ACL 2026 Papers — Page 21
Annual Meeting of the Association for Computational Linguistics · 2296 papers
The Bidirectional Process Reward Model
Lingyin Zhang (Soochow University), Ziqiang Cao (Soochow University)
Reinforcement 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)
Autonomous 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 Confidence Dichotomy: Analyzing and Mitigating Miscalibration in Tool-Use Agents
Weihao Xuan (University of Tokyo), Naoto Yokoya (University of Tokyo)
AI Code AssistantTransformerReinforcement LearningAgentic AIPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: The study investigates the confidence calibration issue in agents using tools, finding that evidence tools lead to overconfidence, while verification tools can alleviate this problem; it proposes the CALIBRATION Agentic RL (CAR) framework, which simultaneously optimizes task accuracy and confidence calibration through the Margin-Separated Calibration Reward (MSCR); it verifies the effectiveness of the framework on two types of tools—retrieval and code explanation—and demonstrates its robustness across domains and environments.
The Digital Dunning-Kruger Effect: Decoupling Hallucinations via Geometric Hidden-state Observation for Semantic Truthfulness
Yueheng Mao (Chinese Academy of Sciences), Ming Liu (Deakin University)
Anomaly DetectionExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelContrastive LearningTextFinance Related
🎯 What it does: Propose a white-box hallucination detection framework called GHOST based on internal geometric trajectories, used to real-time determine whether hallucinations exist in the text generated by LLMs;
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.)
Representation 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 Evolution of Thought: Tracking LLM Overthinking via Reasoning Dynamics Analysis
Zihao Wei (University of Chinese Academy of Sciences), Xueqi Cheng (Institute of Computing Technology)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: This paper investigates and quantifies the overthinking phenomenon that occurs in large language models during the reasoning process, proposing to analyze its two-stage characteristics through Reasoning Length Dynamics and Semantic Dynamics, defining the Reasoning Completion Point (RCP), and implementing the RCPD early stopping method.
The GaoYao Benchmark: A Comprehensive Framework for Evaluating Multilingual and Multicultural Abilities of Large Language Models
Yilun Liu (Huawei), Yanghua Xiao (Huawei)
TransformerLarge 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 GDN-CC Dataset: Automatic Corpus Clarification for AI-enhanced Democratic Citizen Consultations
Pierre-Antoine Lequeu (Sorbonne Université), Benjamin Piwowarski (Sorbonne Université)
Explainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Clarify the citizen consultation data from France's Grand Débat National by decomposing noisy, multi-topic statements into single-topic argument units, and structuring and clarifying each unit.
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)
Explainability 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 Imperfective Paradox in Large Language Models
Bolei Ma (LMU Munich), Yusuke Miyao (University of Tokyo)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: Evaluate the reasoning ability of large language models in handling the Imperfective Paradox, and construct a specialized diagnostic dataset called IMPERFECTIVENLI.
The Paradox of Outcome Optimization: A Causal Information-Theoretic Bound on Reasoning Shortcuts in LLMs
Zihan Chen (HFIPS, Chinese Academy of Sciences), Yining Sun (HFIPS, Chinese Academy of Sciences)
OptimizationExplainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose and verify the phenomenon of 'reward-induced manifold collapse,' clarifying that large language models prioritize learning low-complexity shortcuts over high-complexity causal reasoning paths during training due to optimization bias; and use a process reward model (PRM) as a topological filter to eliminate shortcuts, thereby enhancing the model's robustness to out-of-distribution (OOD) scenarios.
The Path Not Taken: Duality in Reasoning about Program Execution
Eshgin Hasanov (University of Central Florida), Aashish Yadavally (University of Central Florida)
Explainability 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 Personalization Trap: How User Memory Alters Emotional Reasoning in LLMs
Xi Fang (Amazon), Chandan K. Reddy (Amazon)
Recommendation SystemOptimizationFederated LearningExplainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Evaluate the impact of incorporating user memory on the emotional reasoning of large language models, and systematically measure how personalized memory introduces social bias.
The Pitfalls of KV Cache Compression
Alex Chen (University of California, Los Angeles), Daniel Mingyi Israel (University of California, Los Angeles)
CompressionExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Explore the potential risks and performance degradation of KV cache compression in multi-instruction prompts (such as system prompts), discovering issues of 'selective forgetting' and information leakage between instructions;
The Prosody of Emojis
Giulio Zhou (University of Edinburgh), Barry Haddow (Aveni)
RecognitionTransformerSupervised Fine-TuningContrastive LearningTextMultimodalityAudio
🎯 What it does: Investigate how emojis affect the prosodic expression of speech, and explore how listeners recover the intended meaning of emojis through prosodic cues.
The Proxy Presumption: From Semantic Embeddings to Valid Social Measures
Baishi Li (National University of Singapore), Ke-Wei Huang (National University of Singapore)
Explainability and InterpretabilityRepresentation LearningAdversarial AttackData-Centric LearningTransformerLarge Language ModelContrastive LearningTextReview/Survey PaperRetrieval-Augmented Generation
🎯 What it does: Propose the construction of a Construct Validity Protocol (CVP) that builds from semantic embeddings to effective social measurements, and validate its stability and deconfounding effectiveness through a case study on the GoEmotions dataset.
