EMNLP 2025 Papers — Page 3
Conference on Empirical Methods in Natural Language Processing · 1809 papers
BYOKG-RAG: Multi-Strategy Graph Retrieval for Knowledge Graph Question Answering
Costas Mavromatis, George Karypis (Amazon)
RetrievalLarge Language ModelGraphRetrieval-Augmented Generation
🎯 What it does: Propose the BYOKG-RAG framework, which leverages LLM to generate graph retrieval artifacts such as entities, paths, and queries, combined with multiple specialized graph retrieval tools for iterative context retrieval, ultimately producing knowledge graph question-answering results.
C3: A Bilingual Benchmark for Spoken Dialogue Models Exploring Challenges in Complex Conversations
Chengqian Ma (Peking University), Steven Y. Guo
Data-Centric LearningTransformerLarge Language ModelTextBenchmarkAudio
🎯 What it does: Proposed the C3 bilingual benchmark, covering five complex dialogue phenomena including speech ambiguity, semantic ambiguity, omissions, coreference, and multi-turn interactions, and provided LLM evaluation methods; conducted systematic evaluation of six end-to-end speech dialogue models.
Cache-Efficient Posterior Sampling for Reinforcement Learning with LLM-Derived Priors Across Discrete and Continuous Domains
Ibne Farabi Shihab (Iowa State University), Anuj Sharma (Iowa State University)
Computational EfficiencyMeta LearningTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextTabularRetrieval-Augmented Generation
🎯 What it does: Proposes a meta-learning regulated caching framework that uses a large language model (LLM) as an action suggester for reinforcement learning, significantly reducing LLM query frequency while maintaining near-optimal performance.
Cache-of-Thought: Master-Apprentice Framework for Cost-Effective Vision Language Model Reasoning
Mingyuan Wu (University of Illinois Urbana Champaign), Klara Nahrstedt (University of Illinois Urbana Champaign)
Computational EfficiencyKnowledge DistillationVision Language ModelMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes the Cache-of-Thought (CoT) framework, which stores answers generated by a large VLM (master) in a cache, and subsequently enhances the reasoning quality of a small VLM (apprentice) by using multi-modal retrieval and context learning to retrieve similar QA pairs from the cache as prompts.
Cacheback: Speculative Decoding With Nothing But Cache
Zhiyao Ma (Yale University), Lin Zhong (Yale University)
Computational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper proposes Cacheback Decoding, a training- and model-agnostic inference acceleration method that leverages n-gram records in LRU cache tables to generate draft sequences, thereby achieving Speculative Decoding.
CAFE: Retrieval Head-based Coarse-to-Fine Information Seeking to Enhance Multi-Document QA Capability
Han Peng (Renmin University of China), Lei Fang (DataCanvas Alaya NeW)
RetrievalTransformerLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed the CAFE two-stage coarse-to-fine information seeking framework, which leverages the LLM's internal retrieval head to filter background documents and uses attention mechanisms to focus on key evidence, thereby enhancing multi-document question answering performance.
CAIR: Counterfactual-based Agent Influence Ranker for Agentic AI Workflows
Amit Giloni (Fujitsu Research of Europe), Roman Vainshtein (Fujitsu Research of Europe)
Explainability and InterpretabilityComputational EfficiencyAdversarial AttackLarge Language ModelAgentic AIText
🎯 What it does: Proposes the CAIR method, which uses adversarial analysis to evaluate the impact of each agent on the final output in multi-agent workflows.
CaKE: Circuit-aware Editing Enables Generalizable Knowledge Learners
Yunzhi Yao (Zhejiang University), Nanyun Peng (University of California, Los Angeles)
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Proposes the Circuit-aware Knowledge Editing (CaKE) method, which analyzes the internal reasoning circuits of LLMs and combines circuit-aware training data to ensure that edited knowledge is correctly utilized in multi-hop reasoning.
Calibrating LLM Confidence by Probing Perturbed Representation Stability
Reza Khanmohammadi (Michigan State University), Mohammad M. Ghassemi (Michigan State University)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningAdversarial AttackTransformerLarge Language ModelContrastive LearningText
🎯 What it does: This paper proposes a CCPS method based on target adversarial perturbation of the final hidden states of LLMs and extracting stability features, achieving confidence estimation of LLM answer correctness with a lightweight classifier;
Calibrating LLMs for Text-to-SQL Parsing by Leveraging Sub-clause Frequencies
Terrance Liu (Carnegie Mellon University), Chirag Gupta (Bloomberg)
AI Code AssistantTransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper investigates the calibration issue of large language models in text-to-SQL parsing tasks, and proposes a multivariate Platt scaling method based on clause frequency to achieve more reliable confidence predictions.
