ACL 2025 Papers — Page 3
Annual Meeting of the Association for Computational Linguistics · 1699 papers
Bone Soups: A Seek-and-Soup Model Merging Approach for Controllable Multi-Objective Generation
Guofu Xie (Renmin University of China), Yunsheng Shi (Tencent)
GenerationTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Propose a novel model fusion method called Bone Soup, which first uses multi-objective reinforcement learning to find a series of backbone models, and then dynamically merges them based on user preferences, addressing the Pareto optimality and controllability issues in controllable multi-objective generation.
BOOKCOREF: Coreference Resolution at Book Scale
Giuliano Martinelli (Sapienza University of Rome), Roberto Navigli (Sapienza University of Rome)
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Investigated the book-level coreference resolution task, proposed and implemented an automated annotation pipeline (BookCoref Pipeline), and constructed the first book-level coreference benchmark BOOKCOREF based on this.
BOOKWORLD: From Novels to Interactive Agent Societies for Story Creation
Yiting Ran (Fudan University), Deqing Yang (Fudan University)
GenerationRecurrent Neural NetworkTransformerLarge Language ModelAgentic AIText
🎯 What it does: Construct a multi-agent society based on novels, simulating character interactions and generating coherent stories.
Boosting LLM’s Molecular Structure Elucidation with Knowledge Enhanced Tree Search Reasoning
Xiang Zhuang (Zhejiang University), Huajun Chen (Zhejiang University)
Drug DiscoveryGraph Neural NetworkTransformerLarge Language ModelContrastive LearningBiomedical DataBenchmarkRetrieval-Augmented Generation
🎯 What it does: Developed a knowledge-enhanced Monte Carlo Tree Search framework called K-MSE to improve the reasoning and accuracy of large language models in molecular structure interpretation tasks.
Boosting Long-Context Information Seeking via Query-Guided Activation Refilling
Hongjin Qian, Defu Lian (Gaoling School of Artificial Intelligence Renmin University of China)
RetrievalComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: The ACRE method is proposed for long-text information retrieval tasks, achieving efficient fusion of global and local information through the construction of a dual-layer KV cache and the adoption of query-guided activation filling.
BPP-Search: Enhancing Tree of Thought Reasoning for Mathematical Modeling Problem Solving
Teng Wang (University of Hong Kong), Tao Zhong (Noah's Ark Lab, Huawei)
OptimizationTransformerLarge Language ModelContrastive LearningTextBenchmarkChain-of-Thought
🎯 What it does: This paper proposes a new structured operations research dataset, StructuredOR, and develops the BPP-Search algorithm based on it to improve the accuracy and efficiency of Tree of Thought in solving mathematical modeling problems.
BQA: Body Language Question Answering Dataset for Video Large Language Models
Shintaro Ozaki (Nara Institute of Science and Technology), Taro Watanabe (Nara Institute of Science and Technology)
RecognitionLarge Language ModelSupervised Fine-TuningVision Language ModelVideoTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposed and constructed the BQA (Body Language Question Answering) dataset, converting BoLD video emotion labels into multiple-choice QA, and subsequently using this dataset to evaluate the ability of various VideoLLMs to understand body language emotions.
Bregman Conditional Random Fields: Sequence Labeling with Parallelizable Inference Algorithms
Caio Corro (INSA Rennes, IRISA, Inria, CNRS, Université de Rennes), Joseph Le Roux (Université Sorbonne Paris Nord, CNRS, LIPN)
ClassificationText
🎯 What it does: Propose Bregman Conditional Random Fields (BCRF) and a parallel inference algorithm based on iterative Bregman projections for sequence labeling.
Bridging the Language Gaps in Large Language Models with Inference-Time Cross-Lingual Intervention
Weixuan Wang (Monash University), Alexandra Birch (University of Edinburgh)
Domain AdaptationComputational EfficiencyRepresentation LearningLarge Language ModelText
🎯 What it does: This paper proposes learning a cross-lingual alignment matrix during inference to map the internal representations of low-performance languages into the high-performance language space, thereby improving the performance of large language models on low-resource languages.
BRIGHTER: BRIdging the Gap in Human-Annotated Textual Emotion Recognition Datasets for 28 Languages
Shamsuddeen Hassan Muhammad (Imperial College London), Saif M. Mohammad (National Research Council Canada)
RecognitionTransformerLarge Language ModelTextBenchmark
🎯 What it does: Constructed the BRIGHTER dataset, collecting and multi-label annotating emotional texts in 28 languages, covering low-resource languages and providing intensity levels;
Browsing Like Human: A Multimodal Web Agent with Experiential Fast-and-Slow Thinking
Haohao Luo (Sun Yat-sen University), Yang Deng (Xiaomi Inc)
TransformerLarge Language ModelAgentic AIVision-Language-Action ModelTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Propose a multimodal Web agent called WebExperT, capable of completing complex web interaction tasks based on user instructions.
Browsing Lost Unformed Recollections: A Benchmark for Tip-of-the-Tongue Search and Reasoning
Sky CH-Wang (Columbia University), Rebecca Qian (Patronus AI)
TransformerLarge Language ModelAgentic AIMultimodalityBenchmarkChain-of-Thought
🎯 What it does: This paper constructs the BLUR benchmark, which contains 573 real tip-of-the-tongue (ToT) questions, covering multimodal inputs such as text, images, and audio, as well as multilingual descriptions, requiring models to perform multi-hop reasoning, tool usage, and uncertainty handling.
