Annual Meeting of the Association for Computational Linguistics · 557 papers
Collaborative Chain-of-Agents for Parametric-Retrieved Knowledge Synergy
Yi Jiang (Harbin Institute of Technology), Bing Qin (Harbin Institute of Technology)
CodeRetrievalRecommendation SystemExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningAgentic AITextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the CoCoA framework, which adopts multi-agent collaborative reasoning and long-chain training, explicitly fusing model parameter knowledge with retrieval knowledge;
Combining Distantly Supervised Models with In Context Learning for Monolingual and Cross-Lingual Relation Extraction
Vipul Kumar Rathore (Indian Institute of Technology New Delhi), Mausam (Indian Institute of Technology New Delhi)
CodeTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextReview/Survey PaperBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes the HYDRE framework, which combines candidate relation generation based on DSRE with context learning of LLMs, utilizing dynamically retrieved high-quality examples for relation discrimination.
🎯 What it does: This paper constructs rule-based senders and receivers using a morphology dictionary induced by CSAR in a reconstruction game, verifying their communication effectiveness in Emergent Language (EL), and compares them with senders/receivers based on neural network genetics; meanwhile, through ablation experiments on the dictionary and sender/receiver algorithms, it reveals the morphological and syntactic features in EL and proposes a new compositional metric called morpheme bijectivity.
CodeComputational EfficiencyRepresentation LearningData-Centric LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper investigates the impact of various UTF-8-based tokenizers on the performance of pre-trained language models for Hindi and Marathi in downstream tasks (question answering, transcription, phonetic transcription, noise robustness), and proposes a tokenizer based on visual units (grapheme clusters).
Compressing then Matching: An Efficient Pre-training Paradigm for Multimodal Embedding
Da Li (State Key Laboratory of AI Safety Institute of Computing Technology Chinese Academy of Sciences), Guorui Zhou (Kuaishou Technology)
CodeData SynthesisComputational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringVision Language ModelAuto EncoderContrastive LearningImageTextMultimodality
🎯 What it does: Propose a two-stage pre-training framework called CoMa, which first uses compression pre-training to enable multi-modal large models to generate comprehensive compressed tokens, and then uses contrastive learning to map these tokens into a unified cross-modal embedding space.
Consolidation or Adaptation? PRISM: Disentangling SFT and RL Data via Gradient Concentration
Yang Zhao (Harbin Institute of Technology), Ting Liu (Harbin Institute of Technology)
CodeComputational EfficiencyData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AITextBenchmark
🎯 What it does: Propose a data allocation framework called PRISM, which uses gradient concentration to distinguish between SFT and RL training data, in order to achieve efficient training of LLM agents.
Continuous Interpretive Steering for Scalar Diversity
Ye-eun Cho (Sungkyunkwan University)
CodeExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Studied a mechanism for fine-tuning reasoning in large language models that probes scalar diversity by continuously regulating the mediation direction in the activation space.
CodeAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose a configurable warehouse-level evaluation benchmark called CoreCodeBench, and decompose public Python repositories into adjustable difficulty and fine-grained single-function and multi-function tasks through the automated COREPIPE framework, thereby splitting different cognitive dimensions of code intelligence.
Corpus-Dependent Subcharacter Encoding via HMM-Guided Code Assignment
Tatsuya Hiraoka (Mohamed bin Zayed University of Artificial Intelligence)
CodeClassificationCompressionData-Centric LearningRecurrent Neural NetworkPrompt EngineeringAuto EncoderContrastive LearningTextReview/Survey Paper
🎯 What it does: This paper proposes a corpus-based sub-character encoding framework called Latom, which models the context of character occurrences using HMM and maps characters to fixed-length, reversible atomic codes (atom code) via the Hungarian algorithm;
CoSToM: Causal-oriented Steering for Intrinsic Theory-of-Mind Alignment in Large Language Models
Mengfan Li (Huazhong University of Science and Technology), Yang Deng (Singapore Management University)
CodeExplainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Propose the COSTOM framework, which first locates the layers containing Theory of Mind (ToM) information in LLMs through causal tracing, and then performs lightweight alignment with activation-oriented adjustments on these layers;
CTTA-T: Continual Test-Time Adaptation for Text Understanding via Teacher-Student with a Domain-aware and Generalized Teacher
Tianlun Liu, Dongsheng Li (National University Of Defense Technology)
CodeDomain AdaptationKnowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextBenchmark
🎯 What it does: In text understanding tasks, a Continuous Test-Time Adaptation (CTTA) framework called CTTA‑T is proposed to online adapt in continuously changing target domains.
CURA: Clinical Uncertainty Risk Alignment for Language Model–Based Risk Prediction
Sizhe Wang (Washington University in St. Louis), Chenyang Lu (Washington University in St. Louis)
CodeClassificationExplainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningTextBiomedical DataElectronic Health Records
🎯 What it does: Propose a dual-layer uncertainty calibration framework called CURA for risk prediction based on clinical language models;
d-TreeRPO: Towards More Reliable Policy Optimization for Diffusion Language Models
Leyi Pan (Tsinghua University), Lijie Wen (Tsinghua University)
CodeOptimizationExplainability and InterpretabilityKnowledge DistillationTransformerLarge Language ModelReinforcement LearningDiffusion modelTextChain-of-Thought
🎯 What it does: Proposes d-TreeRPO, a reliable reinforcement learning framework for diffusion-based large language models, which utilizes tree-structured rollout to achieve fine-grained verifiable rewards and improves prediction confidence through temporal self-distillation, thereby reducing probability estimation errors.