Calibrating Pseudo-Labeling with Class Distribution for Semi-supervised Text Classification
Weiyi Yang, Jiawei Sheng (Chinese Academy Of Sciences)
ClassificationTransformerTextBenchmark
🎯 What it does: This paper proposes a pseudo-label calibration framework called PL-POT based on optimal transport, aimed at addressing the pseudo-label bias caused by class imbalance in semi-supervised text classification.
Calibrating Verbal Uncertainty as a Linear Feature to Reduce Hallucinations
Ziwei Ji (Meta FAIR), Nicola Cancedda (Meta FAIR)
Anomaly DetectionExplainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: This paper discovers and utilizes a single linear direction within LLMs to regulate the model's verbal uncertainty, thereby achieving detection and mitigation of confident hallucinations.
Calibration Across Layers: Understanding Calibration Evolution in LLMs
Abhinav Joshi (Indian Institute of Technology Kanpur), Ashutosh Modi (Indian Institute of Technology Kanpur)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: Investigated the calibration evolution process of large language models across different layers, revealing a self-correction phase in later layers and uncovering low-dimensional calibration directions;
CalligraphicOCR for Chinese Calligraphy Recognition
Xiaoyi Bao (Hong Kong Polytechnic University), Chu-Ren Huang (Hong Kong Polytechnic University)
RecognitionLarge Language ModelSupervised Fine-TuningVision Language ModelImageText
🎯 What it does: Propose an end-to-end Chinese calligraphy recognition framework (COCR) that directly converts complete calligraphy images into readable, punctuation-complete text sentences.
Can an Individual Manipulate the Collective Decisions of Multi-Agents?
Fengyuan Liu (Tencent Robotics X), Jindong Gu (University of Oxford)
OptimizationSafty and PrivacyAdversarial AttackTransformerLarge Language ModelAgentic AIPrompt EngineeringText
🎯 What it does: Studying how to exploit attacks on a single agent in a multi-agent system to mislead the collective decision-making of the entire system when only information about that single agent is known.
Can GRPO Boost Complex Multimodal Table Understanding?
Xiaoqiang Kang (Xi'an Jiaotong-Liverpool University), Qiufeng Wang (Xi'an Jiaotong-Liverpool University)
Supervised Fine-TuningReinforcement LearningPrompt EngineeringVision Language ModelMultimodalityTabular
🎯 What it does: Proposed a three-stage reinforcement learning framework Table-R1 to enhance multimodal table understanding capabilities, first preheating through SFT, then aligning table structures using continuous TEDS reward-based Perception-Alignment GRPO, and finally refining remaining reasoning steps with fine-grained rewards via Hint-Completion GRPO.
Can Large Language Models Act as Ensembler for Multi-GNNs?
Hanqi Duan (East China Normal University), Xiang Li (East China Normal University)
ClassificationGraph Neural NetworkLarge Language ModelSupervised Fine-TuningPrompt EngineeringGraphBenchmark
🎯 What it does: Propose LensGNN, which utilizes a large language model (LLM) as an integrator for multiple GNNs, achieving deep fusion of node attribute text and graph structural information through two-stage alignment between multiple GNNs and LLM, to complete node and graph classification tasks.
Can Large Language Models be Effective Online Opinion Miners?
Ryang Heo (Yonsei University), Dongha Lee (Yonsei University)
Data-Centric LearningTransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper proposes the OOMB benchmark to evaluate the opinion mining capability of large language models in diverse, real online texts.
Can Large Language Models Be Good Language Teachers?
LiQing Xu, Ping Wang (Shanghai Jiao Tong University)
TransformerSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Proposed and implemented the CLTE (Chinese Language Teaching Evaluation) benchmark for systematically evaluating the capabilities of large language models (LLMs) in Chinese language teaching.
Can Large Language Models Outperform Non-Experts in Poetry Evaluation? A Comparative Study Using the Consensual Assessment Technique
Piotr Sawicki (University of Kent), Fabricio Goes (University of Leicester)
TransformerLarge Language ModelText
🎯 What it does: This study transfers the Consensual Assessment Technique (CAT) to large language models (Claude-3-Oppus and GPT-4o), automatically evaluating 90 poems through batch assessment and forced ranking, and comparing the results with non-expert human evaluations;
Can Large Language Models Tackle Graph Partitioning?