The Reasoning Trap: How Enhancing LLM Reasoning Amplifies Tool Hallucination
Chenlong Yin, Zechao Li (Independent Researcher)
Federated 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 Retrieval Bottleneck: Scaling Laws for Reinforcement Learning in RAG
Shu Zhou (Nanjing University), Hao Wang (Nanjing University)
RetrievalTransformerLarge Language ModelReinforcement LearningTextRetrieval-Augmented Generation
🎯 What it does: Systematically studied the scalability of retrieval-augmented generation (RAG) in reinforcement learning, proposed the retrieval bottleneck hypothesis, and derived the sigmoidal scaling law, based on which a RAG-SCALERL scheme with predictable performance was designed.
The Role of Mixed-Language Documents for Multilingual Large Language Model Pretraining
Jiandong Shao (Nanyang Technological University), Yao Lu (University College London)
Explainability and InterpretabilityRepresentation LearningData-Centric LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextMultimodality
🎯 What it does: This paper constructs the MONOWEB corpus, which contains only monolingual web pages (removing all bilingual documents, accounting for about 2%), and trains a multilingual LLM from scratch to systematically evaluate and compare four pre-training configurations: including full bilingual documents (FINEWEB), removing bilingual documents, adding only parallel documents, and adding only code-switching documents; differences between the two are assessed using multiple task metrics such as multilingual machine translation, cross-lingual QA, understanding, and reasoning.
The Side Effects of Being Smart: Safety Risks in MLLMs’ Multi-Image Reasoning
Renmiao Chen (Tsinghua University), Minlie Huang (Tsinghua University)
Safty 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 Stackelberg Speaker: Optimizing Persuasive Communication in Social Deduction Games
Zhang Zheng, Hao Wang (Shanghai Jiao Tong University)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringText
🎯 What it does: Model the dialogue in social reasoning games as Stackelberg competition and use reinforcement learning to train LLM agents to generate persuasive statements that can influence the behavior of subsequent players.
The Subjectivity of Respect in Police Traffic Stops: Modeling Community Perspectives in Body-Worn Camera Footage
Preni Golazizian (University of Southern California), Morteza Dehghani (University of Southern California)
Explainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVideoText
🎯 What it does: Fine-grained annotation of respect during traffic stops in a large number of law enforcement videos from the Los Angeles Police Department (LAPD), and the construction of a dataset evaluating respect and its reasons from multiple perspectives (law enforcement officers, affected individuals in the judicial system, and ordinary residents); meanwhile, a specialized respect evaluation rule based on procedural justice theory is proposed, and this rule is used to evaluate and align the reasons generated by the model; finally, an opinion-oriented language model is trained and aligned to predict and explain respect evaluations from different groups.
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)
RecognitionExplainability 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.
Theory-optimal Quantization Based on Flatness
Xiusheng Huang (Chinese Academy of Sciences), Kang Liu (Chinese Academy of Sciences)
Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Proposes the Bidirectional Diagonal Quantization (BDQ) framework, which utilizes a combination of diagonal scaling and Hadamard rotation to achieve low-bit post-training quantization, along with recursive cross-entropy loss to prevent overfitting.
Thermometer of Thoughts: Enhancing LLM’s Exploration via Attention Temperature Modulation
Zhiyuan Yu (Nanjing University), Sanglu Lu (Nanjing University)
OptimizationExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: Propose dynamically controlling the internal attention distribution of LLMs by adjusting the attention temperature, thereby enabling the exploration and switching between modes of thinking, and designing a two-stage reasoning strategy: first generating diverse reasoning trajectories by adjusting the attention temperature, and then aggregating based on problem difficulty.
Think before Go: Hierarchical Reasoning for Image-goal Navigation
Pengna Li (Xi'an Jiaotong University), Nanning Zheng (Xi'an Jiaotong University)
Robotic IntelligenceTransformerReinforcement LearningPrompt EngineeringVision Language ModelImageVideoTextMultimodalityRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose a hierarchical reasoning navigation framework, HRNav, which decomposes the image goal navigation task into two stages: high-level short-term planning and low-level execution.
Think Before Writing: Feature-Level Multi-Objective Optimization for Generative Citation Visibility
Zikang Liu (Nanjing University of Information Science and Technology), Peilan Xu (Nanjing University of Information Science and Technology)
GenerationRetrievalOptimizationTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes the FeatsGEO framework, which enhances citation visibility in generative retrieval systems through multi-objective optimization based on webpage structure, content, and linguistic features.
Think Better, Not Longer: Token-Level Marginal Utility for Efficient Reasoning in Large Reasoning Models
Jiawei Li (Beijing Institute of Technology), Chong Feng (Beijing Institute of Technology)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelReinforcement LearningTextBenchmarkChain-of-Thought
🎯 What it does: Propose a token-level marginal utility metric, and design the MUTO framework based on this metric as a signal, automatically suppressing redundant reasoning steps to achieve more concise Chain-of-Thought generation.
Think Faster Than Words: Efficient LLM Chain-of-Thought Reasoning via Dynamic Shortcut Decoding
Fan Liu (East China Normal University), Jiahui Pan (South China Normal University)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: Proposed the Shortcut Decoding framework, which adaptively exits the chain-of-thought reasoning steps during inference by utilizing internal hidden state probes and external entropy information, significantly reducing token usage;
Think How to Think: Mitigating Overthinking with Autonomous Difficulty Cognition in Large Reasoning Models
Yongjiang Liu (Hong Kong University of Science and Technology), Song Guo (Hong Kong University of Science and Technology)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextChain-of-Thought
🎯 What it does: This study proposes the TH2T framework, a two-stage fine-tuning approach, which utilizes difficulty hypnosis and redundancy hypnosis mechanisms to enable large reasoning models (LRMs) to automatically perceive problem difficulty and prune redundant steps during reasoning, significantly reducing overthinking.
