ACL 2025 Papers — Page 16
Annual Meeting of the Association for Computational Linguistics · 1699 papers
The Mirage of Model Editing: Revisiting Evaluation in the Wild
Wanli Yang (Institute of Computing Technology Chinese Academy of Sciences), Xueqi Cheng (Institute of Computing Technology Chinese Academy of Sciences)
Large Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Developed the QAEdit benchmark and the WILD evaluation framework to re-evaluate the effectiveness of model editing methods in real-world question-answering tasks.
The Nature of NLP: Analyzing Contributions in NLP Papers
Aniket Pramanick (Technische Universität Darmstadt), Iryna Gurevych (Technische Universität Darmstadt)
ClassificationRecognitionTransformerLarge Language ModelPrompt EngineeringTextReview/Survey Paper
🎯 What it does: Created a classification system for NLP contribution types and constructed the NLPContributions dataset, which annotated and automatically identified contribution statements in abstracts of nearly 2,000 ACL papers. Subsequently, the dataset was used to conduct a longitudinal analysis of contribution trends across 28,937 papers.
The Noisy Path from Source to Citation: Measuring How Scholars Engage with Past Research
Hong Chen (University of Michigan), David Jurgens (University of Michigan)
TransformerLarge Language ModelText
🎯 What it does: Built a large-scale computational pipeline to automatically assess citation fidelity on 13,000,000 pairs of cited sentences, revealing systematic patterns of information loss and misrepresentation in academic citations.
The Role of Abstract Representations and Observed Preferences in the Ordering of Binomials in Large Language Models
Zachary Nicholas Houghton (University of California, Davis), Emily Morgan (University of California, Davis)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: Investigate whether large language models can learn abstract preferences for binomial orderings from training data, rather than merely replicating observed frequencies.
The Role of Deductive and Inductive Reasoning in Large Language Models
Chengkun Cai (University of Edinburgh), Lei Li (University of Washington)
Computational EfficiencyTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: Proposed the DID framework, which enhances the reasoning capabilities of LLMs by dynamically fusing inductive and deductive reasoning, and using dual metrics to evaluate task complexity.
The Role of Visual Modality in Multimodal Mathematical Reasoning: Challenges and Insights
Yufang Liu (East China Normal University), Xunliang Cai (Meituan Inc)
Representation LearningTransformerLarge Language ModelVision Language ModelContrastive LearningMultimodalityBenchmark
🎯 What it does: Investigate the practical role of the visual modality in multimodal math reasoning, propose the HC-M3D benchmark dataset, and systematically evaluate the robustness of existing models under image perturbations.
The Task Shield: Enforcing Task Alignment to Defend Against Indirect Prompt Injection in LLM Agents
Feiran Jia (Pennsylvania State University), Anna Squicciarini (Pennsylvania State University)
Safty and PrivacyTransformerLarge Language ModelAgentic AITextBenchmark
🎯 What it does: Implement real-time defense against indirect prompt injection attacks on LLM agents through a task alignment mechanism and Task Shield
The time scale of redundancy between prosody and linguistic context
Tamar I Regev, Tiago Pimentel (MIT)
TransformerLarge Language ModelTextAudio
🎯 What it does: This study systematically explores the redundancy of information duration between prosody and lexical context, using mutual information (MI) to measure the correlation between a word's prosody features and its past and future contexts;
The TIP of the Iceberg: Revealing a Hidden Class of Task-in-Prompt Adversarial Attacks on LLMs
Sergey Berezin (Télécom SudParis Institut Polytechnique de Paris), Noel Crespi (Télécom SudParis Institut Polytechnique de Paris)
Adversarial AttackLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Proposed and evaluated a new class of task-internal prompt (TIP) adversarial attacks, demonstrating their ability to bypass the security mechanisms of multiple LLMs.
The Tug of War Within: Mitigating the Fairness-Privacy Conflicts in Large Language Models
Chen Qian (Renmin University of China), Jing Shao (Shanghai Artificial Intelligence Laboratory)
Safty and PrivacyExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Propose a post-processing method called SPIN, which addresses the fairness-privacy trade-off in LLMs during SFT by identifying and suppressing neurons that simultaneously affect fairness and privacy.
The UD-NewsCrawl Treebank: Reflections and Challenges from a Large-scale Tagalog Syntactic Annotation Project
Angelina Aspra Aquino (University of the Philippines Diliman), Elsie Marie T. Or (University of the Philippines Diliman)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: Constructed and released the largest Tagalog UD treebank, UD-NEWSCRAWL, containing 15.6k sentences and 360.8k tokens, using manual annotation combined with semi-automated quality control; also provided a baseline model based on Transformer.
Theorem Prover as a Judge for Synthetic Data Generation
Joshua Ong Jun Leang (University of Edinburgh), Shay B. Cohen (University of Edinburgh)
Data SynthesisTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmarkChain-of-Thought
🎯 What it does: Built a complete workflow utilizing the theorem prover (Lean) to verify intermediate reasoning steps, incorporating automatic formalization, iterative correction, and a reinforcement learning (RLTPF) feedback mechanism to generate high-quality mathematical reasoning data.
