ACL 2025 Papers — Page 4
Annual Meeting of the Association for Computational Linguistics · 1699 papers
CoE: A Clue of Emotion Framework for Emotion Recognition in Conversations
Zhiyu Shen (Sun Yat-sen University), Jianxing Yu (Sun Yat-sen University)
RecognitionTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed the CoE framework, which enhances dialogue emotion recognition performance by progressively integrating character personality, scene information, and auxiliary tasks.
CogniBench: A Legal-inspired Framework and Dataset for Assessing Cognitive Faithfulness of Large Language Models
Xiaqiang Tang (Hong Kong University of Science and Technology (Guangzhou)), Sihong Xie (Hong Kong University of Science and Technology (Guangzhou))
Anomaly DetectionTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Proposed the CogniBench framework and dataset, which employs a law-informed hierarchical evaluation standard to perform fine-grained assessment of the credibility of factual and cognitive claims generated by LLMs in multi-turn knowledge-driven dialogues. Subsequently, the CogniBench‑L dataset was expanded using an automatic annotation pipeline based on LLMs; followed by training the CogniDet model for efficient detection of factual and cognitive hallucinations.
CoIR: A Comprehensive Benchmark for Code Information Retrieval Models
Xiangyang Li (HUAWEI NOAH'S ARK LAB), Ruiming Tang (HUAWEI NOAH'S ARK LAB)
RetrievalTransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose COIR (Code Information Retrieval Benchmark) for end-to-end evaluation of code retrieval, covering four major retrieval tasks and ten multilingual datasets;
Collapse of Dense Retrievers: Short, Early, and Literal Biases Outranking Factual Evidence
Mohsen Fayyaz (University of California Los Angeles), Nanyun Peng (University of California Los Angeles)
RetrievalTransformerTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Design a series of controlled experiments using Re-DocRED documents to quantify various biases in dense retrievers (short texts, early positions, repeated entities, literal matches, etc.), and reveal the mechanisms by which these biases lead to retrieval errors and RAG failure.
Colloquial Singaporean English Style Transfer with Fine-Grained Explainable Control
Jinggui Liang (Singapore Management University), Lizi Liao (Singapore Management University)
GenerationExplainability and InterpretabilityTransformerLarge Language ModelAgentic AIText
🎯 What it does: Constructed the ExpCSEST fine-grained explanation dataset containing approximately 140K sentences, and proposed the MACoE-Style interpretable and controllable style transfer framework based on multi-agent large language models, achieving fine-grained style migration between Singapore English and Standard English.
Com^2 : A Causal-Guided Benchmark for Exploring Complex Commonsense Reasoning in Large Language Models
Kai Xiong (Harbin Institute of Technology), Ting Liu (Harbin Institute of Technology)
Data SynthesisLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Constructed a complex common-sense reasoning benchmark called Com 2, and used it to evaluate and analyze the reasoning capabilities of various large language models.
Combining Domain and Alignment Vectors Provides Better Knowledge-Safety Trade-offs in LLMs
Megh Thakkar (Chandar Research Lab), Sarath Chandar (Chandar Research Lab)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextBiomedical DataFinance Related
🎯 What it does: Propose the MERGEALIGN method, leveraging interpolation between domain vectors and alignment vectors to achieve safe alignment of domain-specific LLMs;
CoMet: Metaphor-Driven Covert Communication for Multi-Agent Language Games
Shuhang Xu (Beijing Normal University), Fangwei Zhong (Beijing Normal University)
TransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: Propose the CoMet framework to enhance LLMs' ability to use metaphors for covert communication and semantic evasion in multi-agent language games.
Commonsense Reasoning in Arab Culture
Abdelrahman Sadallah (Mohamed bin Zayed University of Artificial Intelligence), Fajri Koto (Mohamed bin Zayed University of Artificial Intelligence)
Large Language ModelTextBenchmark
🎯 What it does: This study constructs the ArabCulture dataset for evaluating common sense reasoning in the Arab world.
Comparing LLM-generated and human-authored news text using formal syntactic theory
Olga Zamaraeva (Universidade da Coruña), Carlos Gómez-Rodríguez (Universidade da Coruña)
Large Language ModelText
🎯 What it does: This paper utilizes the formal syntactic theory HPSG to compare the syntactic structure differences between LLM-generated texts in the New York Times style and texts written by human authors.
Comparing Moral Values in Western English-speaking societies and LLMs with Word Associations
Chaoyi Xiang (University of Melbourne), Lea Frermann (University of Melbourne)
Explainability and InterpretabilityGraph Neural NetworkLarge Language ModelText
🎯 What it does: By comparing the results of Western English native speakers and large language models (Llama-3.1-8B) in word association experiments, a global moral network is constructed to evaluate differences in moral value cognition between the two.
Comparison-based Active Preference Learning for Multi-dimensional Personalization
Minhyeon Oh (POSTECH), Jungseul Ok (POSTECH)
Recommendation SystemReinforcement Learning from Human FeedbackLarge Language ModelText
🎯 What it does: Propose an active multi-dimensional preference learning framework named AMPLe, which leverages pairwise comparison feedback from users on language model outputs. By modifying posterior updates and employing active query selection based on Bayesian inference, it accurately captures latent multi-dimensional user preferences and personalizes large language models (LLMs).