Building a Long Text Privacy Policy Corpus with Multi-Class Labels
Florencia Marotta-Wurgler (NYU School of Law), David Stein (Northeastern University)
Safty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Constructed a manually annotated corpus covering 149 privacy policies and their referenced documents, containing 64-dimensional multi-class labels.
Building Better: Avoiding Pitfalls in Developing Language Resources when Data is Scarce
Nedjma Ousidhoum (Cardiff University), Saif M. Mohammad (National Research Council Canada)
Data-Centric LearningText
🎯 What it does: Conduct a questionnaire survey on practical and ethical issues in developing NLP resources for low-resource languages, collect quantitative and qualitative feedback from 81 researchers and annotators, and subsequently summarize the main challenges and improvement suggestions.
Bypass Back-propagation: Optimization-based Structural Pruning for Large Language Models via Policy Gradient
Yuan Gao (Wuhan University), Gui-Song Xia (Wuhan University)
OptimizationComputational EfficiencyTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Proposes a structured pruning method that optimizes large language models without backpropagation, utilizing only forward propagation. It prunes model channels, attention heads, and layers by learning a Bernoulli distribution to generate binary masks.
Byte Latent Transformer: Patches Scale Better Than Tokens
Artidoro Pagnoni (University of Washington), Srini Iyer
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: This paper proposes Byte Latent Transformer (BLT), a vocabulary-free large language model based on dynamic byte patches, which can match BPE-based models in scale while achieving significant improvements in inference efficiency and robustness.
CADReview: Automatically Reviewing CAD Programs with Error Detection and Correction
Jiali Chen (South China University of Technology), Li Qing
Anomaly DetectionAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelImageTextMultimodality
🎯 What it does: This paper introduces the CAD program review task, which automatically detects and corrects errors in CAD code to ensure consistency between 3D models and reference images.
CalibraEval: Calibrating Prediction Distribution to Mitigate Selection Bias in LLMs-as-Judges
Haitao Li (Tsinghua University), Yiqun Liu (Tsinghua University)
OptimizationExplainability and InterpretabilityTransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper proposes CalibraEval, a label-free, inference-phase-executable calibration method designed to eliminate selection bias in LLM-as-Judges.
Call for Rigor in Reporting Quality of Instruction Tuning Data
Hyeonseok Moon (Korea University), Heuiseok Lim (Korea University)
Hyperparameter SearchData-Centric LearningTransformerLarge Language ModelText
🎯 What it does: In the paper, the authors investigate how the arbitrariness of hyperparameter selection affects the consistency of conclusions when evaluating the quality of instruction tuning data, and verify this through experiments on two 1K datasets, LIMA and Alpaca-Longest, under various hyperparameter settings.
CaLMQA: Exploring culturally specific long-form question answering across 23 languages
Shane Arora (University of Texas at Austin), Eunsol Choi (New York University)
Data-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Constructed the first Chinese dataset CALMQA spanning 23 languages with 51.7K culturally specific long-form question-answer pairs, and conducted systematic evaluation of seven large language models (LLMs) on this dataset.
CAMI: A Counselor Agent Supporting Motivational Interviewing through State Inference and Topic Exploration
Yizhe Yang (Beijing Institute of Technology), Ee-Peng Lim (Singapore Management University)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper designs and implements a conversational tutoring agent called CAMI based on Motivational Interviewing, utilizing the STAR framework to accomplish client state inference, topic exploration, strategy selection, and response generation.
Can a Single Model Master Both Multi-turn Conversations and Tool Use? CoALM: A Unified Conversational Agentic Language Model
Emre Can Acikgoz (University of Illinois Urbana Champaign), Gokhan Tur (University of Illinois Urbana Champaign)
TransformerLarge Language ModelSupervised Fine-TuningAgentic AITextChain-of-Thought
🎯 What it does: Propose CoALM, a unified conversational language model capable of simultaneously handling multi-turn task-oriented dialogue (TOD) and complex tool calls (LA)
Can Community Notes Replace Professional Fact-Checkers?
Nadav Borenstein (University of Copenhagen), Isabelle Augenstein (University of Copenhagen)
TransformerLarge Language ModelText
🎯 What it does: This paper automatically annotates 1.5M Twitter/X community notes using a language model, generating a large-scale dataset containing attributes such as theme, source type, and whether professional fact-checking is cited. Based on this dataset, the paper analyzes the dependency and characteristics of community notes on professional fact-checking.
Can External Validation Tools Improve Annotation Quality for LLM-as-a-Judge?
Arduin Findeis (University of Cambridge), Tom Gunter (Apple)
TransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This study proposes an scalable agentic framework that integrates external validation tools (web search and code execution) into the LLM-as-a-Judge system to improve the quality of bilateral preference annotation in challenging domains such as long-form facts, programming, and mathematics.
Can Graph Descriptive Order Affect Solving Graph Problems with LLMs?
Yuyao Ge (Institute Of Computing Technology Chinese Academy Of Sciences), Xueqi Cheng (Institute Of Computing Technology Chinese Academy Of Sciences)
ClassificationTransformerLarge Language ModelPrompt EngineeringGraphBenchmarkChain-of-Thought
🎯 What it does: Studied the impact of graph description order on the performance of large language models (LLMs) in solving graph problems, and systematically evaluated the effectiveness of four graph description orders across six graph tasks.
Can Graph Neural Networks Learn Language with Extremely Weak Text Supervision?