TheoremExplainAgent: Towards Video-based Multimodal Explanations for LLM Theorem Understanding
Max Ku (University Of Waterloo), Wenhu Chen (University Of Waterloo)
Explainability and InterpretabilityLarge Language ModelAgentic AIVideoTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose an agent-based long video reasoning framework called TheoremExplainAgent for generating multimodal theorem explanation videos.
Theoretical Analysis of Hierarchical Language Recognition and Generation by Transformers without Positional Encoding
Daichi Hayakawa (University of Tokyo), Issei Sato (University of Tokyo)
RecognitionGenerationTransformerText
🎯 What it does: Demonstrate that Transformer can efficiently recognize and generate hierarchical languages such as Dyck k and Shuffle-Dyck k without absolute position encoding, using causal masking and start tokens; provide conditional generation of pseudo-start signals when start tokens are missing; conduct experiments on layer normalization positions;
Theoretical Guarantees for Minimum Bayes Risk Decoding
Yuki Ichihara (Nara Institute of Science and Technology), Eiji Uchibe (Advanced Telecommunications Research Institute International)
Optimization
🎯 What it does: Analyze the theoretical convergence of minimum Bayes risk (MBR) decoding, providing high-probability and expected regret upper bounds, and compare it with maximum a posteriori (MAP) decoding.
Theory of Mind in Large Language Models: Assessment and Enhancement
Ruirui Chen, Cheston Tan (Institute of High Performance Computing)
TransformerSupervised Fine-TuningPrompt EngineeringVision Language ModelMultimodalityReview/Survey PaperBenchmark
🎯 What it does: This paper reviews the evaluation and enhancement methods of large language models (LLMs) in theory of mind (ToM), systematically organizes the latest story-based and multimodal benchmarks, and summarizes various enhancement strategies based on prompt engineering and fine-tuning;
Think&Cite: Improving Attributed Text Generation with Self-Guided Tree Search and Progress Reward Modeling
Junyi Li (National University of Singapore), Hwee Tou Ng (National University of Singapore)
GenerationTransformerLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: Propose the Think&Cite framework, achieving multi-step citation-aware text generation through self-guided Monte Carlo Tree Search (SG-MCTS) and progress reward modeling.
THOR-MoE: Hierarchical Task-Guided and Context-Responsive Routing for Neural Machine Translation
Yunlong Liang (Tencent Inc), Jie Zhou (Tencent Inc)
GenerationTransformerMixture of ExpertsText
🎯 What it does: Propose a hierarchical task-guided and context-responsive routing sparse expert (THOR-MoE) framework for neural machine translation
TiC-LM: A Web-Scale Benchmark for Time-Continual LLM Pretraining
Jeffrey Li (University of Washington), Fartash Faghri (Apple)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposed TiC-LM, a benchmark for time-continuous pretraining of large language models, covering 114 months of Common Crawl data and multi-domain dynamic evaluations;
TigerLLM - A Family of Bangla Large Language Models
Nishat Raihan (George Mason University), Marcos Zampieri (George Mason University)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Developed the TigerLLM series of Bangla language models and open-sourced a high-quality Bangla-TextBook corpus and 100k instruction-response pairs (Bangla-Instruct).
Time-MQA: Time Series Multi-Task Question Answering with Context Enhancement
Yaxuan Kong (University of Oxford), Qingsong Wen (University of Oxford)
ClassificationAnomaly DetectionExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningTime SeriesBenchmarkFinance Related
🎯 What it does: Proposed the Time-MQA framework, unifying multiple time series tasks (forecasting, imputation, anomaly detection, classification, open-ended reasoning) into natural language question answering; constructed the TSQA dataset containing approximately 200k question-answer pairs across 12 domains and 5 tasks; utilized this dataset for continued pre-training and parameter-efficient fine-tuning of LLMs such as Mistral-7B, Llama-3-8B, and Qwen-2.5-7B.
TokAlign: Efficient Vocabulary Adaptation via Token Alignment
Chong Li (Institute of Automation, Chinese Academy of Sciences), Chengqing Zong (Institute of Automation, Chinese Academy of Sciences)
Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: This paper proposes an unsupervised method called TokAlign based on token-token co-occurrence for efficiently replacing and adapting vocabularies in large language models.
Token Prepending: A Training-Free Approach for Eliciting Better Sentence Embeddings from LLMs
Yuchen Fu (Nanjing University), Qing Gu (Nanjing University)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: Propose a training-agnostic Token Prepending (TP) technique that compensates for the missing backward dependencies under the unidirectional attention of decoder-only LLMs by inserting sentence embeddings before the input of each layer, achieving higher-quality sentence embeddings without modifying model parameters.
Tokenisation is NP-Complete
Philip Whittington (ETH Zürich), Tiago Pimentel (ETH Zürich)
🎯 What it does: Demonstrated that both direct and bottom-up tokenization schemes are NP-complete under compression objectives
ToolCoder: A Systematic Code-Empowered Tool Learning Framework for Large Language Models
Hanxing Ding (State Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences), Xueqi Cheng (State Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences)
AI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Reformulate the tool learning task as a code generation task by converting natural language queries into structured Python function templates, then generating and executing runnable code to invoke tools.