CompileAgent: Automated Real-World Repo-Level Compilation with Tool-Integrated LLM-based Agent System
Li Hu (University of Science and Technology of China), Nenghai Yu (University of Science and Technology of China)
AI Code AssistantTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Explored the CompileAgent framework, based on large language models, for automating repository-level compilation
Completing A Systematic Review in Hours instead of Months with Interactive AI Agents
Rui Qiu (Ohio State University), Han Wei Shen
TransformerLarge Language ModelAgentic AIBiomedical DataReview/Survey Paper
🎯 What it does: Developed Insight Agent, an interactive AI agent system that helps medical experts complete high-quality systematic reviews within a few hours.
Computation Mechanism Behind LLM Position Generalization
Chi Han (University of Illinois Urbana Champaign), Heng Ji (University of Illinois Urbana Champaign)
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: This paper reveals the "position generalization" capability of large language models (LLMs) in handling text position perturbations and ultra-long sequences through in-depth analysis of their self-attention mechanism, proving that attention logits can be approximately decomposed into position-related and semantic-related components, and verifying this phenomenon theoretically and experimentally;
Con Instruction: Universal Jailbreaking of Multimodal Large Language Models via Non-Textual Modalities
Jiahui Geng (Mohamed bin Zayed University of Artificial Intelligence), Iryna Gurevych (Mohamed bin Zayed University of Artificial Intelligence)
Adversarial AttackTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodalityAudio
🎯 What it does: Propose a generic gray-box attack method called Con Instruction, which generates adversarial examples with non-text modalities (image or audio) that are highly similar to malicious text instructions in the embedding space, directly deceiving multi-modal large language models (MLLMs) into jailbreaking.
ConceptCarve: Dynamic Realization of Evidence
Eylon Caplan (Purdue University), Dan Goldwasser (Purdue University)
RetrievalTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Propose the CONCEPTCARVE framework, which dynamically enables trend evidence retrieval within communities by leveraging LLMs and traditional retrievers to construct a concept tree.
Conditional Dichotomy Quantification via Geometric Embedding
Shaobo Cui (EPFL), Boi Faltings (EPFL)
Representation LearningTransformerContrastive LearningText
🎯 What it does: Proposed and implemented a conditional binary quantization task (ConDQ), which directly measures the opposition between two outputs under the same context through geometric embedding; developed the Dichotomy-oriented Geometric Embedding (DoGE) framework to capture such oppositional relationships.
Condor: Enhance LLM Alignment with Knowledge-Driven Data Synthesis and Refinement
Maosong Cao (Shanghai AI Laboratory), Kai Chen (Shanghai AI Laboratory)
Data SynthesisReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Propose a two-stage data generation pipeline (Condor Void + Condor Refine), generating high-quality SFT data through knowledge trees and self-reflection to enhance the dialogue and knowledge answering capabilities of LLMs.
ConECT Dataset: Overcoming Data Scarcity in Context-Aware E-Commerce MT
Mikołaj Pokrywka (Allegro.com), Mikołaj Koszowski (Allegro.com)
Data-Centric LearningTransformerSupervised Fine-TuningPrompt EngineeringVision Language ModelTextMultimodalityBenchmark
🎯 What it does: This paper constructs the ConECT dataset, containing 11,400 Czech-Polish e-commerce product sentence pairs, combined with product images and category paths; subsequently, several context-aware machine translation methods (visual-language models, category path injection, image description prefixes) were fine-tuned and evaluated.
CONFETTI: Conversational Function-Calling Evaluation Through Turn-Level Interactions
Tamer Alkhouli (Amazon), Yi Zhang (Amazon)
Large Language ModelTextSequentialBenchmark
🎯 What it does: Propose the CONFETTI benchmark, which evaluates the multi-round function calling and answer quality of LLMs using 109 artificial dialogues.
Confidence v.s. Critique: A Decomposition of Self-Correction Capability for LLMs
Zhe Yang (Peking University), Zhifang Sui (Peking University)
Explainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Investigated the self-correction ability of large language models (LLMs) and decomposed it into two capabilities: confidence (Confidence) and critique (Critique).
Conformity in Large Language Models
Xiaochen Zhu (University of Cambridge), Andreas Vlachos (University of Cambridge)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Investigated the conformity effects in large language models during dialogues and explored their mechanisms and mitigation strategies.
ConLoan: A Contrastive Multilingual Dataset for Evaluating Loanwords
Sina Ahmadi (University of Zurich), Rico Sennrich (University of Zurich)
Data SynthesisTransformerLarge Language ModelContrastive LearningTextBenchmark
🎯 What it does: Created and annotated a contrastive sentence set called ConLoan in ten languages, where each sentence pair differs only by the loanword and its native alternative, and used this dataset to evaluate LLM surprise and NMT translation performance.
ConSim: Measuring Concept-Based Explanations’ Effectiveness with Automated Simulatability
Antonin Poché (IRT Saint Exupéry), Fanny Jourdan (IRT Saint Exupéry)
ClassificationExplainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: This paper proposes a framework that evaluates conceptual explanation methods using automatic simulatability, employing large language models (LLMs) as simulators to scale and replicate the assessment of different explanation techniques.
ConsistencyChecker: Tree-based Evaluation of LLM Generalization Capabilities
Zhaochen Hong (University of Illinois Urbana Champaign), Jiaxuan You (University of Illinois Urbana Champaign)
AI Code AssistantLarge Language ModelTextChain-of-Thought
🎯 What it does: Propose the ConsistencyChecker framework, which utilizes a tree-based self-consistency structure to evaluate the semantic or functional consistency of LLMs in multi-step transformations (translation, code editing) without requiring reference data.