Zihao Li (University Of Illinois Urbana Champaign), Jiawei Han (University Of Illinois Urbana Champaign)
ClassificationRepresentation LearningGraph Neural NetworkLarge Language ModelPrompt EngineeringContrastive LearningTextMultimodalityGraph
🎯 What it does: Proposed Morpher, a multimodal prompt learning framework, which aligns the semantic embedding spaces of pre-trained graph neural networks (GNN) and large language models (LLM) under extremely weak text supervision, achieving few-shot learning, cross-domain transfer, and zero-shot classification.
Can Indirect Prompt Injection Attacks Be Detected and Removed?
Yulin Chen (National University of Singapore), Bryan Hooi (National University of Singapore)
Anomaly DetectionAdversarial AttackTransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper studies how to detect and eliminate indirect prompt injection attacks, constructing a benchmark dataset targeting various attack purposes such as deception, advertising, and propaganda, and systematically evaluates the effectiveness of detection models and two elimination methods.
Can Knowledge Graphs Make Large Language Models More Trustworthy? An Empirical Study Over Open-ended Question Answering
Yuan Sui (National University of Singapore), Bryan Hooi (National University of Singapore)
Explainability and InterpretabilityTransformerLarge Language ModelTextGraphBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Introduces a new open-domain knowledge graph question answering benchmark OKGQA (and its noisy version OKGQA-P), and systematically evaluates the reliability of large language models in open-domain question answering, as well as the impact of KG enhancement on hallucinations.
Can Language Models Reason about Individualistic Human Values and Preferences?
Liwei Jiang (University of Washington), Yejin Choi (Stanford University)
TransformerLarge Language ModelPrompt EngineeringTextTabular
🎯 What it does: The study investigates whether language models can infer an individual's value judgments in new scenarios based on their expressed value statements, and proposes the INDIEVALUECATALOG dataset and the INDIEVALUEREASONER model grounded in the World Values Survey (WVS).
Can Language Models Replace Programmers for Coding? REPOCOD Says ‘Not Yet’
Shanchao Liang (Purdue University), Lin Tan (Purdue University)
AI Code AssistantLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes the REPOCOD benchmark to evaluate the performance of large language models in real-world repository-level code generation tasks;
Can Large Language Models Accurately Generate Answer Keys for Health-related Questions?
Davis Bartels, Dina Demner-Fushman (National Library of Medicine)
TransformerLarge Language ModelPrompt EngineeringBiomedical DataBenchmark
🎯 What it does: Investigate the ability of large language models to automatically generate answer nuggets in medical question answering and evaluate their alignment with human-generated nuggets.
Can Large Language Models Detect Errors in Long Chain-of-Thought Reasoning?
Yancheng He (Alibaba Group), Bo Zheng (Alibaba Group)
TransformerLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: Create the DeltaBench dataset and conduct a fine-grained evaluation of long Chain-of-Thought (CoT) generated by o1-like models, exploring the critical error detection capability of LLMs.
Can Large Language Models Understand Internet Buzzwords Through User-Generated Content
Chen Huang (Sichuan University), Jiancheng Lv (Sichuan University)
GenerationTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: Construct a task for generating definitions of Chinese internet slang, propose a definition generation framework based on user-generated content (UGC), and collect and build the first Chinese slang dataset, CHEER.
Can LLM Watermarks Robustly Prevent Unauthorized Knowledge Distillation?
Leyi Pan (Tsinghua University), Philip S. Yu (University of Illinois at Chicago)
Knowledge DistillationAdversarial AttackTransformerLarge Language ModelText
🎯 What it does: Systematically study the robustness of LLM watermarks in preventing unauthorized knowledge distillation, proposing three de-watermarking methods: pre-distillation (unoriented rewriting, oriented rewriting) and post-distillation (inference-time watermark neutralization).
Can LLMs Deceive CLIP? Benchmarking Adversarial Compositionality of Pre-trained Multimodal Representation via Text Updates
Jaewoo Ahn (Seoul National University), Gunhee Kim (Seoul National University)
Representation LearningAdversarial AttackLarge Language ModelSupervised Fine-TuningImageVideoMultimodalityBenchmarkAudio
🎯 What it does: This paper proposes the Multimodal Adversarial Compositionality (MAC) benchmark, which uses large language models to generate deceptive text to expose combination vulnerabilities in pre-trained multimodal representations.
Can LLMs Evaluate Complex Attribution in QA? Automatic Benchmarking using Knowledge Graphs
Nan Hu (Southeast University), Jeff Z. Pan (Monash University)
TransformerLarge Language ModelSupervised Fine-TuningTextGraphBenchmark
🎯 What it does: Constructed a large-scale Complex Attributed Question Answering (CAQA) benchmark by automatically generating fine-grained attribution categories and multi-level reasoning complexity question-answer samples using knowledge graphs, evaluating the performance of LLMs in attribution assessment and further enhancing their capabilities.
Can LLMs Generate High-Quality Test Cases for Algorithm Problems? TestCase-Eval: A Systematic Evaluation of Fault Coverage and Exposure
Zheyuan Yang, Yilun Zhao (HKUST)
AI Code AssistantTransformerLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: Propose the TestCase-Eval benchmark to evaluate the Fault Coverage and Fault Exposure of LLMs in generating test cases for algorithm problems.
Can LLMs Ground when they (Don’t) Know: A Study on Direct and Loaded Political Questions
Clara Lachenmaier (Bielefeld University), Sina Zarrieß (Bielefeld University)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper investigates the grounding ability of large language models (LLMs) in political contexts, exploring how they identify and reject loaded questions containing false premises.
