for Fine-Tuning LLMs

Wu Li (Harbin Institute of Technology), Jing Li (Harbin Institute of Technology)

CodeOptimizationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringText

🎯 What it does: The paper proposes a fully self-supervised team-based self-play algorithm called TPAW, which uses historical checkpoints as both opponents and main players to align large language models.

TEMA: Anchor the Image, Follow the Text for Multi-Modification Composed Image Retrieval

Zixu Li (Shandong University), Liqiang Nie (Harbin Institute of Technology)

CodeRetrievalTransformerLarge Language ModelPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodalityRetrieval-Augmented Generation

🎯 What it does: This paper proposes the TEMA image retrieval framework assisted by multi-modified semantic text (MMT), and constructs two multi-modified datasets, M-FashionIQ and M-CIRR.

Temporal Leakage in Search-Engine Date-Filtered Web Retrieval: A Retrospective Forecasting Case Study

Ali El Lahib (University of California, San Diego), Xinyu Pi (University of California, San Diego)

CodeRetrievalExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextTime SeriesBenchmarkRetrieval-Augmented Generation

🎯 What it does: Systematically audit the date filtering functions of Google Search and DuckDuckGo, discovering that most search results contain later information leakage, and evaluate the impact of these leaks on the prediction accuracy of LLMs.

TeRA: Vector-based Random Tensor Network for High-Rank Adaptation of Large Language Models

Yuxuan Gu (Imperial College London), Danilo Mandic (Imperial College London)

CodeComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelText

🎯 What it does: Designed TeRA, a method for efficiently fine-tuning large-scale language models using tensor networks to achieve high-rank parameter updates.

Text-Attributed Knowledge Graph Enrichment with Large Language Models for Medical Concept Representation

Mohsen Nayebi Kerdabadi (University of Kansas), Zijun Yao (University of Kansas)

CodeRepresentation LearningDrug DiscoveryGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringContrastive LearningTextGraphElectronic Health Records

🎯 What it does: Construct a global heterogeneous medical knowledge graph (diagnosis, drugs, surgeries), generate node descriptions and edge reasoning texts using large language models, and produce unified medical concept embeddings through LLM-GNN co-learning for improving EHR sequence prediction.

Text2Tabular – Reconstructing Tabular Research Data from Scientific Publications

Jonas Gottal (Technical University of Munich), Florian Matthes (Technical University of Munich)

CodeRestorationData SynthesisTransformerLarge Language ModelGaussian SplattingTextTabularBenchmarkChain-of-Thought

🎯 What it does: This paper proposes a system called Text2Tabular that automatically extracts statistical summaries from scientific papers and reconstructs complete tabular data.

The Bidirectional Process Reward Model

Lingyin Zhang (Soochow University), Ziqiang Cao (Soochow University)

CodeReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText

🎯 What it does: Propose BiPRM, a bidirectional evaluation process reward model to improve the reasoning quality of large language models (LLMs)

The Bitter Lesson of Diffusion Language Models for Agentic Workflows: A Comprehensive Reality Check

Qingyu Lu (Southeast University), Dacheng Tao (Nanyang Technological University)

CodeAutonomous DrivingOptimizationFederated LearningRobotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringDiffusion modelTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought

🎯 What it does: Systematically evaluate the performance of diffusion-based large language models (dLLMs) in multi-turn agent tasks and propose a pluggable multi-agent evaluation framework called DiffuAgent;

The Dominance of Text Space: Unveiling the Asymmetric Nature of Cross-Modal Alignment in Large Language Models

Linqing Chen (PatSnap Co., LTD.), Peng Zhou (PatSnap Co., LTD.)

CodeRepresentation LearningData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningImageTextMultimodality

🎯 What it does: Investigate the geometric properties of cross-modal alignment, propose the hypothesis that 'the text space serves as an anchor,' and design the Anchor-Preserving Projection (APP) method based on this hypothesis;

The GaoYao Benchmark: A Comprehensive Framework for Evaluating Multilingual and Multicultural Abilities of Large Language Models

Yilun Liu (Huawei), Yanghua Xiao (Huawei)

CodeTransformerLarge Language ModelPrompt EngineeringTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought

🎯 What it does: Designed and released the GaoYao multilingual and multicultural evaluation benchmark, covering 182.3k samples, 26 languages, and 51 regions, constructing a three-layer cultural framework (general, multicultural, monolingual) and nine sublayers; expanded instruction following and dialogue in 19 languages through expert localization, and built a cross-cultural test set SUPERBLEND with 34 cultures.

The Illusion of Specialization: Unveiling the Domain-Invariant "Standing Committee" in Mixture-of-Experts Models

Yan Wang (Fin AI), Zining Zhu (Stevens Institute of Technology)

CodeExplainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerSupervised Fine-TuningMixture of ExpertsContrastive LearningTextBenchmark

🎯 What it does: This paper proposes the COMMITTEEAUDIT framework, which conducts group-level analysis of routing behavior in Mixture-of-Experts models, revealing the existence of a cross-domain invariant 'Standing Committee' of core experts;

The Path Not Taken: Duality in Reasoning about Program Execution

Eshgin Hasanov (University of Central Florida), Aashish Yadavally (University of Central Florida)

CodeExplainability and InterpretabilityAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextBenchmark

🎯 What it does: Proposed the concept of 'bidirectional reasoning during program execution' and constructed the DEXBENCH benchmark to evaluate the ability of large language models in forward execution reasoning and backward reverse reasoning.

The Reasoning Trap: How Enhancing LLM Reasoning Amplifies Tool Hallucination

Chenlong Yin, Zechao Li (Independent Researcher)

CodeFederated LearningSafty and PrivacyExplainability and InterpretabilityKnowledge DistillationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought

🎯 What it does: This paper systematically investigates the phenomenon of hallucination (false tool invocation) during tool usage after enhancing LLM inference capabilities (such as RL, distillation, or switch thinking) by constructing the SIMPLETOOLHALLUBENCH benchmark, and reveals that inference reinforcement is the main cause of such hallucinations;

The Side Effects of Being Smart: Safety Risks in MLLMs’ Multi-Image Reasoning

Renmiao Chen (Tsinghua University), Minlie Huang (Tsinghua University)

CodeSafty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringVision Language ModelDiffusion modelImageTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought

🎯 What it does: Constructed MIR-SafetyBench, specifically designed to evaluate the safety risks of multi-modal large language models (MLLMs) in multi-graph reasoning scenarios, and conducted large-scale experiments on 19 mainstream MLLMs.

The Visual Iconicity Challenge: Evaluating Vision-Language Models on Sign Language Form–Meaning Mapping

Onur Keleş (Max Planck Institute for Psycholinguistics), Esam Ghaleb (Max Planck Institute for Psycholinguistics)

CodeRecognitionExplainability and InterpretabilityTransformerPrompt EngineeringVision Language ModelVideoTextMultimodality

🎯 What it does: Proposes the Visual Iconicity Challenge, evaluating the capabilities of vision-language models (VLMs) on three tasks: phoneme prediction, meaning transparency, and iconicity scoring, using NGT (Dutch Sign Language) videos.