Consistent Client Simulation for Motivational Interviewing-based Counseling
Yizhe Yang (Beijing Institute of Technology), Ee-Peng Lim (Singapore Management University)
TransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: A framework based on state tracking, action selection, information retrieval, and response generation was constructed to simulate human clients with consistency in Motivational Interviewing (MI) psychological counseling.
Conspiracy Theories and Where to Find Them on TikTok
Francesco Corso (Politecnico di Milano), Gianmarco De Francisci Morales (Politecnico di Milano)
TransformerLarge Language ModelPrompt EngineeringVideoText
🎯 What it does: This paper systematically analyzes conspiracy theory videos on the American TikTok platform from 2021-2023, estimating their frequency, exploring the impact of creators' incentive mechanisms on content duration, and evaluating the zero-shot detection performance of open-weight large language models (LLMs) on text transcription.
Context-Aware Sentiment Forecasting via LLM-based Multi-Perspective Role-Playing Agents
Fanhang Man (Tsinghua University), Yong Li (Tsinghua University)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningAgentic AIText
🎯 What it does: Propose a multi-perspective role-playing framework based on large language models (LLMs) to predict future emotions of social media users in ongoing events, with personalized modeling through implicit features (tone, attitude).
Contextual Experience Replay for Self-Improvement of Language Agents
Yitao Liu (Princeton University), Shunyu Yao (Stanford University)
TransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Proposes Contextual Experience Replay (CER), a training-free, experience replay-based framework that dynamically accumulates, refines, and retrieves past trajectories within the language model context, enabling self-improvement of language agents;
Continual Gradient Low-Rank Projection Fine-Tuning for LLMs
Chenxu Wang (Beijing Jiaotong University), Liping Jing (Beijing Jiaotong University)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Proposed the GORP strategy for large language models, jointly updating full parameters and low-rank parameters within a unified low-rank gradient subspace to achieve continuous fine-tuning;
Contrastive Learning on LLM Back Generation Treebank for Cross-domain Constituency Parsing
Peiming Guo (Harbin Institute of Technology), Yue Zhang (Westlake University)
Domain AdaptationTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Leverage large language models (LLM) to reverse generate cross-domain trees, constructing a new treebank, and propose span-level contrastive learning pre-training to enhance cross-domain constituent syntactic parsing performance.
Contrastive Perplexity for Controlled Generation: An Application in Detoxifying Large Language Models
Tassilo Klein (SAP SE), Moin Nabi (SAP SE)
Safty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringContrastive LearningText
🎯 What it does: Proposes a structure-agnostic fine-tuning framework based on Contrastive Perplexity (CP) for toxicity elimination and attribute control in large language models without altering the model architecture.
Contrastive Prompting Enhances Sentence Embeddings in LLMs through Inference-Time Steering
Zifeng Cheng (Nanjing University), Qing Gu (Nanjing University)
Representation LearningTransformerPrompt EngineeringContrastive LearningText
🎯 What it does: This paper proposes a training-free and data-free inference-time intervention method called Contrastive Prompting (CP), which introduces auxiliary prompts and contrasts them with the original prompts to drive LLMs to better encode the core semantics of sentences in the activation vectors of multi-head attention layers, thereby obtaining higher quality sentence embeddings.
Controllable and Reliable Knowledge-Intensive Task-Oriented Conversational Agents with Declarative Genie Worksheets
Harshit Joshi (Stanford University), Monica Lam (Stanford University)
TransformerLarge Language ModelAgentic AITextRetrieval-Augmented Generation
🎯 What it does: Propose the Genie framework, which defines tasks and knowledge queries using a declarative Genie Worksheet, separating the LLM's semantic parsing from runtime strategy execution, achieving reliable knowledge-intensive task-oriented dialogues.
Controllable Style Arithmetic with Language Models
Weiqi Wang (University of Science and Technology of China), Houqiang Li (University of Science and Technology of China)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed a style arithmetic method based on parameter space, performing addition and subtraction operations on language model parameters using style vectors to achieve fine-grained control over text style, cross-task transfer, and multi-dimensional style combination.
Controlled Low-Rank Adaptation with Subspace Regularization for Continued Training on Large Language Models
Yuheng Lu (Beijing University of Posts and Telecommunications), Xiaojie Wang (LI Auto Inc)
OptimizationComputational EfficiencyTransformerSupervised Fine-TuningText
🎯 What it does: Propose Controlled LoRA (CLoRA), a parameter-efficient fine-tuning method that incorporates subspace regularization into the LoRA structure, which can alleviate catastrophic forgetting during LLM continued training.
ControlSpeech: Towards Simultaneous and Independent Zero-shot Speaker Cloning and Zero-shot Language Style Control
Shengpeng Ji (Zhejiang University), Zhou Zhao (Zhejiang University)
GenerationTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsAuto EncoderTextAudio
🎯 What it does: Developed a TTS system called ControlSpeech that can simultaneously achieve zero-shot voice cloning and text-prompted style control, supporting independent adjustment of timbre, content, and style.
Cool-Fusion: Fuse Large Language Models without Training
Cong Liu (Sun Yat-sen University), Liang Lin (Sun Yat-sen University)
GenerationComputational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposed a training-free multi-large language model (LLM) fusion method called Cool-Fusion, which can share knowledge and generate text between models with different tokenizers.