Think in Latent Thoughts: A New Paradigm for Gloss-Free Sign Language Translation
Yiyang Jiang (Hong Kong Polytechnic University), Li Qing (Hong Kong Polytechnic University)
Explainability and InterpretabilityRepresentation LearningConvolutional Neural NetworkTransformerVision Language ModelVideoTextMultimodalityRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the SignThought framework, which uses cross-modal latent thought chains to achieve explicit intermediate reasoning, addressing the problem of sign language translation without gloss.
Think in Sentences: Explicit Sentence Boundaries Enhance Language Model’s Capabilities
Zhichen Liu (Southern University of Science and Technology), Yang Xu (Southern University of Science and Technology)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Explicitly inserting sentence boundary separators into the input of LLMs encourages the model to split text into sentences during inference and think sentence by sentence. Two implementation methods are provided: context insertion during inference (ICL) and enabling the model to have built-in sentence separation capabilities through supervised fine-tuning (SFT).
Think Natively: Unlocking Multilingual Reasoning with Consistency-Enhanced Reinforcement Learning
Xue Zhang (Beijing Jiaotong University), Jie Zhou (Tencent Inc)
Computational EfficiencyRepresentation LearningData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Develop the M-Thinker multilingual reasoning model, trained with reinforcement learning to address the issues of language inconsistency and decreased reasoning accuracy when handling non-English inputs.
Think Parallax: Solving Multi-Hop Problems via Multi-View Knowledge-Graph-Based Retrieval-Augmented Generation
Jinliang Liu (University of Electronic Science and Technology of China), Shaoning Zeng (University of Electronic Science and Technology of China)
RetrievalExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextGraphBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes ParallaxRAG, a framework that performs symmetric multi-view splitting for knowledge graph retrieval and generation. It maps queries and graphs into independent semantic subspaces via the multi-head attention of Transformers, and achieves perspective separation and path focusing in multi-hop reasoning by combining Pairwise Similarity Regularization (PSR) with query-aware gating.
Thinking Alignment of Scenario-Oriented User Simulation
Xiaoting Wu (JIUTIAN Research), Junlan Feng (JIUTIAN Research)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Train a thinking user simulator, ThinkingUS, which generates multi-turn dialogues using scenario descriptions and internal user thinking, thereby enhancing the realism of dialogue simulation.
Thinking beyond the anthropomorphic paradigm benefits LLM research
Lujain Ibrahim (University of Oxford), Myra Cheng (Stanford University)
Explainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextReview/Survey PaperRetrieval-Augmented Generation
🎯 What it does: This paper conducts a large-scale text analysis on the abstracts of over 20,000 computer science papers (especially those on LLM research), quantifies the prevalence of anthropomorphic language, and proposes a five-hypothesis framework covering the entire lifecycle of LLM development and deployment. It further identifies non-anthropomorphic research directions under each hypothesis.
Thinking in Schemas: Robust Syllogistic Reasoning in LLMs
Federico Ranaldi (University of Rome Tor Vergata), Shay B Cohen
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Propose the Boethius framework, which decomposes syllogistic reasoning into two stages: first identifying the semantic structure (Schema Identification), and then mapping the structure to specific propositions and determining validity (Schema Instantiation).
Thinking-Based Non-Thinking: Solving the Reward Hacking Problem in Training Hybrid Reasoning Models via Reinforcement Learning
Siyuan Gan (Nanjing University), Yang Gao (Nanjing University)
Explainability and InterpretabilityComputational EfficiencyTransformerReinforcement LearningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Proposed a hybrid reasoning training framework TNT (Thinking-Based Non-Thinking) based on reinforcement learning, which solves the reward hacking problem by dynamically determining the maximum token upper limit in the non-thinking mode.
ThinkPersona: Thinking with Persona Graphs for Faithful Individualized Role-Playing
Yichen Cai (Zhejiang University), Lingyun Sun (Zhejiang University)
GenerationExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextGraphRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Designed and implemented ThinkPersona, a role-playing agent (RPA), by constructing a Persona Graph (individual identity knowledge graph) and generating a Question-Reasoning-Answer (QRA) dataset to train LLMs to perform logical and coherent reasoning and answering in dialogues based on individual identity;
THOR: A Theta-Gamma Hierarchical Oscillatory Reasoning Framework for Multi-hop QA
Ziyang Ling (University of Science and Technology of China), Mingzhai Sun (University of Science and Technology of China)
RetrievalExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose a multi-hop question answering reasoning framework called THOR based on Theta-Gamma hierarchical oscillations, aiming to address the issues of attention decay and error accumulation through the separation of global planning and local retrieval, visual error feedback, and adaptive re-retrieval.
ThreadSumm: Summarization of Nested Discourse Threads Using Tree of Thoughts
Olubusayo Olabisi (Portland State University), Ameeta Agrawal (Portland State University)
GenerationExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes a multi-stage LLM framework called ThreadSumm, which utilizes Tree of Thoughts search to perform multi-level reasoning-based summarization on nested discussion threads;
Threshold Differential Attention for Sink-Free, Ultra-Sparse, and Non-Dispersive Language Modeling
Xingyue Huang (University of Oxford), Tong Zhao (Snap Inc)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: A novel attention mechanism called Threshold Differential Attention (TDA) is studied, aiming to solve the attention sink and dispersion issues of Softmax in long contexts, and to achieve extremely sparse and sink-free attention.