DART: Disambiguation-Aware Reasoning for Video-guided Machine Translation
Boyu Guan (Chinese Academy of Sciences), Chengqing Zong (University of Chinese Academy of Sciences)
CodeComputational EfficiencyRepresentation LearningData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringVision Language ModelContrastive LearningVideoTextMultimodality
🎯 What it does: This paper proposes the DART framework, which uses reinforcement learning and the TVRF mechanism to dynamically determine whether video context is needed and perform corresponding reasoning in video-assisted machine translation;
Data Efficient RLVR via Off-Policy Influence Guidance
Erle Zhu (Tsinghua University), Hongning Wang (Tsinghua University)
CodeData-Centric LearningTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: This paper proposes a data selection framework for RLVR based on offline gradient estimation using influence functions and sparse random projection (CROPI), aimed at accelerating reinforcement learning in large language models.
🎯 What it does: Propose the DEBAR framework, which uses dual-stream graph reasoning and bridge-aware input construction to alleviate relation transfer bias in cross-document relation extraction, and introduces a curriculum learning-based ranking optimization loss to enhance discriminativeness.
Hao Li (Washington University in St. Louis), Lianli Gao (University of Electronic Science and Technology of China)
CodeClassificationRetrievalData-Centric LearningTransformerVision Language ModelScore-based ModelContrastive LearningImageTextMultimodality
🎯 What it does: Propose the One-Step Anti-Noise (OSA) method, which identifies and mitigates noisy labels in a single forward pass by leveraging the high-dimensional orthogonality of pre-trained multi-modal models.
DecIF: Improving Instruction-Following through Decomposition
Tingfeng Hui (Beijing University of Posts and Telecommunications), Sen Su (Beijing University of Posts and Telecommunications)
CodeData SynthesisReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Propose DecIF, a decomposition-based automated data synthesis framework that can generate verifiable, controllable, and diverse instruction-following data from scratch for supervised fine-tuning (SFT) and reinforcement learning (RL) training.
DecoCal: Decoding with Calibration in Diffusion Large Language Models
Fan Xu (Peking University), Xiaojun Wan (Peking University)
CodeExplainability and InterpretabilityComputational EfficiencyAI Code AssistantTransformerLarge Language ModelPrompt EngineeringDiffusion modelContrastive LearningText
🎯 What it does: Proposes the DecoCal decoding framework, which utilizes the historical predictions across diffusion steps through KL-weighted aggregation for token confidence calibration, and determines when to decode or re-mask based on the calibrated confidence, thereby achieving more reliable early decoding and error correction in DLLMs.
CodeExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Designed an evolvable evaluation framework called DeepFact, which uses the Audit-then-Score mechanism to verify the truthfulness of claims in deep research reports (DRR), and released the corresponding benchmark and evaluator DeepFact-Eval.
DEFT: Demystifying VLN Failures via a Unified Dual-View Explainability Framework for LLM-based Agents
Yawen Wang (Chinese Academy of Sciences), Qing Wang (Chinese Academy of Sciences)
CodeExplainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelMixture of ExpertsVision Language ModelVision-Language-Action ModelImageTextMultimodality
🎯 What it does: Propose the DEFT framework for joint temporal and visual interpretation of LLM-driven vision-language navigation (VLN) agents;
DeReA: Improving Idiom Translation with Detect-Retrieve-Arbitrate Reasoning
Rongqing Jiang (Harbin Institute of Technology), Min Zhang (Harbin Institute of Technology)
CodeGenerationRetrievalTransformerLarge Language ModelReinforcement LearningTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed a three-stage (Detect‑Retrieve‑Arbitrate) framework named DEREA, simulating the human translation process. It first uses a preference-aligned detector to identify idioms in sentences, then employs a refined retriever to obtain corresponding explanations from a knowledge base, and finally selects the optimal result between direct translation and retrieval-enhanced translation through a dual-path arbitration mechanism. Based on this, a new high-difficulty idiom translation benchmark, LoMI, and an Emerging Slang dataset were constructed to verify the model's performance on novel and rapidly evolving idioms.
Deriving Character Logic from Storyline as Codified Decision Trees
Letian Peng (University of California, San Diego), Jingbo Shang (University of California, San Diego)
CodeExplainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes the Codified Decision Tree (CDT) framework, which automatically generates executable, context-aware character behavior decision trees from large-scale narrative text, aiming to enhance the consistency and interpretability of role-playing agents.