Yiheng Wu (National University of Defense Technology), Jibing Wu (National University of Defense Technology)
OptimizationTransformerLarge Language ModelPrompt EngineeringGraph
🎯 What it does: This paper explores the feasibility of large language models (LLM) in graph partitioning tasks and proposes a three-stage pipeline (graph coarsening, reasoning, refinement) to enable LLMs to partition large-scale graphs;
Can Large Language Models Translate Spoken-Only Languages through International Phonetic Transcription?
Jiale Chen (South China Normal University), Tianyong Hao (South China Normal University)
TransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes the UNILANG framework, which utilizes large language models (LLMs) to translate unwritten writing systems (e.g., Bai language) through intermediary text in the International Phonetic Alphabet (IPA);
Can Large Language Models Translate Unseen Languages in Underrepresented Scripts?
Dianqing Lin (Inner Mongolia University), Guodong Shi (Inner Mongolia University)
GenerationTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the Lotus benchmark to evaluate the translation ability of large language models on low-resource script languages (Mongolian traditional script and I language) that the models have not been exposed to.
Can Large Language Models Unlock Novel Scientific Research Ideas?
Sandeep Kumar (Indian Institute of Technology Patna), Asif Ekbal (Indian Institute of Technology Patna)
TransformerLarge Language ModelPrompt EngineeringTextPhysics RelatedRetrieval-Augmented Generation
🎯 What it does: This paper explores whether large language models (LLMs) can read scientific papers and automatically generate new research directions, and proposes a quantifiable automatic evaluation method for this purpose;
Can Large Language Models Win the International Mathematical Games?
Alessio Cocchieri (University of Bologna), Gianluca Moro (University of Bologna)
ImageTextBenchmarkChain-of-Thought
🎯 What it does: Constructed and released the MATHGAMES benchmark, collecting 2,183 problems from the International Mathematics and Logic Games Championship, and conducted zero-shot evaluations on 28 LLM/LMM; simultaneously compared results with human participants in the 2024 international competition.
Can LLM Agents Maintain a Persona in Discourse?
Pranav Bhandari (University of Western Australia), Mehwish Nasim (Edith Cowan University)
ClassificationTransformerLarge Language ModelAgentic AIText
🎯 What it does: Studied how LLM agents maintain and express specified OCEAN Big Five personality traits (high/low) in dyadic dialogues and assessed their consistency through evaluation.
Can LLMs be Good Graph Judge for Knowledge Graph Construction?
Haoyu Huang (Hong Kong University of Science and Technology), Wentao Zhang (Peking University)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose the GraphJudge framework, leveraging large language models (LLMs) to act as a 'graph judge' during knowledge graph (KG) construction. It first denoises entity-centric text and extracts candidate triplets, then uses a fine-tuned open-source LLM to determine the authenticity of triplets, filtering noise, errors, and hallucinations to ultimately generate high-quality KGs.
Can LLMs be Literary Companions?: Analysing LLMs on Bengali Figures of Speech Identification
Sourav Das (Indian Institute of Information Technology Kalyani), Kripabandhu Ghosh (Indian Institute of Science Education and Research Kolkata)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Created the BengFoS dataset and conducted experiments on Bengali metaphor (Figures of Speech) identification tasks, evaluating the performance of LLMs (Llama-3, DeepSeek R1) after fine-tuning and their internal representations;
Can LLMs Explain Themselves Counterfactually?
Zahra Dehghanighobadi (Ruhr University Bochum), Muhammad Bilal Zafar (Ruhr University Bochum)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Investigate whether large language models (LLMs) can generate self-generated counterfactual explanations (SCEs) and evaluate their effectiveness and reliability.
Can LLMs Extract Frame-Semantic Arguments?
Jacob Devasier, Chengkai Li (University of Texas at Arlington)
RecognitionTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: A systematic evaluation of large language models (LLMs) for argument identification in frame semantic parsing, investigating input formats, model scale, generalization ability, and proposing a novel frame identification method based on predicted frame elements;
Can LLMs Generate and Solve Linguistic Olympiad Puzzles?