Cooperative or Competitive? Understanding the Interaction between Attention Heads From A Game Theory Perspective
Xiaoye Qu (Huazhong University of Science and Technology), Yu Cheng (Chinese University of Hong Kong)
Explainability and InterpretabilityTransformerLarge Language ModelTextMultimodality
🎯 What it does: This paper first quantitatively analyzes the interactions between attention heads using Harsanyi dividend from cooperative game theory, revealing that most combinations yield nearly zero gains, while only a few combinations produce positive (cooperative) or negative (competitive) gains. Subsequently, a training-free Game-theoretic Attention Calibration (GAC) method is proposed, which first identifies attention head groups that significantly contribute to model performance, then performs fine-grained smoothing on the attention distributions of the remaining heads to suppress competition and enhance overall collaboration.
Coordinating Chaos: A Structured Review of Linguistic Coordination Methodologies
Benjamin Roger Litterer (University of Michigan), Dallas Card (University of Michigan)
TextReview/Survey Paper
🎯 What it does: Review and systematize research methods for language coordination, propose a framework, and classify and criticize existing methods.
CORAL: Learning Consistent Representations across Multi-step Training with Lighter Speculative Drafter
Yepeng Weng (AI Lab, Lenovo Research), Zhongchao Shi (AI Lab, Lenovo Research)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelContrastive LearningTextBenchmark
🎯 What it does: This paper proposes an improved Speculative Decoding framework called CORAL, which can significantly enhance the generation speed of large language models while maintaining lossless inference.
CORDIAL: Can Multimodal Large Language Models Effectively Understand Coherence Relationships?
Aashish Anantha Ramakrishnan, Dongwon Lee
ClassificationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: Construct and benchmark CORDIAL, evaluating the prediction and verification of multimodal large language models on discourse coherence relations
CoRe-MMRAG: Cross-Source Knowledge Reconciliation for Multimodal RAG
Yang Tian, Liqiang Nie (Harbin Institute of Technology)
RetrievalLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelMultimodalityRetrieval-Augmented Generation
🎯 What it does: Propose a cross-source knowledge harmonization framework called CoRe-MMRAG, which follows a four-stage process: first generate answers using internal parameter knowledge, then select the most relevant evidence through joint visual and textual similarity, subsequently generate external answers based on retrieved multimodal evidence, and finally fuse internal and external responses to produce the final answer.
CoreEval: Automatically Building Contamination-Resilient Datasets with Real-World Knowledge toward Reliable LLM Evaluation
Jingqian Zhao (Harbin Institute of Technology), Ruifeng Xu (Hong Kong Polytechnic University)
Data SynthesisData-Centric LearningLarge Language ModelTextBenchmark
🎯 What it does: Automatically construct a pollution-free evaluation dataset named CoreEval, which integrates original data with real-time world knowledge for updates.
CoRet: Improved Retriever for Code Editing
Fabio James Fehr, Giovanni Zappella (AWS)
RetrievalAI Code AssistantTransformerTextBenchmark
🎯 What it does: Proposed CoRet—a dense retrieval model for code editing tasks that can retrieve relevant code snippets from repositories based on natural language queries.
COSMMIC: Comment-Sensitive Multimodal Multilingual Indian Corpus for Summarization and Headline Generation
Raghvendra Kumar (IIT Patna), Jose G Moreno
GenerationData-Centric LearningTransformerLarge Language ModelVision Language ModelContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: Constructed a multimodal, multilingual dataset named COSMMIC, encompassing nine major Indian languages (Bengali, Hindi, Gujarati, Marathi, Malayalam, Odia, Tamil, Telugu, Kannada), which includes articles, images, and reader comments, and conducted experiments on the tasks of abstract and title generation using this dataset.
CoT-based Synthesizer: Enhancing LLM Performance through Answer Synthesis
Bohan Zhang (Renmin University of China), Jie Tang (Tsinghua University)
Data SynthesisTransformerSupervised Fine-TuningReinforcement LearningTextBenchmarkChain-of-Thought
🎯 What it does: Propose a CoT-based Synthesizer that analyzes and synthesizes multiple candidate answers through chain-of-thought reasoning to generate more accurate responses, and design an automated data generation pipeline to train small models and enhance large model performance.
CoT-ICL Lab: A Synthetic Framework for Studying Chain-of-Thought Learning from In-Context Demonstrations
Vignesh Kothapalli (New York University), Maziar Sanjabi (LinkedIn AI)
Data SynthesisExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: This paper introduces CoT-ICL Lab, a controllable synthetic data framework for systematically studying the performance of chain-of-thought (CoT) in in-context learning (ICL).
CoT-Valve: Length-Compressible Chain-of-Thought Tuning
Xinyin Ma (National University of Singapore), Xinchao Wang (National University of Singapore)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkChain-of-Thought
🎯 What it does: Propose the CoT-Valve mechanism, enabling a single model to dynamically generate chain-of-thought reasoning of varying lengths (from short to long) based on reasoning difficulty, and achieve fine-grained length control and compression through the construction of the MixChain dataset.
Counterfactual-Consistency Prompting for Relative Temporal Understanding in Large Language Models
Jongho Kim (Seoul National University), Seung-won Hwang (Seoul National University)
Large Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Proposes adversarial consistency prompts (CCP), which enhance the temporal consistency and reasoning accuracy of large language models by generating time counterfactual questions and aggregating answers.