ToolHop: A Query-Driven Benchmark for Evaluating Large Language Models in Multi-Hop Tool Use
Junjie Ye (Fudan University), Jiecao Chen (ByteDance)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose the ToolHop dataset for evaluating large language models (LLMs) in multi-hop tool usage scenarios, focusing on their understanding, reasoning, and function calling capabilities.
Top-n\sigma: Eliminating Noise in Logit Space for Robust Token Sampling of LLM
Chenxia Tang (University of Science and Technology of China), Liusheng Huang (University of Science and Technology of China)
GenerationLarge Language ModelText
🎯 What it does: Proposed a sampling method called topnσ based on logit space denoising, which can maintain stable diversity and quality at any temperature
Toward Automatic Discovery of a Canine Phonetic Alphabet
Theron S. Wang (University of Texas at Arlington), Kenny Q. Zhu (University of Texas at Arlington)
ClassificationRecognitionTransformerVideoAudio
🎯 What it does: A set of phonemes and word patterns in canine vocalizations was constructed through iterative algorithms and minimal pair analysis, achieving automatic discovery of the canine vocal alphabet.
Towards a More Generalized Approach in Open Relation Extraction
Qing Wang (Iowa State University), Qi Li (Iowa State University)
ClassificationRepresentation LearningTransformerAuto EncoderContrastive LearningText
🎯 What it does: Proposes a two-stage MixORE framework for open relation extraction (OpenRE) tasks involving a mixture of known and unknown relations, simultaneously achieving classification of known relations and clustering of unknown relations.
Towards Better Evaluation for Generated Patent Claims
Lekang Jiang (University of Cambridge), Stefan Goetz
TransformerContrastive LearningTextBenchmark
🎯 What it does: Proposed an automatic evaluation method for patent claims called PatClaimEval, and constructed the first patent claim evaluation benchmark named Patent-CE;
Towards Better Value Principles for Large Language Model Alignment: A Systematic Evaluation and Enhancement
Bingbing Xu (Renmin University of China), Xiaofeng Meng (Renmin University of China)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This study proposes a systematic framework for evaluating value principles and constructs a hierarchical set of value principles called HiVaP based on this framework to enhance value alignment in large language models.
Towards Building Large Scale Datasets and State-of-the-Art Automatic Speech Translation Systems for 14 Indian Languages
Ashwin Sankar (Nilekani Centre at AI4Bharat), Raj Dabre (Nilekani Centre at AI4Bharat)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningTextAudio
🎯 What it does: This paper constructs the largest Indian language speech translation dataset, BHASAANUVAAD (44,000 hours, 17 million parallel texts), and trains an end-to-end speech translation model for Indian languages, INDICSEAMLESS (SDBA), which outperforms existing state-of-the-art models on multilingual test sets.
Towards Comprehensive Argument Analysis in Education: Dataset, Tasks, and Method
Yupei Ren (East China Normal University), Xiaopeng Bai (East China Normal University)
ClassificationRecognitionTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Constructed the CEAMC-v2 dataset containing 226 high school English argumentative essays, and proposed 14 fine-grained longitudinal and lateral argumentation relation annotation schemes to comprehensively characterize argument structures; conducted experiments on three tasks: argument component identification, relation prediction, and automatic essay scoring on this dataset.
Towards Context-Robust LLMs: A Gated Representation Fine-tuning Approach
Shenglai Zeng (Michigan State University), Hui Liu (Amazon.com)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: Trained a lightweight gated representation fine-tuning method called Grft to enhance the context robustness of LLMs under retrieval-augmented scenarios.
Towards Dynamic Theory of Mind: Evaluating LLM Adaptation to Temporal Evolution of Human States
Yang Xiao (Hong Kong Polytechnic University), Pengfei Liu (Shanghai Jiao Tong University)
TransformerLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: Developed the DYNTOM benchmark to evaluate LLMs' ability to track and reason about human mental states in dynamic contexts.
Towards Economical Inference: Enabling DeepSeek’s Multi-Head Latent Attention in Any Transformer-based LLMs
Tao Ji (Fudan University), Tao Gui (Fudan University)
Computational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Propose a data-efficient fine-tuning framework named MHA2MLA, which migrates a pre-trained multi-head attention (MHA) LLM to a low-rank key-value (KV) cache multi-head latent attention (MLA) architecture, achieving a significant reduction in inference cost;
Towards Effective and Efficient Continual Pre-training of Large Language Models
Jie Chen (Renmin University of China), Ji-Rong Wen (Renmin University of China)
Data SynthesisComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Continuously pre-train Llama-3 (8B) through refined data mixing, curriculum learning, and synthetic high-quality science and code QA to enhance bilingual (Chinese-English) and multidisciplinary scientific reasoning capabilities.
Towards Effective Extraction and Evaluation of Factual Claims
Dasha Metropolitansky (Microsoft Research), Jonathan Larson (Microsoft Research)
TransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Proposes a claim extraction evaluation framework tailored for fact-checking tasks, and develops Claimify, a novel LLM-based claim extraction method.
Towards Enhanced Immersion and Agency for LLM-based Interactive Drama
Hongqiu Wu (Shanghai Jiao Tong University), Hai Zhao (University of Zurich)
TransformerLarge Language ModelAgentic AIPrompt EngineeringText
🎯 What it does: This study explores interactive drama based on large language models (LLMs), enhancing immersion and agency through script generation and character agents.