Neh Majmudar (City University Of New York), Elena Filatova (City University Of New York)
GenerationTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Investigate the capability of large language models (LLMs) to solve and generate high school and intermediate-level language Olympiad puzzles (primarily from UKLO and writing system puzzles), and attempt to automatically generate puzzles that comply with Olympiad rules.
Can LLMs Help You at Work? A Sandbox for Evaluating LLM Agents in Enterprise Environments
Harsh Vishwakarma (Fujitsu Research), Mahesh Chandran (Fujitsu Research)
TransformerLarge Language ModelAgentic AITextBenchmarkChain-of-Thought
🎯 What it does: This paper constructs an enterprise-level LLM agent evaluation benchmark called EnterpriseBench, containing 500 tasks spanning HR, IT, SWE, Sales, Finance, and other domains, involving multi-source data, access control, and cross-functional workflows, and provides a simulated enterprise sandbox and an automatic task generation pipeline.
Can LLMs Reason Abstractly Over Math Word Problems Without CoT? Disentangling Abstract Formulation From Arithmetic Computation
Ziling Cheng (Mila - Quebec AI Institute), Jackie CK Cheung (Mila - Quebec AI Institute)
Explainability and InterpretabilityTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: This study proposes a decoupled evaluation framework by decomposing the solving of math word problems into two subskills: abstraction (constructing mathematical expressions) and arithmetic operations. It uses mechanism-interpretable methods to reveal that LLMs internally perform reasoning in the order of 'first abstraction, then computation,' proving that arithmetic calculation is the primary bottleneck for final answer accuracy.
Can LLMs simulate the same correct solutions to free-response math problems as real students?
Yuya Asano (University of Pittsburgh), Erin Walker (University of Pittsburgh)
TransformerLarge Language ModelAgentic AIPrompt EngineeringTextChain-of-Thought
🎯 What it does: Compare the correct solutions of large language models (LLMs) on open-response math problems with those of real students, investigating their differences in idea diversity, coverage, and distribution similarity.
Can Prompts Rewind Time for LLMs? Evaluating the Effectiveness of Prompted Knowledge Cutoffs
Xin Gao (University Of California San Diego), Pengtao Xie (University Of California San Diego)
Explainability and InterpretabilityLarge Language ModelPrompt EngineeringText
🎯 What it does: Investigating whether injecting simulated knowledge cutoff prompts into large language models can enable the model to answer questions based on its knowledge state prior to a given time point, thereby reducing misjudgments caused by training data leakage.
Can Vision-Language Models Solve Visual Math Equations?
Monjoy Narayan Choudhury (IIIT Bangalore), Mrinmaya Sachan (ETH Zürich)
Vision Language ModelImageMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Investigated the performance of vision-language models (VLMs) in visual equation solving tasks, evaluating their integration of perception and symbolic reasoning capabilities.
CARD: Cross-modal Agent Framework for Generative and Editable Residential Design
Pengyu Zeng (Tsinghua University), Shuai Lu (Tsinghua University)
GenerationLarge Language ModelAgentic AIDiffusion modelMultimodalityRetrieval-Augmented Generation
🎯 What it does: Proposed the CARD framework, which utilizes a cross-modal agent system to achieve text-based residential floor plan generation and editing;
Cardiverse: Harnessing LLMs for Novel Card Game Prototyping
Danrui Li (Rutgers University), Mubbasir Kapadia (Roblox)
GenerationData SynthesisTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringText
🎯 What it does: Construct a complete card game prototype generation pipeline by extracting game mechanics through graph structures, generating code with LLM, and building game AI based on heuristic functions.
CARE: A Disagreement Detection Framework with Concept Alignment and Reasoning Enhancement
Jiyuan Liu (Sun Yat-sen University), Yanghui Rao (Sun Yat-sen University)
ClassificationData SynthesisExplainability and InterpretabilityKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Propose the CARE framework, consisting of two stages: Concept Alignment (aligning LLM and expert conceptual understanding through a sub-concept vocabulary) and Reasoning Enhancement (enhancing LLM's causal reasoning ability via a curriculum-based reasoning process, including rationale-to-critique and counterfactual-to-detection).
CARE: Multilingual Human Preference Learning for Cultural Awareness
Geyang Guo (Georgia Institute of Technology), Wei Xu (Georgia Institute of Technology)
Data-Centric LearningReinforcement Learning from Human FeedbackLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: This study constructs a multilingual cultural preference dataset called CARE and enhances the cultural awareness of multilingual language models through preference learning.