Counterspeech the ultimate shield! Multi-Conditioned Counterspeech Generation through Attributed Prefix Learning
Aswini Kumar Padhi (IIT Delhi), Tanmoy Chakraborty (Logically.ai)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringText
🎯 What it does: To address online hate speech, this paper proposes a two-phase multi-attribute adversarial speech generation framework called HiPPrO, which can simultaneously control the strategy of responses (e.g., positive, informative, questioning, condemning) and emotions (anger, disgust, joy, sadness, surprise).
Crab: A Novel Configurable Role-Playing LLM with Assessing Benchmark
Kai He (National University of Singapore), Mengling Feng (National University of Singapore)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBenchmark
🎯 What it does: This paper proposes the Crab framework, which supports configurable role-playing LLMs and is accompanied by a dedicated evaluation benchmark.
Cracking Factual Knowledge: A Comprehensive Analysis of Degenerate Knowledge Neurons in Large Language Models
Yuheng Chen (Key Laboratory of Cognition and Decision Intelligence for Complex Systems), Jun Zhao (Key Laboratory of Cognition and Decision Intelligence for Complex Systems)
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: Study the neural structure of factual knowledge in large language models, propose and systematically define and identify degenerate knowledge neurons (DKN), and explore their role in knowledge updating and robustness.
Cracking the Code of Hallucination in LVLMs with Vision-aware Head Divergence
Jinghan He (Chinese Academy of Sciences), Jinqiao Wang (Chinese Academy of Sciences)
Explainability and InterpretabilityRepresentation LearningTransformerMultimodality
🎯 What it does: This paper analyzes the multi-head attention mechanism and proposes the Vision-aware Head Divergence (VHD) metric to quantify the sensitivity of each attention head to visual information. Based on this, a training-free Vision-aware Head Reinforcement (VHR) method is developed to enhance visually sensitive attention heads during generation, thereby reducing hallucinations in LVLMs.
CrafText Benchmark: Advancing Instruction Following in Complex Multimodal Open-Ended World
Zoya Volovikova (AIRI), Alexey Skrynnik (AIRI)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningVision-Language-Action ModelWorld ModelTextMultimodalityBenchmark
🎯 What it does: Proposed and implemented the CrafText benchmark to evaluate agents' ability to understand and execute natural language instructions in dynamic, multimodal, open-world environments.
Cramming 1568 Tokens into a Single Vector and Back Again: Exploring the Limits of Embedding Space Capacity
Yuri Kuratov (AIRI), Mikhail Burtsev (London Institute for Mathematical Sciences)
CompressionRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: Achieve a lossless compression scheme that can decode thousands of tokens with a single vector by inserting a trainable [mem] vector before a frozen LLM.
CrisisTS: Coupling Social Media Textual Data and Meteorological Time Series for Urgency Classification
Romain Meunier (IRIT, Université de Toulouse, CNRS, Toulouse INP), Savitha Ramasamy (Institute for Infocomm Research)
ClassificationTransformerLarge Language ModelTextMultimodalityTime Series
🎯 What it does: Proposes the CRISISTS multimodal multilingual crisis urgency classification dataset, improving the discriminability of crisis information urgency by aligning social media text with meteorological time series.
CRiskEval: A Chinese Multi-Level Risk Evaluation Benchmark Dataset for Large Language Models
Ling Shi (Tianjin University), Deyi Xiong (Tianjin University)
Safty and PrivacyLarge Language ModelTextBenchmark
🎯 What it does: Constructed a Chinese-oriented frontier risk assessment benchmark dataset CRiskEval, using a fine-grained risk grading system and multiple-choice questions to evaluate the risk propensity of LLMs.
CritiQ: Mining Data Quality Criteria from Human Preferences
Honglin Guo (Fudan University), Tao Gui (Fudan University)
Data-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIText
🎯 What it does: Propose the CRITIQ method, which automatically mines and evolves data quality evaluation criteria using only about 30 manually annotated comparisons, then trains a lightweight scoring model with the generated criteria to select high-quality samples from large-scale text data.
Croppable Knowledge Graph Embedding
Yushan Zhu (Zhejiang University), Huajun Chen (Zhejiang University)
Knowledge DistillationRepresentation LearningTransformerGraph
🎯 What it does: Propose the MED framework, which trains trimmable KGE models in one go, enabling different dimensional sub-models to be used directly.
Cross-Lingual Auto Evaluation for Assessing Multilingual LLMs
Sumanth Doddapaneni (Nilekani Centre at AI4Bharat), Mitesh M Khapra
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Proposed a cross-lingual automatic evaluation framework (CIA Suite) and trained a series of evaluator models (HERCULE) based on Llama-3.1-8B to assess the generated text of multilingual LLMs.
Cross-Lingual Generalization and Compression: From Language-Specific to Shared Neurons
Frederick Riemenschneider (Heidelberg University), Anette Frank (Heidelberg University)
CompressionExplainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: Analyze the pre-training process of multilingual language models, tracking the evolution from language-specific to shared neurons.
Cross-Lingual Optimization for Language Transfer in Large Language Models
Jungseob Lee (Korea University), Heuiseok Lim (Korea University)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmark
🎯 What it does: This paper proposes a cross-lingual optimization (CLO) method that efficiently transfers large language models to target languages using publicly available English SFT data and translation models, while maintaining English capabilities.