Towards Fully Exploiting LLM Internal States to Enhance Knowledge Boundary Perception
Shiyu Ni (Chinese Academy of Sciences), Xueqi Cheng (Chinese Academy of Sciences)
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: Investigate how large language models (LLMs) estimate confidence using internal states before and after generation, and propose a confidence calibration method based on consistency called C3
Towards Geo-Culturally Grounded LLM Generations
Piyawat Lertvittayakumjorn (Google), Sunipa Dev (Google)
TransformerLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Explores the effectiveness of retrieval-augmented generation (RAG) and search-enhancement techniques in improving large language models (LLMs) in cross-cultural knowledge and cultural sensitivity. Experiments use multiple-choice cultural question-answering and stereotype avoidance as benchmarks, with human evaluations to measure cultural fluency.
Towards Harmonized Uncertainty Estimation for Large Language Models
Rui Li (Peking University), Zhifang Sui (Peking University)
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: Proposes the CUE framework, which utilizes a lightweight corrector to calibrate uncertainty scores of large language models (LLMs).
Towards LLM-powered Attentive Listener: A Pragmatic Approach through Quantity Self-Repair
Junlin Li (Hong Kong Polytechnic University), Yu-Yin Hsu (Hong Kong Polytechnic University)
GenerationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes two practical methods based on large language models (LLMs) — Q‑Tuning and Q‑Traveling — to enable LLMs to automatically self-correct information quantity in dialogues to comply with Grice's Quantity Maxim, thereby enhancing their ability to 'listen carefully' and engage in human-like conversations.
Towards Multi-dimensional Evaluation of LLM Summarization across Domains and Languages
Hyangsuk Min (Korea Advanced Institute of Science and Technology), Hwanjun Song (Korea Advanced Institute of Science and Technology)
GenerationTransformerLarge Language ModelAgentic AITextBenchmark
🎯 What it does: Proposes a multidimensional, multi-domain, bilingual (English-Chinese) summarization evaluation benchmark called MSumBench, covering domain-specific key facts and multi-agent debate-style human annotations.
Towards Objective Fine-tuning: How LLMs’ Prior Knowledge Causes Potential Poor Calibration?
Ziming Wang (Beihang University), Jianxin Li (Beihang University)
Explainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Studied the calibration distortion caused by prior knowledge in large language models during fine-tuning, and proposed the real-time cognitive-aware calibration framework CogCalib.
Towards Omni-RAG: Comprehensive Retrieval-Augmented Generation for Large Language Models in Medical Applications
Zhe Chen (Shanghai Jiao Tong University), Yu Wang (Shanghai Jiao Tong University)
TransformerLarge Language ModelSupervised Fine-TuningBiomedical DataRetrieval-Augmented Generation
🎯 What it does: This paper constructs a multi-source, multi-structure medical knowledge base called MedOmniKB and proposes the Source Planning Optimisation (SPO) method to help large language models efficiently plan multi-source retrieval in medical question answering.
Towards Reward Fairness in RLHF: From a Resource Allocation Perspective
Sheng Ouyang (Renmin University of China), Yong Liu (Renmin University of China)
Reinforcement Learning from Human FeedbackReinforcement LearningTextBenchmark
🎯 What it does: This paper proposes a reward fairness framework, treating reward bias as a resource allocation problem, and designs two methods: fair regularization and fair coefficient, achieving fair reward models and strategies in two stages of reward learning and reinforcement learning in RLHF.
Towards Robust and Efficient Federated Low-Rank Adaptation with Heterogeneous Clients
Jabin Koo (POSTECH), Jungseul Ok (POSTECH)
ClassificationFederated LearningComputational EfficiencyTransformerSupervised Fine-TuningText
🎯 What it does: Propose LoRA-A2, achieving efficient and robust fine-tuning of large language models in federated learning through alternating freezing of LoRA modules and adaptive rank selection.
Towards Robust ESG Analysis Against Greenwashing Risks: Aspect-Action Analysis with Cross-Category Generalization
Keane Ong (National University of Singapore), Gianmarco Mengaldo (National University of Singapore)
ClassificationAdversarial AttackTransformerLarge Language ModelContrastive LearningTextFinance Related
🎯 What it does: Construct the A3CG dataset and evaluate the performance of supervised models and LLMs in cross-category robust ESG analysis
Towards Robust Universal Information Extraction: Dataset, Evaluation, and Solution
Jizhao Zhu (Key Laboratory of Network Data Science and Technology, Institute of Computing Technology, Chinese Academy of Sciences), Xueqi Cheng (Key Laboratory of Network Data Science and Technology, Institute of Computing Technology, Chinese Academy of Sciences)
Adversarial AttackData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: This paper proposes a benchmark dataset called RUIE-Bench for evaluating the robustness of a universal information extraction (UIE) model, covering 14 types of adversarial perturbations, and comprehensively evaluates existing UIE models, traditional IE models, and LLMs on this dataset; meanwhile, a loss-guided data augmentation (LDA) method is proposed to enhance model robustness with a small number of difficult samples.