Through the Magnifying Glass: Adaptive Perception Magnification for Hallucination-Free VLM Decoding
Shunqi Mao (University Of Sydney), Weidong Cai (University Of Sydney)
Explainability and InterpretabilityComputational EfficiencyTransformerPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes a decoding method that adaptively magnifies visual regions during inference based on attention, called Perception Magnifier (PM), to reduce visual hallucinations in vision-language models (VLMs);
TIGER: Text-Informed Generalized Enzyme-Reaction Retrieval
Yuhang Zhang (University of Science and Technology of China), Qi Song (University of Science and Technology of China)
RetrievalDrug DiscoveryTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextBiomedical DataRetrieval-Augmented Generation
🎯 What it does: Proposed a text information-driven enzyme-reaction retrieval framework called TIGER, for achieving bidirectional retrieval between enzymes and reactions.
TiKMiX: Efficient Semi-Dynamic Data Mixture via Data Influence for LLM Pre-training
Yifan Wang (ByteDance), Taifeng Wang (ByteDance)
Computational EfficiencyRepresentation LearningData-Centric LearningTransformerLarge Language ModelMixture of ExpertsContrastive LearningText
🎯 What it does: Propose a semi-dynamic data mixing framework called TiKMiX, which combines Group Influence to quantify the domain-level impact, achieving dynamic adjustment of the data proportion during the pre-training phase.
Timely Machine: Awareness of Time Makes Test-Time Scaling Agentic
Yichuan Ma (Shanghai Artificial Intelligence Laboratory), Kai Chen (Shanghai Artificial Intelligence Laboratory)
Autonomous DrivingComputational EfficiencyKnowledge DistillationRepresentation LearningAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AITextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed the concept of Timely Machine, defining test latency as actual wall-clock time, and constructed the TimelyEval benchmark for agentic scenarios. Introduced the Timely-RL reinforcement learning method, enabling LLMs to adaptively adjust their reasoning strategies under a given time budget.
TimeSAF: Towards LLM-Guided Semantic Asynchronous Fusion for Time Series Forecasting
Fan Zhang (Shandong Technology and Business University), Hua Wang (Shandong Technology and Business University)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringTextMultimodalityTime Series
🎯 What it does: Proposed a multi-modal long-term time series forecasting framework called TimeSAF, which utilizes hierarchical asynchronous fusion of text prompts generated by LLMs and numerical sequences.
TInR: Exploring Tool-Internalized Reasoning in Large Language Models
Qiancheng Xu (Hong Kong Polytechnic University), Wenjie Li (Southeast University)
Autonomous DrivingOptimizationFederated LearningExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringMixture of ExpertsTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes a tool internalization reasoning framework called TInR-U, which internalizes tool knowledge into large language models, achieving the unification of tool identification, invocation, and reasoning through a three-stage training process.
Tiny Scales, Great Challenges: The Limits of Multimodal LLMs in Scale Recognition
Jihang Jin (East China University of Science and Technology), Jingping Liu (Sun Yat-sen University)
RecognitionTransformerSupervised Fine-TuningPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Proposed and constructed a multiple-choice benchmark called ScaleBench for evaluating the visual scale recognition capability of multi-modal large language models (MLLMs), covering 13 physical quantities, 33 measuring objects, and 38 units, with a total of 6,574 high-quality samples;
TinyJudge: Unverifiable Constraint Alignment via Lightweight Specialist Ensembles
Yirong Zeng (Harbin Institute of Technology), Ting Liu (Harbin Institute of Technology)
Computational EfficiencyKnowledge DistillationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningMixture of ExpertsTextBenchmark
🎯 What it does: This study investigates the evaluation of large language models (LLMs) in instruction following (IF) with respect to verifiable and non-verifiable constraints, and proposes the TinyJudge framework. This framework replaces the traditional LLM-as-a-judge approach with an integrated lightweight expert model to improve reward accuracy and reduce computational costs.
TLoRA: Task-aware Low Rank Adaptation of Large Language Models
Weicheng Lin (Shenzhen University), Liang-Jie Zhang (Shenzhen University)
Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningMixture of ExpertsDiffusion modelScore-based ModelText
🎯 What it does: Proposed a task-aware low-rank adaptation method called TLoRA, which initializes by using SVD on the product of pre-trained weights and activation covariance, freezes the A matrix, and only trains the B matrix; meanwhile, it adaptively allocates hierarchical ranks and scaling factors based on sensitivity importance.
TLPO: Token-Level Policy Optimization for Mitigating Language Confusion in Large Language Models
Jinho Choo (Samsung SDS), Yeong-Dae Kwon (Samsung SDS)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextMultimodality
🎯 What it does: Propose Token-Level Policy Optimization (TLPO), which alleviates language confusion in large language models by identifying erroneous tokens during the generation process and updating policies only for those tokens.
TLSA: LLM-Guided Text-Label Space Alignment with Contrastive Learning for Generalized Category Discovery
Wenxi Xu (Computer Network Information Center, Chinese Academy of Sciences), Hengshu Zhu (Computer Network Information Center, Chinese Academy of Sciences)
ClassificationRepresentation LearningData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringContrastive LearningText
🎯 What it does: Propose the TLSA framework, which aligns text and labels in a shared semantic space using dual encoders, and completes the task of generalized category discovery through label induction, graph structure refinement, and confidence sampling with LLM.