Cross-Lingual Pitfalls: Automatic Probing Cross-Lingual Weakness of Multilingual Large Language Models
Zixiang Xu (MBZUAI), Xiangliang Zhang (University of Notre Dame)
Explainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: This paper proposes an automated method combining Beam Search and LLM-Simulation to efficiently generate bilingual question pairs, aiming to accurately reveal performance limitations of multilingual large language models (LLMs) on non-English languages. Based on this, a dataset containing over 6,000 bilingual samples (covering 16 languages) was constructed.
Cross-Lingual Representation Alignment Through Contrastive Image-Caption Tuning
Nathaniel Krasner (George Mason University), Antonios Anastasopoulos (George Mason University)
RetrievalRepresentation LearningTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This paper achieves cross-lingual alignment of multilingual text representations through contrastive learning on image-caption data.
Cross-Lingual Transfer of Cultural Knowledge: An Asymmetric Phenomenon
Chen Zhang (Peking University), Yansong Feng (Peking University)
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: Investigate the mechanism of cultural knowledge transfer in large language models across multilingual environments, constructing an interpretable experimental framework and conducting continuous pre-training and bilingual evaluations on four non-English cultures (Korean, Chinese, Tibetan, Mongolian);
Cross-model Transferability among Large Language Models on the Platonic Representations of Concepts
Youcheng Huang (Sichuan University), Jiancheng Lv (Sichuan University)
Representation LearningTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Investigated the cross-model transferability of concept representations (Steering Vectors, SVs) in large language models (LLMs), proposing the L-Cross Modulation method that transfers SVs from a source model to a target model via linear transformation, and conducting experiments to validate the approach.
Crowd Comparative Reasoning: Unlocking Comprehensive Evaluations for LLM-as-a-Judge
Qiyuan Zhang (City University of Hong Kong), Chen Ma (City University of Hong Kong)
Explainability and InterpretabilityKnowledge DistillationLarge Language ModelSupervised Fine-TuningTextBenchmarkChain-of-Thought
🎯 What it does: Propose a population-based comparison evaluation method (CCE), which enriches the chain-of-thought (CoT) assessment of LLM-as-a-Judge by generating diverse group responses and comparing them with candidate answers.
Crowdsource, Crawl, or Generate? Creating SEA-VL, a Multicultural Vision-Language Dataset for Southeast Asia
Samuel Cahyawijaya (Cohere), Peerat Limkonchotiwat (AI Singapore)
Data SynthesisTransformerVision Language ModelDiffusion modelContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: Construct the SEA-VL multilingual visual-language dataset using three methods: community crowdsourcing, web crawling, and AI generation, accumulating over 1.28 million culturally relevant images with corresponding multilingual annotations;
CRUXEVAL-X: A Benchmark for Multilingual Code Reasoning, Understanding and Execution
Ruiyang Xu (Chinese Academy of Sciences), Le Sun (Chinese Academy of Sciences)
AI Code AssistantLarge Language ModelTextBenchmark
🎯 What it does: Built and released CRUXEVAL-X, a multilingual code reasoning benchmark covering 19 programming languages, focusing on input/output reasoning tasks.
CSTree-SRI: Introspection-Driven Cognitive Semantic Tree for Multi-Turn Question Answering over Extra-Long Contexts
Zhaowen Wang (Central South University), Li Kuang (Central South University)
RetrievalTransformerLarge Language ModelMixture of ExpertsTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes the CSTree-SRI framework to address logical consistency and retrieval efficiency in multi-turn question answering over ultra-long contexts.
CU-MAM: Coherence-Driven Unified Macro-Structures for Argument Mining
Debela Gemechu (University of Dundee), Chris Reed (University of Dundee)
ClassificationRecognitionGraph Neural NetworkTransformerText
🎯 What it does: This paper proposes a unified macro-structure driven argument mining model, CU-MAM, which achieves joint prediction of argument component types, relation identification, and their types by integrating local and global consistency of argument relations into a multi-task learning framework.
Cuckoo: An IE Free Rider Hatched by Massive Nutrition in LLM’s Nest
Letian Peng (University of California San Diego), Jingbo Shang (University of California San Diego)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Transform the next-word prediction (NTP) of LLMs into next-word extraction (NTE), automatically generating BIO labeling using pre-training and fine-tuning data from C4 and TuluV3 to train a large-scale, few-shot, and instruction-driven IE model named Cuckoo.
CULEMO: Cultural Lenses on Emotion - Benchmarking LLMs for Cross-Cultural Emotion Understanding
Tadesse Destaw Belay (Instituto Politécnico Nacional), Seid Muhie Yimam (University of Hamburg)
ClassificationRecognitionTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: This paper develops the CuLEmo benchmark to evaluate the performance of large language models in cross-cultural emotion recognition and sentiment analysis.
CulFiT: A Fine-grained Cultural-aware LLM Training Paradigm via Multilingual Critique Data Synthesis
Ruixiang Feng (University of Electronic Science and Technology of China), Shuo Shang (University of Electronic Science and Technology of China)
Explainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmark
🎯 What it does: Propose a multi-lingual, multi-dimensional cultural awareness training paradigm named CulFiT, leveraging target-aware multi-lingual criticism data and fine-grained reward mechanisms to enhance LLM's cultural sensitivity and diversity;
Cultivating Gaming Sense for Yourself: Making VLMs Gaming Experts
Wenxuan Lu (Institute of Information Engineering, Chinese Academy of Sciences), Tianning Zang (Institute of Information Engineering, Chinese Academy of Sciences)
TransformerReinforcement LearningVision Language ModelImage
🎯 What it does: Propose the GameSense framework, enabling Vision Language Models (VLMs) to act as developers of gameplay modules, generating real-time execution modules (GSMs) to achieve uninterrupted, real-time game control.