Towards Style Alignment in Cross-Cultural Translation
Shreya Havaldar (University of Pennsylvania), Lyle Ungar (University of Pennsylvania)
GenerationDomain AdaptationLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Proposed and implemented a style alignment objective in cross-cultural machine translation, developing the retrieval-based RASTA method to enable translations that better retain the source language's style while adapting to the target language's cultural norms.
Towards Text-Image Interleaved Retrieval
Xin Zhang (Harbin Institute of Technology), Min Zhang (Harbin Institute of Technology)
RetrievalLarge Language ModelVision Language ModelContrastive LearningMultimodalityBenchmark
🎯 What it does: Proposed the text-image interleaved retrieval (TIIR) task, constructed a TIIR benchmark dataset based on WikiHow tutorials, and introduced a Matryoshka Multimodal Embedder (MME) with compressible visual tokens to enhance interleaved retrieval performance.
Towards the Law of Capacity Gap in Distilling Language Models
Chen Zhang (Beijing Institute Of Technology), Yao Hu
Knowledge DistillationTransformerTextBenchmark
🎯 What it does: Proposed and validated the 'capacity gap law,' which states that in language model distillation, the optimal teacher size is linearly related to the student size. Based on this law, we distilled large language models to obtain the efficient model MINIMA and its instruction-tuned version MINICHAT.
Tracing and Dissecting How LLMs Recall Factual Knowledge for Real World Questions
Yiqun Wang (Zhejiang University), Jieping Ye (Alibaba Cloud)
Explainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Propose a 2D interpretable framework (token back-tracing and single token decoding), conducting fine-grained analysis of the process of LLM recalling factual knowledge in real-world problems, and discovering a three-stage reasoning pattern: subject expansion broadcasting → object retrieval re-ranking → conclusion fusion generation.
Tracking Life’s Ups and Downs: Mining Life Events from Social Media Posts for Mental Health Analysis
Minghao Lv (Shanghai Jiao Tong University), Mengyue Wu (University of Texas at Arlington)
ClassificationExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper constructs a social media life event annotation dataset called PsyEvent, trains life event detection and self-state determination models using this dataset, and integrates life event features into early depression risk detection and suicide risk prediction tasks, significantly improving model performance.
TRACT: Regression-Aware Fine-tuning Meets Chain-of-Thought Reasoning for LLM-as-a-Judge
Cheng-Han Chiang (NTU GICE), Michal Lukasik (Google Research)
TransformerSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: Propose a two-stage regression-aware fine-tuning method called TRACT, which combines chain-of-thought (CoT) with regression-aware fine-tuning (RAFT) to enhance the numerical prediction capability of large language models in scoring tasks.
Training Dynamics Underlying Language Model Scaling Laws: Loss Deceleration and Zero-Sum Learning
Andrei Mircea (Mila Quebec AI Institute), Ekaterina Lobacheva (Mila Quebec AI Institute)
OptimizationTransformerLarge Language ModelText
🎯 What it does: Investigate the impact of language model scale on training dynamics, discovering and explaining the phenomena of loss deceleration and zero-sum learning (ZSL).
Training-free LLM Merging for Multi-task Learning
Zichuan Fu (City University of Hong Kong), Xiangyu Zhao (City University of Hong Kong)
TransformerLarge Language ModelText
🎯 What it does: Propose a training-free hierarchical iterative merging method called Hi-Merging, which merges fine-tuned LLMs from different tasks and languages into a single multi-task model;
Transferring Textual Preferences to Vision-Language Understanding through Model Merging
Chen-An Li (National Taiwan University), Hung-yi Lee (National Taiwan University)
TransformerLarge Language ModelVision Language ModelMultimodalityBenchmark
🎯 What it does: Transfer the preference knowledge from the text reward model (RM) to large-scale vision-language models (LVLM) through model merging, constructing a training-free vision-language reward model (VLRM).
Translation and Fusion Improves Cross-lingual Information Extraction
Yang Chen (Georgia Institute of Technology), Alan Ritter (Georgia Institute of Technology)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Proposes the TransFusion framework, which uses machine translation to convert low-resource language text into English, then generates annotations on the translated text, and finally fuses the annotations with the original text to complete the information extraction task.
Tree-KG: An Expandable Knowledge Graph Construction Framework for Knowledge-intensive Domains
Songjie Niu (Tsinghua University), Wenguang Chen (Tsinghua University)
Graph Neural NetworkTransformerLarge Language ModelTextGraphPhysics Related
🎯 What it does: Proposed the Tree-KG framework, which constructs and iteratively expands knowledge graphs by leveraging textual book structures and large language models.
Tree-of-Debate: Multi-Persona Debate Trees Elicit Critical Thinking for Scientific Comparative Analysis
Priyanka Kargupta (University of Illinois at Urbana Champaign), Jiawei Han (University of Illinois at Urbana Champaign)
TransformerLarge Language ModelAgentic AITextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the Tree-of-Debate framework, which uses multi-role LLMs to conduct structured debate trees between papers, generating fine-grained comparative summaries.
Tree-of-Evolution: Tree-Structured Instruction Evolution for Code Generation in Large Language Models
Ziyang Luo (Hong Kong Baptist University), Jing Ma (Hong Kong Baptist University)
GenerationData SynthesisAI Code AssistantTransformerLarge Language ModelText
🎯 What it does: Propose the Tree-of-Evolution framework, which automatically synthesizes high-quality code instruction data using a tree structure and optimization-driven evolutionary methods.