To Judge or Not to Judge: Can Large Language Models Leverage the Dispute Focus in Legal Judgment?
Luoming Hu (Dalian University of Technology), Zijie Xing (Dalian University of Technology)
ClassificationFederated LearningExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Constructed a full-process Chinese civil judicial dataset called FocalLaw, and designed a legal judgment prediction framework based on dispute focuses called FocalJudge.
To Lie or Not to Lie? Investigating The Biased Spread of Global Lies by LLMs
Zohaib Khan (Fatima Fellowship), Tarek Naous (Stanford University)
GenerationSafty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Investigates the behavior of large language models in generating misinformation under multilingual and multicultural contexts, and constructs a cross-lingual, cross-national misinformation generation template dataset called GlobalLies
To Think or Not to Think: The Hidden Cost of Meta-Training with Excessive CoT Examples
Vignesh Kothapalli (Stanford University), Maziar Sanjabi (LinkedIn AI)
OptimizationExplainability and InterpretabilityComputational EfficiencyMeta LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Studies how to optimize the combination of chain-of-thought (CoT) and in-context learning (ICL) through meta-training in large language models, aiming to enhance the model's reasoning ability on new tasks.
TokDrift: When LLM Speaks in Subwords but Code Speaks in Grammar
Yinxi Li (University of Waterloo), Pengyu Nie (University of Waterloo)
AI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: This paper systematically evaluates the token drift caused by the inconsistency between subword tokenization and programming language syntax in large code LLMs by designing a set of semantics-preserving rewrite rules. It proposes the TOKDRIFT framework and conducts experiments on nine models and three types of tasks (bug fixing, code summarization, code translation), revealing that even minor changes in formatting or naming can significantly alter model outputs.
Token-Level Policy Optimization: Linking Group-Level Rewards to Token-Level Aggregation via sequence-level likelihood
Xingyu Lin (Jilin University), Chenfu Bao (Baidu Inc)
OptimizationComputational EfficiencyRepresentation LearningTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Proposed a token-level policy optimization framework called TEPO to address the training instability caused by sparse token rewards in chain-of-thought (CoT) reasoning.
TokenTiming: A Dynamic Alignment Method for Universal Speculative Decoding Model Pairs
Sibo Xiao (Zhejiang University), Lidan Shou (Zhejiang University)
GenerationComputational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose the TokenTiming algorithm, which utilizes dynamic time warping to align the inference drafts with the target model's heterogeneous vocabulary, achieving lossless speculative decoding acceleration without vocabulary matching.
ToMMeR - Efficient Entity Mention Detection from Large Language Models
Victor Morand (Sorbonne Universite), Benjamin Piwowarski (Sorbonne Universite)
RecognitionComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Designed and trained a lightweight entity mention detector called ToMMeR, which achieves zero-shot entity mention detection by utilizing the hidden states from any early layer of an LLM;
Too Correct to Learn: Reinforcement Learning on Saturated Reasoning Data
Zhenwen Liang (Tencent AI Lab), Dong Yu (Tencent AI Lab)
OptimizationComputational EfficiencyRepresentation LearningTransformerLarge Language ModelReinforcement LearningPrompt EngineeringMixture of ExpertsTextBenchmarkChain-of-Thought
🎯 What it does: Propose the hybrid CUTS framework, which maintains the advantage signal through structured exploration on saturated reasoning data, thereby improving the reasoning performance of large language models (LLMs).
Too Long, Do Re-weighting for Efficient LLM Reasoning Compression
Zhong-Zhi Li (University of Chinese Academy of Sciences), Cheng-Lin Liu (University of Chinese Academy of Sciences)
CompressionComputational EfficiencyData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Propose the TLDR framework, dynamically reweighting System I and System II CoT data to achieve inference length compression;
Too Nice to Tell the Truth: Quantifying Agreeableness-Driven Sycophancy in Role-Playing Language Models
Arya Shah (IIT Gandhinagar), Chaklam Silpasuwanchai (Asian Institute of Technology)
Recommendation SystemExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Investigated the systematic relationship between agreeableness and sycophancy when large language models (LLMs) assume different personality roles.
ToolOmni: Enabling Open-World Tool Use via Agentic learning with Proactive Retrieval and Grounded Execution
Shouzheng Huang (Harbin Institute of Technology (Shenzhen)), Min Zhang (Harbin Institute of Technology (Shenzhen))
Autonomous DrivingOptimizationFederated LearningComputational EfficiencyRobotic IntelligenceTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AITextRetrieval-Augmented Generation
🎯 What it does: Proposes a unified agent framework called ToolOmni, which can achieve efficient tool usage in open-world large tool libraries through active retrieval and execution based on retrieval results;
ToolPRM: Fine-Grained Inference Scaling of Structured Outputs for Function Calling
Jianghao Lin (Shanghai Jiao Tong University), Ying Wen (Shanghai Jiao Tong University)
AI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Studied an inference expansion framework for structured output function calling, proposing a combination of the fine-grained procedural reward model ToolPRM and fine-grained beam search, significantly improving the performance of large models on function calling tasks.