Diagnosing and Remedying Representation Deficiencies for Deterministic Reasoning in KGQA
Gewen Liang (Harbin Institute of Technology), Min Zhang (Harbin Institute of Technology)
CodeExplainability and InterpretabilityKnowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringContrastive LearningTextGraph
🎯 What it does: Proposed the DR2 method to diagnose and repair the non-deterministic reasoning problems caused by insufficient semantic representation in large language models during knowledge graph question answering (KGQA) tasks.
CodeGenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Propose a controllable difficulty multiple-choice fill-in-the-blank question distractor generation framework, which constructs annotated data by utilizing bidirectional candidate generation, filtering, and difficulty clustering, and trains the model using multi-task learning on this data.
DimABSA: Building Multilingual and Multidomain Datasets for Dimensional Aspect-Based Sentiment Analysis
Lung-Hao Lee (National Yang Ming Chiao Tung University), Saif M. Mohammad (National Research Council Canada)
CodeData-Centric LearningLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmarkFinance Related
🎯 What it does: Proposed the DimABSA multilingual multi-domain dimensional sentiment analysis dataset, and designed three subtasks and a unified evaluation metric cF1.
Discourse Realization of Generics in Human and LLM-generated Texts
Søren Kirkegaard Fomsgaard (University of Caen), Nelleke Oostdijk (Radboud University)
CodeClassificationExplainability and InterpretabilityRecurrent Neural NetworkTransformerLarge Language ModelPrompt EngineeringTextReview/Survey PaperRetrieval-Augmented Generation
🎯 What it does: This paper constructs sentence-level generalization annotations and designs a text-level generalization score metric to compare the generalization usage of humans and large language models in argumentative texts.
DIXITWORLD: Evaluating Multimodal Abductive Reasoning in Vision-Language Models with Multi-Agent Dixit Gameplay
Yunxiang MO, Yangqiu Song (HKUST)
CodeExplainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Designed the DixitWorld evaluation suite, which includes the multi-agent dynamic game DixitArena and the static multi-difficulty QA benchmark DixitBench, for systematic evaluation of the generative and selective capabilities of vision-language models in multi-modal inductive reasoning.
DMHM: Density-aware Manifold Learning and Hybrid Mahalanobis Energy for LLMs-generated Text Detection
Tianle Liu (National University of Defense Technology), Dongsheng Li (National University of Defense Technology)
CodeAnomaly DetectionTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Propose a detection framework for LLM-generated text (DMHM) based on density-aware manifold learning and hybrid Mahalanobis distance energy, which constructs a manifold with more distinct density differentiation in the embedding space and clearly separates generated text (LGT) from human-written text (HWT) through energy learning and contrastive learning.
Does RLVR Extend Reasoning Boundaries? Investigating Capability Expansion in Vision-Language Models
Minghe Shen (University College London), Che Liu (University of California, Los Angeles)
CodeExplainability and InterpretabilityComputational EfficiencyRepresentation LearningReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelContrastive LearningImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Investigate the extension effects of Reinforcement Learning with Verifiable Rewards (RLVR) on the spatial reasoning capabilities of Vision-Language Models (VLM), using the verifiable maze navigation experimental framework Ariadne to evaluate the reasoning boundary.
Don’t Adapt Small Language Models for Tools; Adapt Tool Schemas to the Models
Jonggeun Lee (Seoul National University), Yohan Jo (Seoul National University)
CodeOptimizationComputational EfficiencyRepresentation LearningData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes PA-Tool, an untrained tool architecture optimization method, which reduces pattern mismatch errors during tool usage by remapping tool and parameter names into naming patterns that the model is already familiar with during the pre-training phase.
DPC: Training-Free Text-to-SQL Candidate Selection via Dual-Paradigm Consistency
Boyan Li (Hong Kong University of Science and Technology (Guangzhou)), Yuyu Luo (Hong Kong University of Science and Technology (Guangzhou))
CodeComputational EfficiencyData-Centric LearningAI Code AssistantTransformerLarge Language ModelAgentic AIPrompt EngineeringTextTabularBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes the DPC framework, which selects candidate SQL without training by utilizing adversarial MDD and SQL/Python dual paradigm consistency.
DPWriter: Reinforcement Learning with Diverse Planning Branching for Creative Writing
Qian Cao (Renmin University of China), Han Li (Kuaishou Technology)
CodeGenerationData SynthesisReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Propose the DPWriter framework, which introduces an explicit planning phase by utilizing a semi-structured long Chain-of-Thought (CoT), and employs diverse planning branching and group-based diversity rewards in reinforcement learning to enhance the output diversity of creative writing.
DR-Arena: an Automated Evaluation Framework for Deep Research Agents
Yiwen Gao (National University of Singapore), Wenxuan Zhang (Singapore University of Technology and Design)
CodeLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose DR-Arena, a fully automated evaluation framework that constructs an information tree by real-time web scraping and uses an automated Examiner with an adaptive evolutionary loop to conduct dual-dimensional tests on Deep Research agents, including depth (multi-hop reasoning) and breadth (information aggregation).
DRInQ: Evaluating Conversational Implicature with Controlled Context Variation
Hirona Jacqueline Arai (University of Southern California), Xiang Ren (University of Southern California)
CodeTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes the DRInQ benchmark, which uses a semi-automated pipeline to generate multiple-choice instances with varying contexts that induce different meanings under the same question sentence, and evaluates the performance of LLMs in dialogue meaning reasoning.