Cultural Bias Matters: A Cross-Cultural Benchmark Dataset and Sentiment-Enriched Model for Understanding Multimodal Metaphors
Senqi Yang (Dalian University of Technology), Feng Xia (RMIT University)
ClassificationTransformerVision Language ModelMultimodalityBenchmark
🎯 What it does: Constructed the cross-cultural multimodal metaphor dataset MultiMM and proposed the emotion-enhanced metaphor detection model SEMD;
Cultural Learning-Based Culture Adaptation of Language Models
Chen Cecilia Liu (Technical University of Darmstadt), Iryna Gurevych (Technical University of Darmstadt)
Data SynthesisDomain AdaptationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Propose the CLCA framework, which generates social dialogues through role-playing that simulates cultural adaptation and combines intent understanding for multi-task fine-tuning to improve LLM alignment across different cultural values.
CulturalBench: A Robust, Diverse and Challenging Benchmark for Measuring LMs’ Cultural Knowledge Through Human-AI Red-Teaming
Yu Ying Chiu (University Of Washington), Yejin Choi (Stanford University)
Data-Centric LearningLarge Language ModelTextBenchmark
🎯 What it does: Constructed CULTURALBENCH, a cultural knowledge benchmark containing 1,696 questions verified by five people, spanning 45 regions, covering 17 themes, and collected through human-AI red team collaboration;
Culture is Not Trivia: Sociocultural Theory for Cultural NLP
Naitian Zhou (University of California Berkeley), Isaac L. Bleaman
Review/Survey Paper
🎯 What it does: This paper introduces sociocultural linguistics theory into cultural NLP, advocating for viewing culture as a dynamic, contextual process of social meaning-making, and proposes a practical path aimed at localization.
Culture Matters in Toxic Language Detection in Persian
Zahra Bokaei (University of Edinburgh), Bonnie Webber (University of Edinburgh)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Investigated the toxic language detection task in Persian (Farsi), systematically comparing methods such as fine-tuning, zero/few-shot inference, data augmentation (sparse supervision), and cross-lingual transfer learning, and for the first time evaluated the impact of cultural context on transfer effectiveness.
Curiosity-Driven Reinforcement Learning from Human Feedback
Haoran Sun (Baidu Inc), Haifeng Wang (Baidu Inc)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Propose the CD-RLHF framework, combining curiosity-driven intrinsic rewards with traditional human feedback-based RLHF, encouraging the model to explore less-visited states during generation, thereby enhancing output diversity while maintaining alignment quality.
Curriculum Debiasing: Toward Robust Parameter-Efficient Fine-Tuning Against Dataset Biases
Mingyu Lee (Korea University), SangKeun Lee (Korea University)
Representation LearningData-Centric LearningTransformerSupervised Fine-TuningText
🎯 What it does: Proposes a curriculum learning based bias elimination framework named CURRICULUM DEBIASING to enhance the generalization ability of parameter-efficient fine-tuning (PEFT) on biased data.
CxGGEC: Construction-Guided Grammatical Error Correction
Yayu Cao (Zhejiang University), Ming Cai (Dalian University of Technology)
TransformerLarge Language ModelText
🎯 What it does: CxGGEC, a GEC framework based on construction grammar, first automatically builds a construction dictionary from corpora. It then uses a noise-robust construction prediction model to identify expected constructions in ungrammatical sentences. Finally, it trains a Seq2Seq model on parallel data with construction masks to achieve error correction.
CypherBench: Towards Precise Retrieval over Full-scale Modern Knowledge Graphs in the LLM Era
Yanlin Feng (Megagon Labs), Sajjadur Rahman (Adobe)
RetrievalTransformerLarge Language ModelPrompt EngineeringTextGraphBenchmark
🎯 What it does: Convert large RDF knowledge graphs (e.g., Wikidata) into multi-domain attribute graph views, and construct the CypherBench benchmark based on these views, which includes 11 large-scale attribute graphs and over 10,000 natural language questions, followed by proposing a text-to-Cypher (text-to-Cypher) task generation and evaluation process.
D.Va: Validate Your Demonstration First Before You Use It
Qi Zhang (Zhejiang University), Junbo Zhao (Zhejiang University)
RetrievalData-Centric LearningLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Propose an adaptive example selection method based on demonstration verification (D.Va), which indirectly estimates and calibrates perplexity under the target task using retrieved validation examples to guide context selection in large language models.
DAC: A Dynamic Attention-aware Approach for Task-Agnostic Prompt Compression
Yi Zhao (Shanghai Jiao Tong University), Liu Guoming (Xiaomi)
CompressionComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose a task-agnostic prompt compression method (DAC) based on the fusion of dynamic attention and information entropy, maximizing information retention while reducing computational costs through multi-stage dynamic compression and attention-based key token constraints.
DAPE V2: Process Attention Score as Feature Map for Length Extrapolation
Chuanyang Zheng (Chinese University of Hong Kong), Yu Li (Chinese University of Hong Kong)
Computational EfficiencyRepresentation LearningConvolutional Neural NetworkTransformerTextBenchmark
🎯 What it does: Proposed the CDAPE (Convolutional Data-Adaptive Position Encoding) method, treating attention scores as feature maps and applying convolution on them to enhance the length extrapolation capability of Transformers.