ToolScope: Enhancing LLM Agent Tool Use through Tool Merging and Context-Aware Filtering
Marianne Menglin Liu (Oracle AI), Dan Roth (Oracle AI)
Recommendation SystemOptimizationComputational EfficiencyGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the ToolScope framework, which improves the efficiency of LLM agent tool usage through tool merging and retrieval filtering
Toward Consistent World Models with Multi-Token Prediction and Latent Semantic Enhancement
Qimin Zhong (Microsoft Research Asia), Naipeng Chao (Shenzhen University)
Representation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelMixture of ExpertsContrastive LearningWorld ModelGraphTime SeriesSequential
🎯 What it does: This paper studies the impact of multi-step prediction (MTP) on the internal world models of large language models, and proposes the Latent Semantic Enhancement MTP (LSE-MTP) method to align the latent space and reduce structural hallucinations.
Toward Robust Evaluation for Multilingual Grammatical Error Correction: Can Large Language Models Replace Human References?
Alla Rozovskaya (City University of New York), Dan Roth (University of Pennsylvania)
Data-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed and verified the feasibility of using automatically generated 'closest-gold' (CG) references to evaluate multilingual grammar error correction systems;
Toward Robust In-Context Learning: Leveraging Out-of-distribution Proxies for Target Inaccessible Demonstration Retrieval
Hao Xu (Jilin University), Rui Song (Jilin University)
Domain AdaptationRepresentation LearningData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose a framework called DOPA that achieves robust context learning in the absence of access to the target domain. It uses an OOD agent constructed from source domain instruction-tuned LLMs and non-tuned LLMs to estimate the similarity between samples and the target domain, and combines diversity constraints for example retrieval;
Toward Secure Tuning: Mitigating Security Risks from Instruction Fine-Tuning
Yanrui Du (Harbin Institute of Technology), Bing Qin (Harbin Institute of Technology)
Safty and PrivacyComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextChain-of-Thought
🎯 What it does: To address the security risks caused by instruction fine-tuning, the SWAT method is proposed. During the user optimization phase, it first warms up the safety-robust module (Mods Rob) and then performs standard tuning, thereby reducing the model's vulnerability to attacks.
Towards a Mechanistic Understanding of Large Reasoning Models: A Survey of Training, Inference, and Failures
Yi Hu (Peking University), Liangming Pan (Peking University)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringAuto EncoderTextReview/Survey PaperChain-of-Thought
🎯 What it does: Review the mechanism research of large reasoning models (LRM), systematically divided into three dimensions: training dynamics, reasoning mechanisms, and failure modes, and integrate existing experimental and theoretical findings.
Towards A Scanpath-Conditioned Surprisal Theory: Modeling Reader Information States
Michael Mooney (University of Glasgow), Edmond S. L. Ho (University of Glasgow)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelContrastive LearningText
🎯 What it does: This paper proposes a 'scanpath-conditioned surprisal' model based on readers' scanpaths and memory dynamics, aiming to better capture the nonlinear and memory-limited characteristics of the human reading process;
Towards Efficient and Effective Diffusion Language Model Inference via Semantic-Aware Adaptive Denoising
Fan Li (Northeastern University), Ge Yu (Northeastern University)
Computational EfficiencyAI Code AssistantTransformerLarge Language ModelPrompt EngineeringDiffusion modelText
🎯 What it does: Propose the Ada-DLM framework, which addresses the statistical stagnation problem in diffusion language model inference by actively identifying converged tokens and stopping their computation early;
Towards Explainable Diagnosis: A Self-learned Explanatory Knowledge Base Approach
Dongqi Huang (Key Laboratory of Cognition and Decision Intelligence for Complex Systems, Institute of Automation, Chinese Academy of Sciences), Yubo Chen (Minzu University of China)
Explainability and InterpretabilityTransformerLarge Language ModelTextBiomedical DataElectronic Health RecordsRetrieval-Augmented Generation
🎯 What it does: Propose the SEKAD framework, which self-drivenly constructs an interpretable diagnostic knowledge base using patient medical records, and implements two-stage (differential and definitive) diagnosis on this basis, generating credible and complete diagnostic explanations.
Towards Fast and Accurate Modeling for Cross-Lingual Label Projection
Thang Le (MBZUAI), Thien Huu Nguyen (University of Oregon)
ClassificationDomain AdaptationComputational EfficiencyRepresentation LearningData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringContrastive LearningTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed the PVP framework, which fine-tunes language models using synthesized aligned sequences and span alignment objectives, generates high-quality aligned pairs through a proposal-verification (PV) mechanism, and constructs the EXP benchmark for explicit evaluation.
Towards Fine-Grained and Multi-Granular Contrastive Language-Speech Pre-training
Yifan Yang (Shanghai Jiao Tong University), Xie Chen (Shanghai Jiao Tong University)
RetrievalRepresentation LearningTransformerLarge Language ModelAgentic AIPrompt EngineeringContrastive LearningTextMultimodalityAudio
🎯 What it does: Proposed a large-scale fine-grained speech-text alignment dataset FCaps and a pre-trained model CLSP based on multi-granularity contrastive learning.
Towards Fine-grained Audio Captioning with Multimodal Contextual Fusion
Shunian Chen (Chinese University of Hong Kong), Benyou Wang (Chinese University of Hong Kong)
GenerationRetrievalRepresentation LearningTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsContrastive LearningVideoTextMultimodalityRetrieval-Augmented GenerationAudio
🎯 What it does: Designed and implemented a two-stage multi-modal context fusion pipeline, first using expert models (ASR, OpenMu, GAMA, YamNet, Qwen2.5-VL-72B) to extract speech, music, general audio events, and visual information respectively, and then using large language models (Qwen2.5-32B/QwQ-32B) to fuse them and generate fine-grained audio captions; based on this pipeline, the FusionAudio-1.2M dataset was constructed;
Towards Interpretable Tabular Reasoning: Enhancing LLM Reasoning on Tabular Data with Pre-Constructed Logic Graph
Lirong Gao (Zhejiang University), Sheng Guo (MYbank)
ClassificationExplainability and InterpretabilityTransformerLarge Language ModelContrastive LearningTabularFinance RelatedChain-of-Thought
🎯 What it does: Propose the LogGER framework, which converts the decision paths of tree models into human-readable logic diagrams and uses these diagrams to guide LLMs for step-by-step reasoning, thereby improving the accuracy and interpretability of table data prediction.