DUAL RM: Beyond Rule-based Preference Reward Modeling via Meta-Reward
Xiaobo Liang, Min Zhang (Soochow University)
CodeOptimizationExplainability and InterpretabilityRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Built a dual reward framework (DUAL RM) that integrates discriminative and generative reward models, using non-parametric meta-rewards to guide generative models, solving the recursive validation bottleneck in reward models (Matryoshka Doll Problem);
DyBBT: Dynamic Balance via Bandit-inspired Targeting for Dialog Policy with Cognitive Dual Systems
Shuyu Zhang (Shanghai Jiao Tong University), Bin Li (SIAT Chinese Academy Of Sciences)
CodeKnowledge DistillationRepresentation LearningTransformerReinforcement LearningPrompt EngineeringMixture of ExpertsTextBenchmark
🎯 What it does: Propose the DyBBT framework, which balances dynamic exploration and exploitation through the cognitive state space C and a meta-controller.
Dynamic Infilling Anchors for Format-Constrained Generation in Diffusion Large Language Models
Boyan Han (AGI Lab Westlake University), Chi Zhang (AGI Lab Westlake University)
CodeGenerationTransformerLarge Language ModelDiffusion modelText
🎯 What it does: Propose a no-training, two-stage method called Dynamic Infilling Anchors (DIA) for achieving format-constrained text generation in diffusion large language models.
EASE: Entity-Aware Sub-table Generation for Real-world Multi-table QA
Myunghoon Kang (Korea University), Heuiseok Lim (Korea University)
CodeExplainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextTabularBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose a multi-table question answering framework called EASE based on entity-aware subtable generation, and construct a Noisy Multi-table QA dataset containing noisy tables;
EdgeFormer: Latency-Aware Collaborative Multi-Head Attention of Transformer Inference in Edge Networks
Yiming Yao (Beihang University), Tao Ren (Beihang University)
CodeFederated LearningComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsContrastive LearningTextRetrieval-Augmented Generation
🎯 What it does: Propose the EdgeFormer framework to achieve distributed Transformer inference, dynamically allocate model blocks, and improve parallelism through collaborative multi-head attention.
Editing the Moving World: Model Editing for Video LLMs
Qian Zhang (Harbin Institute of Technology), Dianbo Sui (Nanyang Technological University)
CodeExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelVideoTextMultimodalityBenchmark
🎯 What it does: Constructed the VMEB benchmark to evaluate the performance of six model editing methods on three Vid-LLMs;
EDSD: Entropy-Driven Design for Faster Speculative Decoding
Longkai Cheng (Ant Group), Haixiang Hu (Ant Group)
CodeComputational EfficiencyAI Code AssistantTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Proposes the EDSD framework, which utilizes entropy-driven draft model training and architecture design to achieve more efficient speculative decoding, thereby accelerating the inference of large language models.
EDUMATH: Generating Standards-aligned Educational Math Word Problems
Bryan R Christ (University of Virginia), Thomas Hartvigsen (University of Virginia)
CodeGenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Use large language models (LLM) to automatically generate educational math word problems (MWP) that comply with American standards, and construct a standardized dataset annotated by teachers (STEM) for training and evaluation.
Effects of Collaboration on the Performance of Interactive Theme Discovery Systems
Alvin Po-Chun Chen (University of Colorado Boulder), Maria Leonor Pacheco (University of Colorado Boulder)
CodeExplainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextReview/Survey PaperRetrieval-Augmented Generation
🎯 What it does: Compare the impact of synchronous and asynchronous collaboration modes on the quality of coding results from three types of NLP-assisted qualitative analysis tools (topic models, relational models, LLMs).
Efficient KL Divergence Estimation via Truncated Top-K Integration for Large Language Models
Xinyuan Wang (Fudan University), Xipeng Qiu (Fudan University)
CodeComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningScore-based ModelText
🎯 What it does: Proposed a TIKE method for estimating KL divergence based on Top-k truncation and importance sampling, addressing the memory bottleneck in KL computation during LLM RLHF training.
Efficient Multi-Agent System Training with Data Influence-Oriented Tree Search
Wentao Shi (University of Science and Technology of China), Chenyan Xiong (Carnegie Mellon University)
CodeOptimizationFederated LearningData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: This paper proposes the DITS framework, which utilizes influence scores to guide self-supervised data synthesis and selection in multi-agent systems, thereby improving model training performance.
Efficient Process Reward Modeling via Contrastive Mutual Information
Nakyung Lee (Seoul National University), Jungwoo Lee (Seoul National University)
CodeComputational EfficiencyRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringContrastive LearningTextBenchmark
🎯 What it does: This paper proposes an automatic step reward annotation method based on contrastive point mutual information (CPMI) to construct step-level supervision datasets for process reward models (PRMs), which are then validated across various reasoning tasks.