Can LLMs Help Uncover Insights about LLMs? A Large-Scale, Evolving Literature Analysis of Frontier LLMs
Jungsoo Park (Georgia Institute of Technology), Alan Ritter (Georgia Institute of Technology)
Data-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: This study proposes a semi-automated literature analysis workflow, leveraging large language models (LLMs) to extract experimental records of cutting-edge LLMs from arXiv papers, constructing the LLMEVALDB dataset, and performing automated performance analysis of prompting methods (e.g., CoT, ICL) based on this dataset.
Can LLMs Identify Critical Limitations within Scientific Research? A Systematic Evaluation on AI Research Papers
Zhijian Xu (Yale University), Arman Cohan (Yale University)
Large Language ModelAgentic AITextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed and implemented the LIMITGEN benchmark to systematically evaluate large language models (LLMs) in identifying research limitations in scientific papers, and designed two subsets: LIMITGEN-Syn (synthetically constructed papers with controlled defects) and LIMITGEN-Human (limitations descriptions from real ICLR 2025 reviews).
Can LLMs Reason About Program Semantics? A Comprehensive Evaluation of LLMs on Formal Specification Inference
Thanh Le-Cong (University of Melbourne), Toby Murray (University of Melbourne)
AI Code AssistantLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: This paper designs and implements FormalBench, a benchmark to evaluate the reasoning capabilities of large language models (LLMs) in generating complete and consistent program formal specifications (JML).
Can LLMs Simulate L2-English Dialogue? An Information-Theoretic Analysis of L1-Dependent Biases
Rena Gao (University of Melbourne), Jey Han Lau (University of Melbourne)
GenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper evaluates the performance of large language models (LLMs) in simulating non-native English conversations with different first language (L1) backgrounds, exploring the impact of L1 on second language (L2) output.
Can LLMs Understand Unvoiced Speech? Exploring EMG-to-Text Conversion with LLMs
Payal Mohapatra (Northwestern University), Qi Zhu (Northwestern University)
Representation LearningConvolutional Neural NetworkRecurrent Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextTime SeriesBiomedical Data
🎯 What it does: This paper explores using large language models (LLMs) to directly convert silent electromyography signals (non-vocalized EMG) into text, and proposes a trainable EMG adapter module;
Can MLLMs Understand the Deep Implication Behind Chinese Images?
Chenhao Zhang (Huazhong University of Science and Technology), Shiwen Ni (Shenzhen Institutes of Advanced Technology Chinese Academy of Sciences)
TransformerPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Propose CII-Bench to evaluate the understanding of deep meaning in Chinese images by multimodal large language models
Can Multimodal Large Language Models Understand Spatial Relations?
Jingping Liu (East China University of Science and Technology), Tong Ruan (Fudan University)
Large Language ModelPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Built a bbox-free, multi-perspective, spatial relationship multimodal question answering benchmark called SpatialMQA based on COCO2017.
Can Third Parties Read Our Emotions?
Jiayi Li (Pennsylvania State University), Sarah Rajtmajer (Pennsylvania State University)
RecognitionLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper compares the consistency between third-party (human and LLM) emotion annotations and author-reported emotion labels through human subject experiments, and examines the impact of annotator-author similarity in demographic features (such as race, gender, age) and prompting methods on annotation quality.
Can Uniform Meaning Representation Help GPT-4 Translate from Indigenous Languages?
Shira Wein (Amherst College)
TransformerLarge Language ModelPrompt EngineeringTextGraphRetrieval-Augmented Generation
🎯 What it does: This paper evaluates the effectiveness of embedding Uniform Meaning Representation (UMR) semantic graphs into GPT-4 prompts for translating extremely low-resource indigenous languages (Navajo, Arapaho, Kukama) to English. The authors designed and compared four prompt strategies: zero-shot, zero-shot + UMR, five-shot, and five-shot + UMR, and conducted automatic evaluation using chrF and BERTScore on 1,017 test sentences.
Can Vision-Language Models Evaluate Handwritten Math?
Oikantik Nath (Indian Institute of Technology Madras), Mitesh M Khapra
Vision Language ModelImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Designed and released the FERMAT benchmark to evaluate Vision-Language Models' capabilities in detecting, localizing, and correcting errors in handwritten mathematical answers, covering four error categories: calculation, conceptual, symbolic, and formatting errors.
Can We Further Elicit Reasoning in LLMs? Critic-Guided Planning with Retrieval-Augmentation for Solving Challenging Tasks
Xingxuan Li (MiroMind), Lidong Bing (MiroMind)
AI Code AssistantTransformerLarge Language ModelTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Designed and implemented the CR-Planner framework, which utilizes specialized critics to plan subgoals and execution between reasoning and retrieval, addressing complex tasks requiring deep reasoning and domain-specific knowledge retrieval.
Can we Retrieve Everything All at Once? ARM: An Alignment-Oriented LLM-based Retrieval Method
Peter Baile Chen (MIT), Dan Roth (University of Pennsylvania & Oracle AI)
RetrievalTransformerLarge Language ModelTextMultimodalityTabular
🎯 What it does: Proposes an alignment-oriented LLM retrieval method called ARM, capable of retrieving all data objects required to address complex open-domain questions in a single pass.