Towards Intrinsic Interpretability of Large Language Models: A Survey of Design Principles and Architectures
Yutong Gao (Peking University), Liangming Pan (Peking University)
Explainability and InterpretabilityTransformerLarge Language ModelTextReview/Survey Paper
🎯 What it does: Reviews the methods of intrinsic interpretability for large language models and proposes five design principles.
Towards Mitigating Modality Bias in Vision-Language Models for Temporal Action Localization
Jiaqi Li (University of Warwick), Yu Guan (University of Warwick)
RecognitionObject DetectionExplainability and InterpretabilityComputational EfficiencyConvolutional Neural NetworkTransformerSupervised Fine-TuningMixture of ExpertsVision Language ModelContrastive LearningVideoTextMultimodality
🎯 What it does: Propose the ActionVLM framework, which utilizes vision-centric multi-modal fusion to specifically address the modality bias problem in temporal action localization.
Towards Order Fairness: Mitigating LLMs Order Sensitivity through Dual Group Advantage Optimization
Xu Chu (Peking University), Weiping Li (Peking University)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Significantly improved the model's accuracy and stability when facing different input orders by implementing reinforcement learning based on dual-group advantage on the LLM.
Towards Privacy-Preserving Large Language Model: Text-free Inference Through Alignment and Adaptation
Jeongho Yoon (Korea University), Heuiseok Lim (Korea University)
Domain AdaptationFederated LearningSafty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextBiomedical DataBenchmark
🎯 What it does: Propose Privacy-Preserving Fine-Tuning (PPFT), achieving end-to-end text inference without transmission at the LLM side, utilizing client-side encoding + k-Pooling + noise injection, transmitting only embedded vectors to the server to complete inference and domain adaptation;
Towards Robust Real-World Spreadsheet Understanding with Multi-Agent Multi-Format Reasoning
Houxing Ren (CUHK MMLab), Hongsheng Li (CUHK MMLab)
Federated LearningData-Centric LearningAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringTextMultimodalityTabularBenchmarkFinance RelatedRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed a two-stage multi-agent framework called SpreadsheetAgent for understanding spreadsheets, processing complex spreadsheet data through a step-by-step reading and reasoning approach.
Towards Scalable Lifelong Knowledge Editing with Selective Knowledge Suppression
Dahyun Jung (Korea University), Heuiseok Lim (Korea University)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the LightEdit framework for lifelong knowledge editing without training model parameters.
Towards Stable and Effective Reinforcement Learning for Mixture-of-Experts
Di Zhang (Microsoft Research), Furu Wei (Microsoft Research)
TransformerReinforcement LearningMixture of ExpertsText
🎯 What it does: This paper addresses the reward collapse problem in Mixture-of-Experts (MoE) models during offline reinforcement learning (RLVR) training by proposing the Router-Shift Policy Optimization (RSPO) method. It utilizes the router-shift ratio to softly adjust the importance ratio weights, thereby reducing the impact of distribution drift caused by routing changes on optimization.
Towards Trustworthy Smart Contract Synthesis: A Multi-Agent Framework with Lean-Based Verification
Bowei Zhang (Renmin University of China), Qi Qi (Renmin University of China)
Autonomous DrivingOptimizationFederated LearningSafty and PrivacyExplainability and InterpretabilityComputational EfficiencyMeta LearningDrug DiscoveryAI Code AssistantReinforcement Learning from Human FeedbackNeural Architecture SearchProtein Structure PredictionTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringMixture of ExpertsTextTabularSequentialBenchmarkFinance RelatedRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed and implemented the LeVer framework, combining large language model generation, Lean automatic formalization, and attack simulation to achieve trustworthy smart contract synthesis.
TOWER+: Bridging Generality and Translation Specialization in Multilingual LLMs
Ricardo Rei, Andre Martins
Recommendation SystemOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringMixture of ExpertsTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed the TOWER+ post-training scheme, aiming to achieve a balance between machine translation and general language capabilities for multilingual LLMs, with training scales covering 2B, 9B, and 72B models.
TPA: Next Token Probability Attribution for Detecting Hallucinations in RAG
Pengqian Lu (University of Technology Sydney), Guangquan Zhang (University of Technology Sydney)
Anomaly DetectionExplainability and InterpretabilityComputational EfficiencyTransformerSupervised Fine-TuningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: For Retrieval-Augmented Generation (RAG), the TPA framework is proposed, which decomposes the probability of each generated token into seven sources (query, retrieved context, historical tokens, self tokens, FFN, final LayerNorm, initial embedding), and constructs features for detecting hallucination by aggregating attribution through POS tags.