CARFT: Boosting LLM Reasoning via Contrastive Learning with Annotated Chain-of-Thought-based Reinforced Fine-Tuning
Wenqiao Zhu (HiThink Research), Yulun Zhang (Shanghai Jiao Tong University)
TransformerReinforcement LearningContrastive LearningTextChain-of-Thought
🎯 What it does: Propose the CARFT method, combining contrastive learning with annotated Chain-of-Thought (CoT) based reinforcement fine-tuning to enhance LLM reasoning performance.
CARMA: Enhanced Compositionality in LLMs via Advanced Regularisation and Mutual Information Alignment
Nura Aljaafari (University of Manchester), Andre Freitas
ClassificationTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
🎯 What it does: By incorporating mutual information alignment and layer-wise stability regularization into the multi-layer hidden representations of LLMs, the model's consistency and robustness in compositional reasoning tasks are enhanced, while maintaining performance on downstream tasks after fine-tuning.
Case-Based Decision-Theoretic Decoding with Quality Memories
Hiroyuki Deguchi (NTT Inc), Masaaki Nagata (NTT Inc)
GenerationComputational EfficiencyLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Proposes a case-based decision theory (CBDT) decoding method, combining it with minimum Bayes risk (MBR) decoding to improve text generation quality.
Castle: Causal Cascade Updates in Relational Databases with Large Language Models
Yongye Su (Purdue University), Elisa Bertino (Rutgers University)
AI Code AssistantTransformerLarge Language ModelPrompt EngineeringTabularChain-of-Thought
🎯 What it does: Proposed and implemented the Castle framework, which leverages large language models to generate SQL UPDATE statements with causal cascading characteristics based solely on database schema information, and automatically generates and verifies triggers to ensure data consistency without exposing table data.
CAT: Causal Attention Tuning For Injecting Fine-grained Causal Knowledge into Large Language Models
Kairong Han (Zhejiang University), Kun Kuang (Zhejiang University)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Investigate how to inject fine-grained causal knowledge into the attention mechanism of large language models, and evaluate the model's causal reasoning ability in discrete distribution (OOD) scenarios using a self-made Spurious Token Game (STG) benchmark.
Causal Interventions Reveal Shared Structure Across English Filler–Gap Constructions
Sasha Boguraev (University of Texas at Austin), Kyle Mahowald (University of Texas at Austin)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: This paper investigates whether English fill-gap constructions (such as wh-questions, relative clauses, clefts, etc.) share the same abstract mechanisms within large language models by performing causal intervenability on them, and reveals through experiments that these mechanisms are influenced by factors such as frequency, activity, and fill type.
Causal Representation Learning from Multimodal Clinical Records under Non-Random Modality Missingness
Zihan Liang (Emory University), Ruoxuan Xiong (Emory University)
Representation LearningLarge Language ModelContrastive LearningMultimodalityElectronic Health Records
🎯 What it does: Proposed a causal inference-based multi-modal clinical data representation learning framework, CRL-MMNAR, which can effectively integrate structured data, imaging, and text for multi-task prediction under non-random missing modality scenarios.
Causal Tree Extraction from Medical Case Reports: A Novel Task for Experts-like Text Comprehension
Sakiko Yahata (Kyoto University), Ryozo Nagai (Kyoto University)
TransformerLarge Language ModelSupervised Fine-TuningTextBiomedical Data
🎯 What it does: Proposed and implemented a new task of extracting causal tree structures (CTE) from medical case reports
CausalVLBench: Benchmarking Visual Causal Reasoning in Large Vision-Language Models
Aneesh Komanduri (University of Arkansas), Xintao Wu (University of Arkansas)
TransformerPrompt EngineeringVision Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposes the CausalVLBench benchmark to evaluate the causal reasoning capabilities of large vision-language models (LVLMs), designing three tasks: causal structure inference, intervention target prediction, and counterfactual prediction, and systematically evaluates open-source LVLMs under zero-shot and few-shot scenarios.
CAVE : Detecting and Explaining Commonsense Anomalies in Visual Environments
Rishika Bhagwatkar (EPFL), Antoine Bosselut (EPFL)
Anomaly DetectionExplainability and InterpretabilityPrompt EngineeringVision Language ModelImageMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposed the CAVE benchmark, collecting and annotating real-world visual anomalies to evaluate the capabilities of Vision-Language models in multi-task scenarios such as anomaly detection, description, explanation, and attribution.