TreeCut: A Synthetic Unanswerable Math Word Problem Dataset for LLM Hallucination Evaluation
Jialin Ouyang (Columbia University)
Data SynthesisTransformerLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: Proposed the TREECUT dataset, which generates an unlimited number of answerable and unanswerable math word problems using a tree structure, and constructs unanswerable instances by removing necessary conditions along paths.
TreeRL: LLM Reinforcement Learning with On-Policy Tree Search
Zhenyu Hou (Tsinghua University), Yuxiao Dong (Tsinghua University)
AI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmark
🎯 What it does: Enhancing the performance of large language models in mathematical and code reasoning tasks by introducing tree search and process supervision in reinforcement learning training.
TRIDENT: Enhancing Large Language Model Safety with Tri-Dimensional Diversified Red-Teaming Data Synthesis
Xiaorui Wu (Wuhan University), Zhuang Li (Royal Melbourne Institute of Technology)
Data SynthesisSafty and PrivacyAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Proposed the TRIDENT red team data generation pipeline, systematically generating red team instructions covering lexical diversity, malicious intent diversity, and jailbreak tactic diversity through role-playing, scenario→role mapping, role expansion, and six decoding strategies; paired with security responses to form two large datasets: TRIDENT-CORE (26,311 examples) and TRIDENT-EDGE (18,773 examples).
TrimLLM: Progressive Layer Dropping for Domain-Specific LLMs
Lanxiang Hu (University of California, San Diego), Hao Zhang (University of California, San Diego)
CompressionDomain AdaptationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Compress large language models and enhance inference speed by layer-wise elimination combined with sparse fine-tuning on domain-specific tasks.
TripCraft: A Benchmark for Spatio-Temporally Fine Grained Travel Planning
Soumyabrata Chaudhuri (IIT Bhubaneswar), Shreya Ghosh (IIT Bhubaneswar)
TransformerLarge Language ModelPrompt EngineeringTextTabularBenchmark
🎯 What it does: Propose the TripCraft travel planning benchmark, which includes 1,000 real-world 3/5/7-day trip queries, incorporating multi-dimensional constraints such as public transportation, activity types, and user personas, and provides fine-grained continuous evaluation metrics for LLM-generated itineraries.
TripleFact: Defending Data Contamination in the Evaluation of LLM-driven Fake News Detection
Cheng Xu (University College Dublin), Nan Yan (Georgia Institute of Technology)
Data-Centric LearningTransformerLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the TripleFact evaluation framework, integrating human adversarial testing (HAPT), real-time network verification (RTW-AV), and entity-agnostic environment (ECVE), to assess the real-world capabilities of large language models in fake news detection, resisting benchmark data contamination (BDC) issues.
TROVE: A Challenge for Fine-Grained Text Provenance via Source Sentence Tracing and Relationship Classification
Junnan Zhu (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)
ClassificationRetrievalLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper constructs the TROVE challenge, which requires tracing each sentence in the target text back to specific sentences in the source text and performing fine-grained classification of their relationships.
Truth Knows No Language: Evaluating Truthfulness Beyond English
Blanca Calvo Figueras (University of Basque Country), Rodrigo Agerri (University of Basque Country)
Explainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextMultimodalityBenchmark
🎯 What it does: Professionally translate the original English TruthfulQA dataset into Basque, Catalan, Galician, and Spanish, and conduct truthfulness evaluations on 12 open-source LLMs (Llama 3, Llama 3.1, Gemma 2, and their various size versions) across five languages using human evaluation, the MC2 multiple-choice metric, and LLM-as-a-Judge assessment.
TST: A Schema-Based Top-Down and Dynamic-Aware Agent of Text-to-Table Tasks
Peiwen Jiang (Shanghai Jiao Tong University), Jinhua Cheng (Shanghai Jiao Tong University)
RecognitionTransformerLarge Language ModelSupervised Fine-TuningAgentic AITextTabularFinance RelatedRetrieval-Augmented Generation
🎯 What it does: Proposes a dynamic perception two-stage text-to-table framework TST based on type recognition, which can accurately extract multi-instance dynamic fields under a known domain schema.
TUMLU: A Unified and Native Language Understanding Benchmark for Turkic Languages
Jafar Isbarov (eiLink R&D Center, Khazar University), Duygu Ataman (New York University)
Prompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: This study constructs a multilingual multitask language understanding benchmark called TUMLU and releases its mini version TUMLU-mini; meanwhile, it systematically evaluates multiple open-source and proprietary large language models using this benchmark.
TUNA: Comprehensive Fine-grained Temporal Understanding Evaluation on Dense Dynamic Videos
Fanheng Kong (Northeastern University), Fuzheng Zhang (Kuaishou Technology)
Large Language ModelVision Language ModelVideoTextMultimodalityBenchmark
🎯 What it does: Constructed a fine-grained temporal sequence annotation dataset called TUNA-1K containing 1000 dense dynamic videos, and proposed the TUNA evaluation framework based on it, which includes two tasks: video captioning and multiple-choice questions.