EfficientLLM: Unified Pruning-Aware Pretraining for Auto-Designed Compact Language Models
Xingrun Xing (Beijing Academy of Artificial Intelligence), Jiajun Zhang (University of Chinese Academy of Sciences)
CodeComputational EfficiencyKnowledge DistillationNeural Architecture SearchTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringContrastive LearningText
🎯 What it does: Propose a unified pruning-aware pre-training framework (EfficientLLM), which prunes large language models during the pre-training phase and automatically designs the structure of small models, thus achieving significant compression while maintaining performance.
CodeRobotic IntelligenceReinforcement Learning from Human FeedbackGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringVision Language ModelVision-Language-Action ModelImageTextMultimodalitySequentialRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Built and trained an embodied language-visual model called Embodied-Reasoner, capable of actively observing, reasoning, planning, searching, and executing actions in interactive environments.
Embracing Anisotropy: Turning Massive Activations into Interpretable Control Knobs for Large Language Models
Youngji Roh (Yonsei University), Jaehyung Kim (Yonsei University)
CodeDomain AdaptationSafty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Studied the heterogeneity of internal activations in large language models, finding that a small number of highly activated dimensions are domain-specific functional units. These 'domain-critical dimensions' were identified without training by analyzing activation magnitudes, and then a method called Critical Dimension Steering was proposed to regulate only these dimensions for more precise model control.
EMCEE: Improving Multilingual Capability of LLMs via Bridging Knowledge and Reasoning with Extracted Synthetic Multilingual Context
Hamin Koo (Yonsei University), Jaehyung Kim (Yonsei University)
CodeData SynthesisExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelPrompt EngineeringAuto EncoderTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the EMCEE framework, which leverages synthetic multilingual contexts generated by the LLM itself and integrates them with inference results to improve the quality of LLM responses to non-English queries.
EmoHarbor: Evaluating Personalized Emotional Support by Simulating the User’s Internal World
Jing Ye (State Key Laboratory of Multimodal Artificial Intelligence Systems Institute of Automation Chinese Academy of Sciences), Chengqing Zong (State Key Laboratory of Multimodal Artificial Intelligence Systems Institute of Automation Chinese Academy of Sciences)
CodeRecommendation SystemExplainability and InterpretabilityTransformerLarge Language ModelAgentic AITextTabularBenchmarkChain-of-Thought
🎯 What it does: This paper proposes the EmoHarbor framework, which uses the User-as-a-Judge mode to simulate the user's inner world through a chain-of-agent approach, constructs a benchmark with 100 multidimensional user profiles, and evaluates the personalized emotional support performance of 20 LLMs.
Enabling Agents to Communicate Entirely in Latent Space
Zhuoyun Du (Zhejiang University), Haochao Ying (Zhejiang University)
CodeExplainability and InterpretabilityComputational EfficiencyRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelAuto EncoderContrastive LearningTextChain-of-Thought
🎯 What it does: Propose the Interlat framework, enabling full communication between LLM agents in the latent space, directly transmitting the final hidden layer states of the model and performing compression;
Enhancing Lexical Relation Mining with Structured Sememe Knowledge
Hansi Wang (Key Laboratory of Computational Linguistics, Ministry of Education, Peking University), Yang Liu (School of Computer Science, Peking University)
CodeKnowledge DistillationRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextGraphBenchmark
🎯 What it does: This paper studies how to leverage structured word meaning knowledge (Sememe tree) to improve the lexical relation mining task, which includes two subtasks: lexical relation classification (LRC) and lexical entailment (LE);
Enhancing Linguistic Competence of Language Models through Pre-training with Language Learning Tasks
Atsuki Yamaguchi (University of Sheffield), Nikolaos Aletras (University of Sheffield)
CodeRepresentation LearningData-Centric LearningTransformerLarge Language ModelContrastive LearningTextBenchmark
🎯 What it does: Propose the L2T pre-training framework, adding 14 language learning tasks alongside the CLM objective to enhance the model's linguistic capabilities.
Enhancing LLM-based Search Agents via Contribution Weighted Group Relative Policy Optimization
Junzhe Wang (Fudan University), Qi Zhang (Fudan University)
CodeRetrievalOptimizationTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: A method called Contribution-Weighted GRPO (CW-GRPO) is proposed, which integrates process supervision into the Group Relative Policy Optimization (GRPO) framework. It evaluates the contribution of each retrieval round through an LLM evaluator, thereby achieving fine-grained credit allocation while maintaining the stability of GRPO.
Enhancing Multimodal Large Language Models for Ancient Chinese Character Evolution Analysis via Glyph-Driven Fine-Tuning
Rui Song (Jilin University), Hao Xu (Jilin University)
CodeRecognitionTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: A benchmark for evaluating the evolution of ancient characters in multi-modal large language models (MLLMs) was constructed, and model capabilities were assessed through tasks of character shape comparison and evolutionary reasoning. A stage-wise fine-tuning framework called GEVO, based on character shape comparison, was proposed to enhance the model's evolutionary reasoning and character shape recognition performance.
Enhancing the Transferability of Jailbreak Attacks on Large Language Models via Exploiting Reparameterization Invariance
Ao Wang (China University of Petroleum East China), Weifeng Liu (China University of Petroleum East China)
CodeAdversarial AttackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This study proposes a token-level jailbreaking attack framework called RIGJ based on natural gradients, aiming to enhance cross-model transferability on large language models.