CBP-Tuning: Efficient Local Customization for Black-box Large Language Models
Jiaxuan Zhao (Chinese Academy of Sciences), Weiping Wang (Chinese Academy of Sciences)
Federated LearningSafty and PrivacyComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextBiomedical DataFinance Related
🎯 What it does: Propose the CBP-Tuning framework, which achieves local and efficient customization of black-box large language models by leveraging a prompt generator trained on the server and gradient-free optimization (CMA-ES) on the client side, while ensuring bidirectional privacy protection.
CCQA: Generating Question from Solution Can Improve Inference-Time Reasoning in SLMs
Jinyoung Kim, Ji Won Yoon (Chung-Ang University)
Computational EfficiencyTransformerSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: Propose CCQA, a method that leverages cyclic consistency to generate questions during reasoning and evaluate answers, specifically designed for small language models (SLMs);
CEMTM: Contextual Embedding-based Multimodal Topic Modeling
Amirhossein Abaskohi (University of British Columbia), Giuseppe Carenini (University of British Columbia)
RetrievalExplainability and InterpretabilityRepresentation LearningTransformerVision Language ModelMultimodality
🎯 What it does: Propose an interpretable multimodal topic model CEMTM that leverages pre-trained vision-language models (VL-LM) to generate context embeddings and learns through distributed attention and reconstruction objectives.
Certainty in Uncertainty: Reasoning over Uncertain Knowledge Graphs with Statistical Guarantees
Yuqicheng Zhu (University of Stuttgart), Steffen Staab (University of Stuttgart)
Computational EfficiencyRepresentation LearningGraph Neural NetworkGraph
🎯 What it does: Propose the UNKGCP framework to construct prediction intervals for uncertain knowledge graph embedding models, providing statistical confidence guarantees.
Certified Mitigation of Worst-Case LLM Copyright Infringement
Jingyu Zhang (Johns Hopkins University), Daniel Khashabi (Johns Hopkins University)
Safty and PrivacyTransformerLarge Language ModelText
🎯 What it does: Designed and implemented BLOOMSCRUB, a copyright shielding method that detects and rewrites long references using Bloom filters during inference.
Chain-of-Talkers (CoTalk): Fast Human Annotation of Dense Image Captions
Yijun Shen (Hohai University), Yuhui Zheng (Hohai University)
Data-Centric LearningTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: Propose an AI-assisted multimodal annotation framework called Chain-of-Talkers (COTALK), which achieves efficient generation of dense image captions by first providing a complete description and then allowing subsequent annotators to only supplement residual information.
Chameleon LLMs: User Personas Influence Chatbot Personality Shifts
Jane Xing (University of North Carolina at Chapel Hill), Shashank Srivastava (University of North Carolina at Chapel Hill)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Studied how large language models (LLMs) perceive personality during prolonged interactions with users, constructed a self-assessment measurement framework based on the IPIP 50-item questionnaire, and conducted large-scale simulation experiments to explore adaptation patterns and predictability/regulatability of different Big Five personality traits.
Chart2Code53: A Large-Scale Diverse and Complex Dataset for Enhancing Chart-to-Code Generation
Tianhao Niu (Harbin Institute of Technology), Wanxiang Che (Harbin Institute of Technology)
GenerationData SynthesisAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodalityBenchmark
🎯 What it does: Constructed a large-scale, diverse, and complex Chart2Code dataset called Chart2Code53, and fine-tuned open-source multimodal LLMs on this dataset to achieve state-of-the-art (SOTA) performance on the Chart2Code benchmark.
ChartGaze: Enhancing Chart Understanding in LVLMs with Eye-Tracking Guided Attention Refinement
Ali Salamatian (University of British Columbia), Giuseppe Carenini (University of British Columbia)
Explainability and InterpretabilityTransformerSupervised Fine-TuningVision Language ModelImageMultimodality
🎯 What it does: This study constructs the ChartGaze eye-tracking dataset, systematically analyzes the attention distribution of large vision-language models (LVLM) in chart question answering (CQA) tasks and the deviation from human gaze, and proposes an attention correction method guided by human eye movements;
Charting the Landscape of African NLP: Mapping Progress and Shaping the Road Ahead
Jesujoba Oluwadara Alabi, Dietrich Klakow (Saarland University)
TransformerLarge Language ModelTextReview/Survey PaperAudio
🎯 What it does: Conduct a systematic literature review of natural language processing research on African languages over the past five years, collecting and annotating 884 papers, and analyzing research languages, tasks, techniques, datasets, and research trends.