TPS-Bench: Evaluating AI Agents’ Tool Planning & Scheduling Abilities in Compounding Tasks
Hanwen Xu (Shanghai Jiao Tong University), Zhijie Deng (Shanghai Jiao Tong University)
TransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the TPS-Bench benchmark to evaluate the tool planning and scheduling capabilities of LLM agents in multi-tool composition tasks; collect 200 composite tasks composed of 141 MCP tools, set two difficulty levels; use LLM-as-a-judge to evaluate task completion rate, tool selection score, token and time efficiency; conduct experiments on multiple mainstream LLMs (GPT-4o, GLM-4.5, Qwen3-1.7B, etc.); explore reinforcement learning fine-tuning to improve scheduling efficiency.
TRAC: Teacher-Guided Token Reward with Adaptive Calibration for Robust Policy Optimization
Sitong Wu (Chinese University of Hong Kong), Jiaya Jia (Chinese University of Hong Kong)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningContrastive LearningText
🎯 What it does: Propose the TRAC framework, which performs multi-granularity adaptive calibration on token-level rewards generated by the teacher model, to filter noise and enhance LLM inference performance.
TRACE: A Corpus of Team Creative Discussions
Yixuan Jiang (Zhejiang University), Luning Sun (University of Cambridge)
Data-Centric LearningTransformerSupervised Fine-TuningContrastive LearningTextReview/Survey PaperRetrieval-Augmented Generation
🎯 What it does: This paper creates the TRACE dataset and uses NLP methods to quantify the team discussion process, exploring how team composition affects creative outcomes.
TRACE: Evaluating Execution Efficiency of LLM-Based Code Translation
Zhihao Gong (Peking University), Dan Hao (Peking University)
Computational EfficiencyAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Evaluate the execution efficiency of LLMs in code translation, construct the TRACE benchmark, collect 1,000 efficiency-critical tasks, and systematically evaluate 28 LLMs using stress testing and functional testing.
TRACE: Traversal Retrieval-Augmented Chain of Evidence for Document Understanding
Liqi He, Ping Wang (Wuhan University)
RetrievalExplainability and InterpretabilityComputational EfficiencyRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodalityGraphBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed the TRACE framework (Retrieval + Reasoning Chain via Dual-layer Graphs and Adaptive Path Tracking) and constructed M5BookVQA, a multi-hop, multi-modal, cross-page book question answering benchmark.
Tracing Logit Trajectories Across Layer Depth: Dataset-Level Explainability for Language Models
Jeesu Jung (KAIST), Sangkeun Jung (Chungnam National University)
Safty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Propose the Logit-Trajectory framework, which tracks the updates of logit vectors across layers of a language model and aggregates them at the dataset level to obtain reproducible deep trajectory patterns;
Tracing the Roots: A Multi-Agent Framework for Uncovering Data Lineage in Post-Training LLMs
Yu Li (University of Science and Technology of China), Lijun Wu (Shanghai Artificial Intelligence Laboratory)
OptimizationExplainability and InterpretabilityData-Centric LearningGraph Neural NetworkTransformerLarge Language ModelAgentic AIPrompt EngineeringTextGraphBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper constructs a multi-agent framework to automatically extract and reconstruct the evolution graph of LLM datasets from documents, subsequently analyzing data redundancy and benchmark contamination from topological, cross-domain, and temporal perspectives, and then building a journey-oriented dataset based on this spectrum graph, significantly enhancing semantic diversity.
Traffic-R1: Reinforced LLMs Bring Human-Like Reasoning to Traffic Signal Control Systems
Xingchen Zou (Hong Kong University of Science and Technology), Yuxuan Liang (Hong Kong University of Science and Technology)
Autonomous DrivingOptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextTabularTime Series
🎯 What it does: Proposed Traffic-R1, a 3B-parameter reinforcement learning large language model (LLM), capable of achieving zero-shot, efficient traffic signal control on any traffic network, supporting human-level reasoning and multi-intersection collaboration;
Training LLMs for Divide-and-Conquer Reasoning Elevates Test-Time Scalability
Xiao Liang (University of California, Los Angeles), Weizhu Chen (Microsoft)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningData-Centric LearningMeta LearningAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextSequentialBenchmarkChain-of-Thought
🎯 What it does: Propose an end-to-end reinforcement learning framework (DAC-RL), which simultaneously learns problem division and subproblem solving during the post-training phase of LLMs, enabling the model to reason in a Divide-and-Conquer (DAC) manner.
Trait Activation in Silicon: A Situation-Aware Framework for Psychologically Grounded Role-Playing
Zuolong Li (University Of Science And Technology Of China), Wenbo Zhou (University Of Science And Technology Of China)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringMixture of ExpertsText
🎯 What it does: Propose the PD-LLM framework, achieving dynamic adaptation of psychological traits in role language models through bipolar LoRA and a context-aware module.
Trajectory Signatures of Deception in Large Language Models
Viraaji Mothukuri (Kennesaw State University), Reza M. Parizi (Kennesaw State University)
ClassificationExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper studies the geometric characteristics of different types of deceptive behaviors (such as persuasive, strategic, instructive, and fabrication) in the hidden state space by tracking the activation trajectories of large language models during the reasoning process.
Trajectory2Task: Training Robust Tool-Calling Agents with Synthesized Yet Verifiable Data for Complex User Intents
Ziyi Wang (Northeastern University), Dakuo Wang (Northeastern University)
Data SynthesisAutonomous DrivingRobotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringGenerative Adversarial NetworkTextSequential
🎯 What it does: Proposed the Trajectory2Task data generation pipeline, which can generate verifiable complex user intent tasks from executable tool call trajectories, and perform supervised fine-tuning of LLMs based on these trajectories.