Can You Really Trust Code Copilot? Evaluating Large Language Models from a Code Security Perspective
Yutao Mou (Peking University), Wei Ye (Peking University)
Safty and PrivacyAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Proposes the CoV-Eval multi-task benchmark to evaluate the code security of large language models (LLMs), and develops the VC-Judge evaluation model to automatically identify security vulnerabilities in generated code.
Can’t See the Forest for the Trees: Benchmarking Multimodal Safety Awareness for Multimodal LLMs
Wenxuan Wang (Renmin University of China), Zhaopeng Tu (Tencent)
Safty and PrivacyTransformerLarge Language ModelPrompt EngineeringVision Language ModelMultimodalityBenchmark
🎯 What it does: Proposed the MMSafeAware benchmark, using 1500 image-text pairs to evaluate the safety awareness of multimodal large language models, covering subsets of safety and over-safety across 29 safety scenarios.
Capability Salience Vector: Fine-grained Alignment of Loss and Capabilities for Downstream Task Scaling Law
Qiming Ge (Shanghai AI Laboratory), Kai Chen (Shanghai AI Laboratory)
OptimizationRepresentation LearningTransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper proposes Capability Salience Vector (CSV), which redefines the model's capability score by assigning different importance weights to different tokens in the validation set, thereby achieving predictable scaling laws for downstream task performance.
Capture the Key in Reasoning to Enhance CoT Distillation Generalization
Chengwei Dai (Chinese Academy of Sciences), Songlin Hu (Chinese Academy of Sciences)
Knowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkChain-of-Thought
🎯 What it does: Leverage opposing Chain-of-Thought (CoT) pairs generated by a teacher LLM to extract key reasoning steps, further training small language models to enhance reasoning capabilities.
Capturing Author Self Beliefs in Social Media Language
Siddharth Mangalik (Stony Brook University), Ryan L. Boyd (University of Texas at Dallas)
ClassificationRecognitionTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes and implements a self-belief recognition task, constructs a multi-source dataset containing expert annotations, LLM annotations, and short self-belief paragraphs, and trains a high-performance classification model called SelfAwareNet.
CART: A Generative Cross-Modal Retrieval Framework With Coarse-To-Fine Semantic Modeling
Minghui Fang (Zhejiang University), Zhou Zhao (Zhejiang University)
RetrievalTransformerAuto EncoderImageVideoTextMultimodalityRetrieval-Augmented GenerationAudio
🎯 What it does: Propose the CART framework, transforming cross-modal retrieval into a task of generating candidate identifiers
Causal Estimation of Tokenisation Bias
Pietro Lesci (University of Cambridge), Tiago Pimentel (ETH Zürich)
Explainability and InterpretabilityTransformerText
🎯 What it does: This paper defines tokenisation bias as the impact of whether a subword is included in the vocabulary on the model's prediction probability for the corresponding string, and estimates this bias through a causal inference framework and regression discontinuity design without retraining the model;
Causal Graph based Event Reasoning using Semantic Relation Experts
Mahnaz Koupaee (Stony Brook University), Niranjan Balasubramanian (University of Texas at Austin)
Explainability and InterpretabilityTransformerLarge Language ModelAgentic AIMixture of ExpertsTextSequential
🎯 What it does: This paper proposes a collaborative causal graph generation framework that utilizes four specialized experts (temporal, discourse, preconditions, common sense) within a large language model to engage in multi-round discussions, ultimately generating a global event causal graph. This graph is then applied to various event reasoning tasks, including explainable event likelihood prediction, event prediction, and next-event prediction.
Caution for the Environment: Multimodal LLM Agents are Susceptible to Environmental Distractions
Xinbei Ma (Shanghai Jiao Tong University), Hai Zhao (Shanghai Jiao Tong University)
Safty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision-Language-Action ModelImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: This paper investigates the trustworthiness of multimodal large language models (MLLMs) in graphical user interface (GUI) environments, exploring how distracting information in the environment leads agents to deviate from user goals and perform erroneous actions; it also constructs a simulated dataset with four typical scenarios to systematically evaluate various general and specialized GUI agents, verifying their robustness and safety against environmental disturbances.
CC-Tuning: A Cross-Lingual Connection Mechanism for Improving Joint Multilingual Supervised Fine-Tuning
Yangfan Ye (Harbin Institute of Technology), Bing Qin (Huawei Technologies Co., Ltd)
Representation LearningSupervised Fine-TuningText
🎯 What it does: This paper proposes CC-TUNING, a multilingual fine-tuning paradigm that explicitly establishes cross-lingual connection mechanisms at the latent level.
CCHall: A Novel Benchmark for Joint Cross-Lingual and Cross-Modal Hallucinations Detection in Large Language Models
Yongheng Zhang (Central South University), Libo Qin (Central South University)
Large Language ModelImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Propose the CCHall benchmark to evaluate large language models in joint cross-lingual and cross-modal scenarios for hallucination issues and systematically evaluate various MLLMs.