Erasing Without Remembering: Implicit Knowledge Forgetting in Large Language Models
Huazheng Wang (Beijing University of Posts and Telecommunications), Dacheng Tao (Nanyang Technological University)
CodeSafty and PrivacyExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
🎯 What it does: This study investigates the issue of implicit knowledge forgetting in large language models, finding that existing forgetting methods have poor generalization capabilities for related semantic variations (such as rewrites, relation reversals, etc.); it proposes PERMU, a forgetting framework based on probabilistic perturbation, which injects noise into the most perturbation-sensitive tokens using adversarial samples and model sensitivity metric MSM, thereby suppressing the probability of fact-related tokens without significantly compromising model utility.
CodeGenerationConvolutional Neural NetworkRecurrent Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningAuto EncoderTextMultimodalityRetrieval-Augmented GenerationAudio
🎯 What it does: This paper proposes the ES4R framework for generating empathetic text and speech responses from voice inputs, achieving a complete process in three stages (empathy understanding, empathy generation, and speech synthesis);
Escaping the Echo Trap: On Credit Assignment Failure in Multi-turn LLM Self-Reflection
Linxuan Du (Soochow University), Min Zhang (Soochow University)
CodeOptimizationExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Propose a tree-structured GRPO extension called TRAE for multi-round self-reflection, addressing behavior collapse caused by Echo Trap.
ETR: Entropy Trend Reward for Efficient Chain-of-Thought Reasoning
Xuan Xiong (University of Toronto), Yang Wang (Concordia University)
CodeComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextChain-of-Thought
🎯 What it does: This paper proposes an entropy trend-based reward mechanism (ETR) to encourage gradual reduction of uncertainty in chain-of-thought (CoT) reasoning, thereby achieving efficient, short-chain reasoning while maintaining answer accuracy.
Evaluating Robustness of Large Language Models Against Multilingual Typographical Errors
Raoyuan Zhao (LMU Munich), Michael A. Hedderich (LMU Munich)
CodeTransformerLarge Language ModelPrompt EngineeringTextMultimodalityBenchmark
🎯 What it does: Proposed the MULTYPO multilingual typing error generation algorithm, and systematically evaluated its robustness in multilingual and multitask environments across 18 open-source LLMs
Every Response Counts: Quantifying Uncertainty of LLM-based Multi-Agent Systems through Tensor Decomposition
Tiejin Chen (Arizona State University), Hua Wei (Arizona State University)
CodeOptimizationExplainability and InterpretabilityComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Quantify uncertainty in multi-agent systems (MAS) built with large language models (LLM), proposing the MATU framework.
Evolving Sparsity: Leveraging Token Importance Dynamics for Efficient LLM Decoding with Sparse Attention
Ruizi Han (Harbin Institute of Technology (Shenzhen)), Liqiang Nie (Harbin Institute of Technology (Shenzhen))
CodeComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Propose the EvoSparse framework, which improves sparse attention in LLM long-text reasoning by using cross-step accumulation and cross-layer propagation to dynamically capture token importance, significantly enhancing efficiency and performance.
EVOTOOL: Self-Evolving Tool-Use Policy Optimization in LLM Agents via Blame-Aware Mutation and Diversity-Aware Selection
Shuo Yang (University of Melbourne), Eduard Hovy (University of Melbourne)
CodeOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringTextBenchmark
🎯 What it does: Proposes EVOTOOL, a self-evolving framework for optimizing tool usage strategies of LLM agents;
Execution as Verification: Fine-Grained Self-Correcting Reasoning for Complex KBQA
Minghan Zhang (Anhui University), Shu Zhao (Anhui University)
CodeExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelTextGraphRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes the Execution as Verification (EVER) framework, which reconstructs semantic parsing as an iterative self-correcting reasoning process driven by execution feedback, aiming to improve the executability of queries and the accuracy of answers in complex knowledge base question answering (KBQA).
ExPerT: Personalizing LLM Responses to Users’ Domain Expertise via Query-Wise Semantic and Keystroke Behavioral Cues
Yeji Park (UNIST), Taesik Gong (UNIST)
CodeRecommendation SystemExplainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextSequentialRetrieval-Augmented Generation
🎯 What it does: Propose the ExPerT framework, which infers user domain expertise based on dual clues at the query level: semantic content and keystroke dynamics, and generates personalized LLM responses accordingly.
Explainable and Fine-Grained Safeguarding of LLM Multi-Agent Systems via Bi-Level Graph Anomaly Detection
Junjun Pan (Griffith University), Shirui Pan (Griffith University)
CodeAnomaly DetectionSafty and PrivacyExplainability and InterpretabilityGraph Neural NetworkTransformerLarge Language ModelAgentic AIContrastive LearningTextGraph
🎯 What it does: A framework named XG-Guard for unsupervised graph anomaly detection is proposed to identify and explain malicious agents in LLM multi-agent systems
Explainable Disentangled Representation Learning for Generalizable Authorship Attribution in the Era of Generative AI
Hieu Man (University of Oregon), Thien Huu Nguyen (University of Oregon)
CodeClassificationAnomaly DetectionExplainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelPrompt EngineeringAuto EncoderContrastive LearningText
🎯 What it does: Designed the Explainable Authorship Variational Autoencoder (EAVAE) framework, which achieves explicit and interpretable separation of author writing style and text content through two-stage pre-training and VAE fine-tuning, and applied it to authorship attribution and AI-generated text detection.