DARS: Dynamic Action Re-Sampling to Enhance Coding Agent Performance by Adaptive Tree Traversal
Vaibhav Aggarwal (MPI for Intelligent Systems), Bernhard Schölkopf (MPI for Intelligent Systems)
AI Code AssistantReinforcement Learning from Human FeedbackLarge Language ModelSupervised Fine-TuningAgentic AITextSequential
🎯 What it does: Proposes a Dynamic Action Resampling (DARS) method, which resamples actions of LLM-encoded agents at critical decision points to quickly correct suboptimal decisions and enhance software defect repair effectiveness.
Data Caricatures: On the Representation of African American Language in Pretraining Corpora
Nicholas Deas (Columbia University), Kathleen McKeown (Columbia University)
Representation LearningData-Centric LearningLarge Language ModelText
🎯 What it does: Quantitative and qualitative analysis of African American English (AAL) in 12 publicly pre-trained corpora, evaluating its quantity, quality, and the impact of filters on AAL representations.
Data Laundering: Artificially Boosting Benchmark Results through Knowledge Distillation
Jonibek Mansurov (Mohamed bin Zayed University of Artificial Intelligence), Alham Fikri Aji (Mohamed bin Zayed University of Artificial Intelligence)
Knowledge DistillationTransformerTextBenchmark
🎯 What it does: Utilize knowledge distillation techniques to construct a 'data washing' process, covertly injecting benchmark test set knowledge into the model to artificially enhance benchmark scores.
Data Quality Issues in Multilingual Speech Datasets: The Need for Sociolinguistic Awareness and Proactive Language Planning
Mingfei Lau (Google), Pavel Golik (Google)
Data-Centric LearningAudio
🎯 What it does: Conducted a quality audit of three publicly available multilingual datasets: Mozilla Common Voice 17.0, FLEURS, and VoxPopuli, identifying and summarizing micro-level issues (e.g., extremely short recordings, low speech ratio, topic imbalance, speaker homogenization) and macro-level issues (e.g., multiple writing systems, dialect/register confusion, unclear dialect boundaries). Using the case of Southern Min (nan_tw), the study analyzed data alignment and overlap problems caused by the lack of writing standards. Subsequently, the paper proposed guidelines, including sociolinguistic evaluation, language planning, standardization guidance, multi-tier quality control, and transparent metadata, to guide the construction of future low-resource or non-institutionalized language datasets.
Data Whisperer: Efficient Data Selection for Task-Specific LLM Fine-Tuning via Few-Shot In-Context Learning
Shaobo Wang (Shanghai Jiao Tong University), Linfeng Zhang (Shanghai Jiao Tong University)
Computational EfficiencyData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Construct a training-agnostic, attention-based efficient data selection method called Data Whisperer using few-shot context learning with pre-trained LLMs.
Data-Constrained Synthesis of Training Data for De-Identification
Thomas Vakili (Stockholm University), Hercules Dalianis (Stockholm University)
Data SynthesisDomain AdaptationSafty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningTextBiomedical Data
🎯 What it does: The study investigates the utility of synthetic clinical text generated by domain-adapted LLMs under data-limited conditions, training a NER model with machine-labeled PII entities to evaluate the effectiveness of synthetic data for downstream tasks.
DavIR: Data Selection via Implicit Reward for Large Language Models
Haotian Zhou (ByteDance, Inc), Hongxia Yang (ByteDance, Inc)
OptimizationData-Centric LearningLarge Language ModelReinforcement LearningText
🎯 What it does: Proposes the DavIR method, which measures 'learnability' by comparing the cross-entropy loss differences between pre-trained LLMs and reference models on each data point, and uses length normalization to eliminate sequence length bias, achieving efficient core set data selection.
DCG-SQL: Enhancing In-Context Learning for Text-to-SQL with Deep Contextual Schema Link Graph
Jihyung Lee (Sungkyunkwan University), Jee-Hyong Lee (Sungkyunkwan University)
RetrievalAI Code AssistantGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextTabularBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes a retrieval method based on a deep contextual pattern linking graph, which uses this graph to retrieve relevant demonstration samples in the prompt-based learning of large language models (LLMs), thereby generating more accurate SQL queries.
DDxTutor: Clinical Reasoning Tutoring System with Differential Diagnosis-Based Structured Reasoning
Qian Wu (Chinese University of Hong Kong), Qi Dou (Chinese University of Hong Kong)
TransformerLarge Language ModelSupervised Fine-TuningTextBiomedical Data
🎯 What it does: Propose the DDxTutor framework, utilizing structured reasoning based on differential diagnosis to achieve clinical diagnosis teaching.
DeAL: Decoding-time Alignment for Large Language Models
James Y. Huang (University of Southern California), Dan Roth (Amazon WS AI Labs)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: Propose a framework that aligns the output of large language models during inference through heuristic search (DeAL), allowing users to use custom reward functions to achieve multi-objective alignment during the decoding process.
DebateCoder: Towards Collective Intelligence of LLMs via Test Case Driven LLM Debate for Code Generation
Jizheng Chen (Shanghai Jiao Tong University), Yong Yu (Shanghai Jiao Tong University)
GenerationAI Code AssistantTransformerLarge Language ModelText
🎯 What it does: Proposes the DebateCoder framework, which improves code generation through test case-driven debates between two large language models (LLMs).