CEAES: Bidirectional Reinforcement Learning Optimization for Consistent and Explainable Essay Assessment
Xia Li (School of Information Science and Technology), Wenjing Pan (School of Information Science and Technology)
OptimizationExplainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Developed a model named CEAES that can simultaneously predict essay scores and generate explanatory feedback, employing a bidirectional reinforcement learning framework to mutually enhance scoring and feedback;
CENTAUR: Bridging the Impossible Trinity of Privacy, Efficiency, and Performance in Privacy-Preserving Transformer Inference
Jinglong Luo (Harbin Institute of Technology), Zenglin Xu (Pengcheng Laboratory)
Safty and PrivacyComputational EfficiencyTransformerText
🎯 What it does: Developed the CENTAUR framework, achieving privacy-preserving Transformer inference by combining random permutation with SMPC;
Centurio: On Drivers of Multilingual Ability of Large Vision-Language Model
Gregor Geigle (University of Würzburg), Goran Glavaš (University of Würzburg)
RecognitionGenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Systematically study the training strategies of multilingual vision-language models (LVLM), explore the impact of the number of languages, training distribution, and text-image understanding, ultimately training the Centurio model that supports 100 languages.
CER: Confidence Enhanced Reasoning in LLMs
Ali Razghandi (Sharif University of Technology), Mahdieh Soleymani Baghshah (Sharif University of Technology)
Explainability and InterpretabilityLarge Language ModelTextChain-of-Thought
🎯 What it does: Proposed a confidence-based multi-path reasoning framework called CER, which improves the reasoning accuracy of LLMs by estimating and weighted voting on the confidence of critical intermediate answers in chain-of-thought reasoning.
CFBench: A Comprehensive Constraints-Following Benchmark for LLMs
Tao Zhang (Baichuan Inc), Zenan Zhou (Baichuan Inc)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposed CFBench, a Chinese constraint-following benchmark containing 1,000 samples, covering over 200 real-world scenarios and 50+ NLP tasks, systematically collecting and classifying constraints, and constructing a multi-dimensional evaluation framework.
Chain-of-Reasoning: Towards Unified Mathematical Reasoning in Large Language Models via a Multi-Paradigm Perspective
Yiyao Yu (Tsinghua University), Furu Wei (Microsoft)
AI Code AssistantLarge Language ModelTextChain-of-Thought
🎯 What it does: Proposed the Chain-of-Reasoning (CoR) framework, integrating natural language reasoning (NLR), algorithmic reasoning (AR), and symbolic reasoning (SR) to build a multi-paradigm reasoning model, CoR-Math-7B;
ChainEdit: Propagating Ripple Effects in LLM Knowledge Editing through Logical Rule-Guided Chains
Zilu Dong (Nanjing University of Science and Technology), Rui Xia (Nanjing University of Science and Technology)
TransformerLarge Language ModelPrompt EngineeringTextGraphBenchmarkChain-of-Thought
🎯 What it does: Proposed the ChainEdit framework, using logic rule-guided chain updates to address the ripple effect problem in LLM knowledge editing.
ChartCoder: Advancing Multimodal Large Language Model for Chart-to-Code Generation
Xuanle Zhao (Tsinghua University), Maosong Sun (Tsinghua University)
Data SynthesisAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodalityChain-of-Thought
🎯 What it does: Developed a dedicated multimodal large language model (ChartCoder) specifically for chart-to-code generation, and created a large-scale dataset named Chart2Code-160k containing 160k chart-code pairs by automatically generating and executing code; simultaneously proposed the Snippet-of-Thought (SoT) method, which converts direct generation into step-by-step generation to highlight key information.
ChartLens: Fine-grained Visual Attribution in Charts
Manan Suri (University of Maryland), Dinesh Manocha (University of Maryland)
Explainability and InterpretabilityConvolutional Neural NetworkTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageMultimodalityBenchmarkChain-of-Thought
🎯 What it does: This paper proposes the ChartLens framework to achieve post-hoc fine-grained visual attribution for chart answers;
ChatBench: From Static Benchmarks to Human-AI Evaluation
Serina Chang (Microsoft Research), Jake M. Hofman (Microsoft Research)
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Design and conduct a user study to convert MMLU multiple-choice questions into real human-AI dialogues, collect user answers when answering alone and in conversation with LLMs, and release a large-scale ChatBench dataset (containing 144K answers and 7,336 dialogues).
ChatSOP: An SOP-Guided MCTS Planning Framework for Controllable LLM Dialogue Agents
Zhigen Li (Tianjin University), Deyi Xiong (Tianjin University)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: Propose ChatSOP, an MCTS planning framework guided by SOP, to enhance the controllability of task-oriented dialogues driven by LLMs.
Cheems: A Practical Guidance for Building and Evaluating Chinese Reward Models from Scratch
Xueru Wen (Chinese Academy of Sciences), Debing Zhang (Xiaohongshu Inc)
Reinforcement Learning from Human FeedbackLarge Language ModelTextBenchmark
🎯 What it does: This paper constructs a Chinese reward model (Reward Model) evaluation benchmark named CheemsBench and a large-scale Chinese preference dataset named CheemsPreference, and trains a high-performance Chinese reward model named CheemsRM through human-AI collaboration.
CHEER-Ekman: Fine-grained Embodied Emotion Classification
Phan Anh Duong (University of Cincinnati), Tianyu Jiang (University of Cincinnati)
RecognitionTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: This paper expands the original CHEER dataset to CHEER-Ekman, adding six basic emotion labels (happiness, sadness, anger, disgust, fear, and surprise), and employs a large language model (LLM) to complete emotion recognition through best-worst scaling (BWS) annotation and prompt engineering.
ChemActor: Enhancing Automated Extraction of Chemical Synthesis Actions with LLM-Generated Data
Yu Zhang (Shanghai Jiao Tong University), Yanyan Xu (X-Imaging Intelligent Technology Co. LTD)
Drug DiscoveryTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Developed ChemActor, a 7B LLM fine-tuned chemical actuator that converts unstructured experimental descriptions into executable chemical steps.