Exploring and Distilling Multi-Dimensional Clues for Interpretable Social Bot Detection
Yi Han (Harbin Institute of Technology), Ting Liu (Harbin Institute of Technology)
CodeClassificationAnomaly DetectionExplainability and InterpretabilityKnowledge DistillationTransformerLarge Language ModelReinforcement LearningTextGraphBenchmark
🎯 What it does: Propose a four-dimensional clue framework, train a clue generator based on reinforcement learning, and then fuse multi-dimensional clues to achieve interpretable social robot detection;
Extending First-Order Logic for Factual Reasoning over Knowledge Graphs
Yuanzhen Hao (University of Chinese Academy of Sciences), Desheng Wu (University of Chinese Academy of Sciences)
CodeExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelPrompt EngineeringTextGraphBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes an extended first-order logic FOLX-KG for fact reasoning on knowledge graphs (KG), and constructs the Fact-FOLX-KG dataset based on this.
False Friends in the Shell: Unveiling the Emoticon Semantic Confusion in Large Language Models
Weipeng Jiang (Xi'an Jiaotong University), Yang Liu (Nanyang Technological University)
CodeSafty and PrivacyExplainability and InterpretabilityAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: Studied and systematically evaluated the semantic confusion risks generated by large language models when processing ASCII emoticons, finding that they are prone to misinterpreting emoticons as executable code, constructed a corpus containing 3,757 test cases, and conducted quantitative experiments on six mainstream LLMs, revealing a confusion rate as high as 38.6% and a 'silent failure' risk exceeding 90%.
False Friends or Cognates? A Cross-lingual Semantic Ambiguity Evaluation for Galician, Portuguese and Spanish
Marta Vázquez Abuín (Centro Singular de Investigación en Tecnoloxías Intelixentes (CiTIUS) Universidade de Santiago de Compostela), Marcos Garcia (Centro Singular de Investigación en Tecnoloxías Intelixentes (CiTIUS) Universidade de Santiago de Compostela)
CodeClassificationRecognitionExplainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextBenchmark
🎯 What it does: This paper constructs a cross-lingual word sense disambiguation dataset based on true/false cognates (cognates and false friends) in Galician, Portuguese, and Spanish, and systematically evaluates the performance of various encoders and large language models on this task;
Fast and Accurate Fisher-Guided Quantization via Efficient Kronecker Factorization
Viktoriia A. Chekalina (FusionBrain Lab), Evgeny Frolov (AXXX)
CodeCompressionOptimizationComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelDiffusion modelAuto EncoderContrastive LearningText
🎯 What it does: This paper proposes FastKron, an algorithm that utilizes efficient Kronecker decomposition to compute second-order Hessian information and embeds it into post-training quantization, significantly accelerating the factorization process;
CodeOptimizationExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelReinforcement LearningPrompt EngineeringVision Language ModelDiffusion modelScore-based ModelGenerative Adversarial NetworkTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposed and implemented the FIGR system, which can dynamically construct and interactively update geometric figures through executable code during multi-round reasoning, achieving visual reasoning.
Fine-Mem: Fine-Grained Feedback Alignment for Long-Horizon Memory Management
Weitao Ma (Harbin Institute of Technology), Bing Qin (Harbin Institute of Technology)
CodeExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed the Fine-Mem framework, which uses reinforcement learning to train a learnable memory manager to address memory management issues in large language models during long-sequence tasks.
Fine-tuning vs. In-context Learning in Large Language Models: A Formal Language Learning Perspective
Bishwamittra Ghosh (Max Planck Institute for Software Systems), Evimaria Terzi (Boston University)
CodeExplainability and InterpretabilityComputational EfficiencyRepresentation LearningMeta LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Compare two learning modes of large language models—fine-tuning and in-context learning—by evaluating their language proficiency and inductive bias through formal language learning tasks;
FineSteer: A Unified Framework for Fine-Grained Inference-Time Steering in Large Language Models
Zixuan Weng (University of California, Los Angeles), Yuan Tian (University of California, Los Angeles)
CodeSafty and PrivacyExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsText
🎯 What it does: Propose the FineSteer framework to achieve fine-grained adjustment of the internal representations of large language models during inference, thereby enhancing safety and correctness.