Tunable LLM-based Proactive Recommendation Agent
Mingze Wang (University of Science and Technology of China), Fuli Feng (Academy of Cyber)
Recommendation SystemTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AITabularSequential
🎯 What it does: Proposes an adjustable LLM-driven proactive recommendation agent (T-PRA), which generates recommendations under real-time user feedback through the Actor-Advisor framework and uses Critic to evaluate long-term rewards, combined with DPO for agent fine-tuning.
Turning Trash into Treasure: Accelerating Inference of Large Language Models with Token Recycling
Xianzhen Luo (Harbin Institute of Technology), Dongliang Xu (Du Xiaoman Science Technology Co Ltd)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Proposes an untrained Token Recycling method that stores candidate tokens and accelerates LLM inference by leveraging adjacency matrices and tree attention.
TWIST: Text-encoder Weight-editing for Inserting Secret Trojans in Text-to-Image Models
Xindi Li (Zhejiang University), Shouling Ji (Zhejiang University)
GenerationSafty and PrivacyAdversarial AttackTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Proposes the TWIST method, which injects a secret backdoor into the text encoder by making minimal rank-one weight edits to the bottleneck MLP layer of the text encoder, without requiring any training or data, thereby triggering the preset image generation in text prompts.
Two Intermediate Translations Are Better Than One: Fine-tuning LLMs for Document-level Translation Refinement
Yichen Dong (Soochow University), Hao Yang (Huawei)
GenerationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: The study refines document-level translation using large language models, employing two intermediate translations (Sent2Sent and Doc2Doc) and further refining document-level translation based on these.
Typology-Guided Adaptation in Multilingual Models
Ndapa Nakashole (University of California San Diego)
ClassificationRecognitionDomain AdaptationTransformerMixture of ExpertsText
🎯 What it does: This paper proposes a morphological index (MoI) based on morphology and constructs a MoI-MoE architecture, routing multilingual models according to morphological structure to improve noun class identification performance for Bantu languages.
UAlign: Leveraging Uncertainty Estimations for Factuality Alignment on Large Language Models
Boyang Xue (Chinese University of Hong Kong), Kam-Fai Wong (Chinese University of Hong Kong)
Representation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: By embedding two uncertainty estimates, confidence and semantic entropy, as knowledge boundary information into prompts, the reward model and policy model are trained, enabling LLMs to more reliably answer known questions and reject unknown ones in knowledge Q&A.
uMedSum: A Unified Framework for Clinical Abstractive Summarization
Aishik Nagar (ASUS Intelligent Cloud Services), Robby T. Tan (ASUS Intelligent Cloud Services)
GenerationTransformerLarge Language ModelPrompt EngineeringTextBiomedical DataElectronic Health RecordsRetrieval-Augmented Generation
🎯 What it does: This paper proposes a unified clinical abstract summarization framework called uMedSum, aiming to enhance the faithfulness and informativeness of summaries while addressing issues of hallucination and information omission.
Unanswerability Evaluation for Retrieval Augmented Generation
Xiangyu Peng (Salesforce Research), Chien-Sheng Wu (Salesforce Research)
TransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Proposed and implemented the UAEval4RAG framework for systematic evaluation of retrieval-augmented generation (RAG) models in handling unanswerable queries, while developing an automated pipeline for generating unanswerable queries and defining six categories of unanswerable requests.
Uncertainty in Causality: A New Frontier
Shaobo Cui (École Polytechnique Fédérale de Lausanne), Boi Faltings (École Polytechnique Fédérale de Lausanne)
Explainability and InterpretabilityLarge Language ModelTextReview/Survey Paper
🎯 What it does: Review causal uncertainty, propose a three-way classification (random, knowledge missing, existence), and evaluate the performance of LLMs in causal reasoning
Uncertainty Propagation on LLM Agent
Qiwei Zhao (University of North Carolina at Chapel Hill), Xujiang Zhao (NEC Labs America)
Explainability and InterpretabilityTransformerLarge Language ModelAgentic AIText
🎯 What it does: Proposes the SAUP framework, which propagates and weights uncertainty at each step in multi-step LLM agents to obtain an overall uncertainty estimation.
Uncertainty-Aware Iterative Preference Optimization for Enhanced LLM Reasoning
Lei Li (Tencent), Zang Li (Tencent)
OptimizationReinforcement Learning from Human FeedbackLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Improve the performance of LLMs in mathematical, coding, and SQL reasoning tasks by iteratively sampling and executing feedback to construct preference pairs, and combining model uncertainty for fine-grained direct preference optimization.
Uncovering the Impact of Chain-of-Thought Reasoning for Direct Preference Optimization: Lessons from Text-to-SQL
Hanbing Liu (Renmin University of China), Jing Zhang (Renmin University of China)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkChain-of-Thought
🎯 What it does: This paper investigates the effectiveness of combining Chain-of-Thought (CoT) with Direct Preference Optimization (DPO) in Text-to-SQL tasks, and verifies the validity of this method through experiments on large-scale language models.
Uncovering Visual-Semantic Psycholinguistic Properties from the Distributional Structure of Text Embedding Space
Si Wu (Northeastern University), Sebastian Bruch (Northeastern University)
Representation LearningTransformerLarge Language ModelText
🎯 What it does: Proposed and verified an unsupervised measure based on the neighborhood stability in semantic embedding space (Neighborhood Stability Measure, NSM) to estimate the imagineability and concreteness of vocabulary from text (especially image captions).