CheXalign: Preference fine-tuning in chest X-ray interpretation models without human feedback
Dennis Hein (Stanford University), Akshay S Chaudhari (Stanford University)
GenerationTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelImageTextBiomedical Data
🎯 What it does: Built a fully automated 'preference tuning' workflow using publicly available chest X-ray images and radiologists' reference reports to generate preference pairs for model-generated reports, and performed preference tuning on chest X-ray image interpretation models (CheXagent, CheXagent-2) based on these pairs.
ChildMandarin: A Comprehensive Mandarin Speech Dataset for Young Children Aged 3-5
Jiaming Zhou (Nankai University), Yong Qin (Nankai University)
Data-Centric LearningTransformerSupervised Fine-TuningBenchmarkAudio
🎯 What it does: Constructed and released a 41.25-hour Mandarin speech dataset of children aged 3-5.
Chinese Inertial GAN for Handwriting Signal Generation and Recognition
Yifeng Wang (Harbin Institute of Technology), Yi Zhao (Harbin Institute of Technology)
RecognitionData SynthesisAuto EncoderGenerative Adversarial NetworkTime Series
🎯 What it does: This paper proposes the CI-GAN model to generate inertial sensor signals for Chinese handwriting, thereby expanding training samples and significantly improving handwriting recognition accuracy.
Chinese SafetyQA: A Safety Short-form Factuality Benchmark for Large Language Models
Yingshui Tan (Future Lab Alibaba Group), Kaifu Zhang (Future Lab Alibaba Group)
Safty and PrivacyTransformerLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed and constructed the Chinese SafetyQA benchmark for evaluating the factual accuracy of large language models in safety knowledge domains such as law, policy, and ethics.
Chinese SimpleQA: A Chinese Factuality Evaluation for Large Language Models
Yancheng He (Alibaba Group), Bo Zheng (Alibaba Group)
Large Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Created the Chinese SimpleQA evaluation benchmark for short question-answering, containing 3,000 short Q&A pairs covering six major topics and 99 subtopics, providing high-quality, static, and easily assessable characteristics.
ChronoSense: Exploring Temporal Understanding in Large Language Models with Time Intervals of Events
Duygu Sezen Islakoglu (Utrecht University), Jan-Christoph Kalo (University of Amsterdam)
TransformerLarge Language ModelPrompt EngineeringTextSequentialBenchmarkChain-of-Thought
🎯 What it does: Created the ChronoSense benchmark to evaluate large language models' understanding of temporal interval relations (13 Allen relations) and temporal arithmetic tasks.
Circuit Compositions: Exploring Modular Structures in Transformer-Based Language Models
Philipp Mondorf (LMU Munich), Barbara Plank (LMU Munich)
Explainability and InterpretabilityComputational EfficiencyTransformerTextSequential
🎯 What it does: Investigate the modular structure of reusable subnetworks (circuits) in Transformer language models across ten string editing subtasks, and explore the feasibility of combining circuits.
Circuit Stability Characterizes Language Model Generalization
Alan Sun (Carnegie Mellon University)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: Define the stability of the reasoning process (circuit) in language models, and demonstrate the correlation between circuit stability and model generalization performance through empirical case studies on tasks such as arithmetic summation, Boolean expression evaluation, and sports understanding.
CiteEval: Principle-Driven Citation Evaluation for Source Attribution
Yumo Xu (AWS AI Labs), Zhiguo Wang (AWS AI Labs)
RetrievalLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes the CiteEval framework, constructs the CiteBench benchmark, and develops the CITEEVAL-AUTO automated evaluation method for fine-grained assessment of citation quality in retrieval-augmented generation (RAG) systems.
CityNavAgent: Aerial Vision-and-Language Navigation with Hierarchical Semantic Planning and Global Memory
Weichen Zhang (Tsinghua University), Yong Li (Tsinghua University)
TransformerLarge Language ModelVision-Language-Action ModelWorld ModelImageTextPoint Cloud
🎯 What it does: Developed a drone agent, CityNavAgent, for urban aerial vision-language navigation (Aerial VLN), leveraging large language models (LLMs) and historical trajectory memory to achieve open-vocabulary semantic perception from images, hierarchical semantic planning, and global memory map search, significantly enhancing long-distance urban navigation performance.
CKnowEdit: A New Chinese Knowledge Editing Dataset for Linguistics, Facts, and Logic Error Correction in LLMs
Jizhan Fang (Zhejiang University), Ningyu Zhang (Zhejiang University)
Data-Centric LearningLarge Language ModelTextBenchmark
🎯 What it does: Constructed the first Chinese knowledge editing dataset CKnowEdit, covering three major categories: linguistics, facts, and logic, with a focus on Chinese language characteristics such as classical poetry, idioms, and allusions.
CLAIM: Mitigating Multilingual Object Hallucination in Large Vision-Language Models with Cross-Lingual Attention Intervention
Zekai Ye (Harbin Institute of Technology), Bing Qin (Harbin Institute of Technology)
Explainability and InterpretabilityTransformerLarge Language ModelVision Language ModelMultimodality
🎯 What it does: Proposes the CLAIM method, which mitigates object hallucinations in multilingual vision-language models during inference through cross-lingual attention intervention.