Flattery in Motion: Benchmarking and Analyzing Sycophancy in Video-LLMs
Wenrui Zhou (Provable Responsible AI and Data Analytics Lab), Di Wang (Provable Responsible AI and Data Analytics Lab)
CodeExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringVision Language ModelContrastive LearningVideoTextMultimodalityBenchmark
🎯 What it does: Proposed the VISe benchmark, systematically evaluating and countering the sycophancy behavior of Video-LLMs;
FlowSearch: Advancing Deep Research with Dynamic Structured Knowledge Flow
Yusong Hu (Shanghai Artificial Intelligence Laboratory), Bo Zhang (Shanghai Artificial Intelligence Laboratory)
CodeDrug DiscoveryGraph Neural NetworkTransformerLarge Language ModelAgentic AITextGraphBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes FlowSearch, a multi-agent deep research framework based on dynamic structured knowledge flow, which can automatically plan, execute, and dynamically adjust knowledge flow in scientific exploration to achieve parallel reasoning and recursive refinement.
Follow the Flow: On Information Flow Across Textual Tokens in Text-to-Image Models
Guy Kaplan (Hebrew University of Jerusalem), Roy Schwartz (Hebrew University of Jerusalem)
CodeGenerationExplainability and InterpretabilityData-Centric LearningTransformerPrompt EngineeringVision Language ModelDiffusion modelContrastive LearningImageTextMultimodality
🎯 What it does: Investigate the impact of text encoding on alignment in text-to-image models, using causal masking and visual language models to evaluate the distribution of subword information and cross-item interactions, and propose de-redundant token and patching techniques to improve alignment;
Foresight Optimization for Strategic Reasoning in Large Language Models
Jessie Wang (Hong Kong Polytechnic University), Wenjie Li (Hong Kong Polytechnic University)
CodeOptimizationFederated LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Propose the FoPO (Foresight Policy Optimization) algorithm, which combines opponent modeling with policy optimization, leveraging self-play and offline reinforcement learning to enhance the strategic reasoning capabilities of large language models in multi-agent environments. Based on this, two dialogue datasets, Cooperative RSA and Competitive Taboo, are designed for training and evaluation.
FormalScience: Scalable Human-in-the-Loop Autoformalisation of Science with Agentic Code Generation in Lean
Jordan Meadows (Independent Researcher), Andre Freitas
CodeReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkPhysics Related
🎯 What it does: Proposes a human-in-the-loop semi-automated pipeline called FormalScience for converting informal reasoning in physics into formal proofs in Lean4;
Frame-Semantic Knowledge Injection for Event-Level Inference in LLMs
Shahid Iqbal Rai (University of Rome Tor Vergata), Roberto Basili (University of Rome Tor Vergata)
CodeExplainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Injecting the complete semantic knowledge of FrameNet into large language models using LoRA, and evaluating its effectiveness through NLI and SRL tasks.
Frankentext: Stitching random text fragments into long-form narratives
Chau Minh Pham (University of Maryland College Park), Mohit Iyyer (University of Maryland College Park)
CodeGenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: This work proposes and implements a long-form narrative generation framework called Frankentexts, which generates high-quality stories by having LLMs select, arrange, and connect text segments from a large collection of human-written fragments, resulting in stories that are mostly composed of original text.
From \log \pi to \pi: Taming Divergence in Soft Clipping via Bilateral Decoupled Decay of Probability Gradient Weight
Xiaoliang Fu (Meituan), Ke Zeng (Meituan)
CodeOptimizationRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Propose the DGPO algorithm, which addresses the gradient divergence and exploration limitations caused by soft clipping in RLVR through the use of probabilistic gradients and a bidirectional attenuation soft clipping method; significantly improves mathematical reasoning performance on large language models.
From Answers to Arguments: Toward Trustworthy Clinical Diagnostic Reasoning with Toulmin-Guided Curriculum Goal-Conditioned Learning
Chen Zhan (Shanghai University of Engineering Science), Xihe Qiu (Shanghai University of Engineering Science)
CodeExplainability and InterpretabilityDrug DiscoveryTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBiomedical DataElectronic Health Records
🎯 What it does: Proposes the CGCL course-based goal-oriented learning framework and the T-Eval evaluation tool, enhancing the transparency and reliability of LLMs in clinical diagnosis reasoning.
From Individual to Common: An Early Exploration of Consensus in Non-verifiable Data for Balanced Preference Optimization
Shangjian Yin (University of California Riverside), Zhouxing Shi (University of California Riverside)
CodeOptimizationData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Proposed the Donkey dataset and a training framework based on consensus preference optimization to improve the subjective alignment and objective performance of LLMs on non-verifiable data.
From Inheritance to Saturation: Disentangling the Evolution of Visual Redundancy for Architecture-Aware MLLM Inference Acceleration
Jiaqi Shi (University of Science and Technology of China), Jianzong Wang (Ping An Technology (Shenzhen) Co., Ltd.)
CodeComputational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringVision Language ModelContrastive LearningImageVideoTextMultimodality
🎯 What it does: This study investigates the issue of visual token redundancy during inference in multimodal large language models, and proposes the HalfV framework. It first uniformly eliminates Intrinsic Visual Redundancy (IVR), and then adaptively accelerates inference based on Secondary Saturation Redundancy (SSR) specific to different architectures. By analyzing the truncated matrix entropy, it reveals the three-stage redundancy lifecycle, and achieves a cross-architecture general inference acceleration solution.