Understanding Common Ground Misalignment in Goal-Oriented Dialog: A Case-Study with Ubuntu Chat Logs
Rupak Sarkar (University of Maryland), Philip Resnik (University of Maryland)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Investigated common ground misalignment in goal-oriented conversations within Ubuntu IRC dialogues and its impact on task success, constructed annotated data, and evaluated the ability of LLMs to detect 'session friction'.
Understanding Cross-Domain Adaptation in Low-Resource Topic Modeling
Pritom Saha Akash (University of Illinois at Urbana-Champaign), Kevin Chen-Chuan Chang (University of Illinois at Urbana-Champaign)
Domain AdaptationRepresentation LearningAuto EncoderGenerative Adversarial NetworkText
🎯 What it does: Study cross-domain adaptation in low-resource topic modeling, proposing the DALTA framework to achieve selective transfer of source domain knowledge.
Understanding Impact of Human Feedback via Influence Functions
Taywon Min (KAIST), Kimin Lee (KAIST)
Explainability and InterpretabilityComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Propose an scalable influence function method to quantify the impact of human feedback on reward models, and use this method to detect label bias and guide labelers to improve strategies.
Understanding In-Context Machine Translation for Low-Resource Languages: A Case Study on Manchu
Renhao Pei (Center for Information and Language Processing, LMU Munich), Hinrich Schuetze
GenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: The study explores the use of large language models (LLMs) for context-based machine translation in the low-resource language Manchu, and systematically evaluates the impact of resources such as dictionaries, parallel examples, grammar books, and chain-of-thought (CoT) on translation performance.
Understanding Large Language Model Vulnerabilities to Social Bias Attacks
Jiaxu Zhao (Eindhoven University of Technology), Mykola Pechenizkiy (A*STAR Centre for Frontier AI Research)
Adversarial AttackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This study systematically evaluates the vulnerability of large language models (LLMs) to social bias attacks, constructs three attack methods (prefix injection, rejection suppression, learning-based attack prompts), and conducts experiments on various mainstream models (LLaMA-2, Falcon, Vicuna, Mistral, Pythia, GPT-3.5, GPT-4).
Understanding Silent Data Corruption in LLM Training
Jeffrey Jian Ma (Harvard University), George Karypis (Amazon Web Services)
OptimizationData-Centric LearningTransformerLarge Language ModelText
🎯 What it does: Systematically conduct experiments and analysis on Silent Data Corruption (SDC) occurring during LLM training in real-world environments, evaluating its impact on submodule computations, gradients, and the final model quality.
Understanding the Dark Side of LLMs’ Intrinsic Self-Correction
Qingjie Zhang (Tsinghua University), Han Qiu (Tsinghua University)
Explainability and InterpretabilityLarge Language ModelText
🎯 What it does: Studied the self-correction failure mechanisms of large language models in the absence of external knowledge, and revealed the failure causes using three explanation methods.
Uni-Retrieval: A Multi-Style Retrieval Framework for STEM’s Education
Yanhao Jia (Nanyang Technological University), Wenqi Fan (Hong Kong Polytechnic University)
RetrievalTransformerPrompt EngineeringVision Language ModelContrastive LearningMultimodality
🎯 What it does: Proposed a multimodal and multi-style STEM education retrieval framework named Uni-Retrieval, and constructed the SER dataset containing 24,000 multimodal samples.
UniCodec: Unified Audio Codec with Single Domain-Adaptive Codebook
Yidi Jiang (National University of Singapore), Haizhou Li (Chinese University of Hong Kong)
CompressionDomain AdaptationTransformerMixture of ExpertsAudio
🎯 What it does: Proposed a unified audio codec called UniCodec, designed to support multi-domain audio data, including speech, music, and sound, using a single codebook.
UniConv: Unifying Retrieval and Response Generation for Large Language Models in Conversations
Fengran Mo (University of Montreal), Meng Jiang (University of Notre Dame)
GenerationRetrievalTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
🎯 What it does: Developed a unified large language model called UniConv, capable of both dialogue retrieval and answer generation.
Unifying Continuous and Discrete Text Diffusion with Non-simultaneous Diffusion Processes
Bocheng Li (University of Science and Technology of China), Linli Xu (University of Science and Technology of China)
GenerationTransformerDiffusion modelText
🎯 What it does: Propose NeoDiff, a unified text diffusion model that integrates discrete and continuous noise control, employing a Poisson process in the forward process and a context-aware time predictor in the backward process.
Unifying Uniform and Binary-coding Quantization for Accurate Compression of Large Language Models
Seungcheol Park (Seoul National University), Dongsoo Lee (NAVER Cloud)
CompressionTransformerTextBenchmark
🎯 What it does: Propose UniQuan F, a unified quantization method that combines the strong optimization capability of UQ with the high expressiveness of BCQ, achieving 3/4-bit high-precision quantization for large language models.
UniICL: An Efficient ICL Framework Unifying Compression, Selection, and Generation
Jun Gao (Soochow University), Wenjie Li (Stepfun)
Computational EfficiencyTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Proposes UniICL, a unified ICL framework that integrates demonstration compression, demonstration selection, and final response generation;