CLaSp: In-Context Layer Skip for Self-Speculative Decoding
Longze Chen (Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences), Min Yang (Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences)
Computational EfficiencyTextBenchmark
🎯 What it does: Proposes the CLaSp method, achieving adaptive layer skipping self-speculative decoding, which can accelerate LLM inference without training additional modules.
Class Distillation with Mahalanobis Contrast: An Efficient Training Paradigm for Pragmatic Language Understanding Tasks
Chenlu Wang (Stony Brook University), Ritwik Banerjee (Stony Brook University)
ClassificationComputational EfficiencyKnowledge DistillationTransformerSupervised Fine-TuningContrastive LearningText
🎯 What it does: Proposed the ClaD (Class Distillation) training paradigm, which uses Mahalanobis distance to cluster and separate minority target classes, addressing the binary classification challenge caused by the extreme diversity of non-target classes.
Classifying Unreliable Narrators with Large Language Models
Anneliese Brei (University of North Carolina at Chapel Hill), Snigdha Chaturvedi (University of North Carolina at Chapel Hill)
ClassificationTransformerLarge Language ModelText
🎯 What it does: This paper proposes a method for automatically identifying three types of unreliable narrators (internal narration, cross-narration, cross-text) and constructs the corresponding expert-annotated dataset TUNA.
CLEME2.0: Towards Interpretable Evaluation by Disentangling Edits for Grammatical Error Correction
Jingheng Ye (Tsinghua University), Zifei Shan (Tencent)
Explainability and InterpretabilityLarge Language ModelTextBenchmark
🎯 What it does: Propose CLEME2.0, an interpretable grammar error correction (GEC) evaluation metric, which categorizes system edits into four types (TP, FPne, FPun, and FN) and assigns four corresponding scores: hit-correction, wrong-correction, under-correction, and over-correction;
CLIPErase: Efficient Unlearning of Visual-Textual Associations in CLIP
Tianyu Yang (University of Notre Dame), Xiangliang Zhang (University of Notre Dame)
Safty and PrivacyComputational EfficiencyRepresentation LearningVision Language ModelDiffusion modelContrastive LearningImageTextMultimodality
🎯 What it does: Proposed a multimodal machine forgetting framework called CLIPErase, which can efficiently remove specified visual-text associations from pre-trained CLIP models without retraining;
ClusterAttn: KV Cache Compression under Intrinsic Attention Clustering
Minwei Zhang (Beijing University of Posts and Telecommunications), Jianxin Liao (Beijing University of Posts and Telecommunications)
CompressionComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Compress the KV cache during the inference of large language models, proposing the ClusterAttn method which achieves training-agnostic sparse attention compression by leveraging intrinsic attention clustering.
CMHKF: Cross-Modality Heterogeneous Knowledge Fusion for Weakly Supervised Video Anomaly Detection
Guohua Wang (South China Agricultural University), Yongsen Zheng (South China Agricultural University)
Anomaly DetectionRecurrent Neural NetworkGraph Neural NetworkTransformerVision Language ModelAuto EncoderContrastive LearningVideoTextMultimodalityAudio
🎯 What it does: Proposed the CMHKF framework, which achieves weakly supervised video anomaly detection by leveraging cross-modal heterogeneous knowledge fusion.
CoachMe: Decoding Sport Elements with a Reference-Based Coaching Instruction Generation Model
Wei-Hsin Yeh (Institute of Information Science Academia Sinica), Lun-Wei Ku (Institute of Information Science Academia Sinica)
GenerationPose EstimationGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoText
🎯 What it does: Propose a reference video-based motion teaching generation model called CoachMe, which can quantify the differences between learners' movements and standard movements, and automatically generate targeted, actionable training guidance.
CoAM: Corpus of All-Type Multiword Expressions
Yusuke Ide (Nara Institute of Science and Technology), Taro Watanabe (Nara Institute of Science and Technology)
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Constructed and released the CoAM dataset, containing 1,300 sentences covering all types of multi-word expressions (MWE), with type labels annotated for MWEs in the test set.
CoCoLex: Confidence-guided Copy-based Decoding for Grounded Legal Text Generation
Santosh T.y.s.s, Xiaomo Liu (JPMorgan AI Research)
GenerationRetrievalTextRetrieval-Augmented Generation
🎯 What it does: Proposes a confidence-guided copy-based decoding strategy called CoCoLex to enhance the authenticity and coherence of retrieval-enhanced legal text generation.
Code-Switching Red-Teaming: LLM Evaluation for Safety and Multilingual Understanding
Haneul Yoo (KAIST), Hwaran Lee (Sogang University)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelText
🎯 What it does: This paper investigates the use of code-switching (cross-lingual token mixing) as a red-teaming attack method and proposes the CSRT (Code-Switching Red-Teaming) framework, which utilizes automatically generated multilingual mixed queries to simultaneously evaluate the security (attack success rate ASR, rejection rate RR) and multilingual understanding capability (comprehension Cmp) of large language models (LLMs).
CodeDPO: Aligning Code Models with Self Generated and Verified Source Code
Kechi Zhang (Peking University), Zhi Jin (Peking University)
AI Code AssistantLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: This paper proposes the CodeDPO framework, which employs self-generation and self-validation mechanisms for preference learning in code generation models to enhance code correctness and execution efficiency.
CodeTool: Enhancing Programmatic Tool Invocation of LLMs via Process Supervision
Yifei Lu (Northeastern University), Feiliang Ren (Northeastern University)
AI Code AssistantLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: Developed CodeTool, leveraging step-by-step code generation and process supervision to enhance LLM's tool calling capabilities