ChartMind: A Comprehensive Benchmark for Complex Real-world Multimodal Chart Question Answering
Jingxuan Wei (Beijing Wenge Technology Co Ltd), Lei Wang (Beijing Wenge Technology Co Ltd)
Large Language ModelPrompt EngineeringVision Language ModelMultimodalityBenchmark
🎯 What it does: This paper proposes the ChartMind benchmark for multimodal chart question answering in complex real-world scenarios;
Chat-Driven Text Generation and Interaction for Person Retrieval
Zequn Xie (Zhejiang University), Tao Jin (Zhejiang University)
RetrievalLarge Language ModelVision Language ModelContrastive LearningImageText
🎯 What it does: This paper proposes a fully manual-annotation-free text retrieval person search framework named CTGI, which achieves efficient and scalable text retrieval person search by generating high-quality pseudo labels through multi-turn dialogue (MTG) and dynamically refining queries via multi-turn interaction during inference (MTI).
ChatVLA: Unified Multimodal Understanding and Robot Control with Vision-Language-Action Model
Zhongyi Zhou (East China Normal University), Yi Xu (Midea Group)
Robotic IntelligenceSupervised Fine-TuningMixture of ExpertsVision-Language-Action ModelDiffusion modelImageTextMultimodality
🎯 What it does: Propose ChatVLA, a unified Vision-Language-Action model that integrates multimodal understanding, dialogue capabilities, and robot control.
CheckEval: A reliable LLM-as-a-Judge framework for evaluating text generation using checklists
Yukyung Lee (Boston University), Najoung Kim (Boston University)
GenerationExplainability and InterpretabilityLarge Language ModelTextBenchmark
🎯 What it does: Proposes CheckEval, a binary question checklist-based LLM-as-a-Judge evaluation framework.
CHENGYU-BENCH: Benchmarking Large Language Models for Chinese Idiom Understanding and Use
Yicheng Fu (Stanford University), Zhongdongming Dai (University of California San Diego)
Large Language ModelTextBenchmark
🎯 What it does: Propose CHENGYU-BENCH, covering three types of idiom tasks: evaluation, applicability, and open-ended fill-in.
Child-Directed Language Does Not Consistently Boost Syntax Learning in Language Models
Francesca Padovani (University of Groningen), Arianna Bisazza (University of Groningen)
TransformerLarge Language ModelText
🎯 What it does: This paper systematically evaluates the differences in syntactic learning performance between language models trained on child-directed language (CDL) and adult-directed text (ADL);
Chinese Toxic Language Mitigation via Sentiment Polarity Consistent Rewrites
Xintong Wang (Universität Hamburg), Chris Biemann (Universität Hamburg)
GenerationData SynthesisTransformerLarge Language ModelTextBenchmark
🎯 What it does: Developed the TOXIREWRITECN Chinese toxic language detoxification dataset and evaluated the detoxification effectiveness of 17 LLMs under emotion preservation.
CHURRO: Making History Readable with an Open-Weight Large Vision-Language Model for High-Accuracy, Low-Cost Historical Text Recognition
Sina Semnani (Stanford University), Monica Lam (Stanford University)
RecognitionTransformerSupervised Fine-TuningVision Language ModelTextMultimodalityBenchmark
🎯 What it does: Propose the CHURRO model and the CHURRO-DS dataset for historical text recognition; by fine-tuning on the 3B-parameter Qwen 2.5 VL, a cost-effective and high-performance historical OCR VLM is achieved.
CIE: Controlling Language Model Text Generations Using Continuous Signals
Vinay Samuel (University of Maryland), Daphne Ippolito (Carnegie Mellon University)
GenerationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This paper proposes the CIE (Control through Interpolated Embeddings) method, which utilizes interpolated control vectors to achieve continuous regulation of language model output attributes, particularly focusing on precise control over answer length.
CIFLEX: Contextual Instruction Flow for Sub-task Execution in Multi-Turn Interactions with a Single On-Device LLM
Juntae Lee (Qualcomm AI Research), Simyung Chang (Qualcomm AI Research)
Computational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextSequential
🎯 What it does: Propose CIFLEX, a framework that efficiently executes multi-task subtasks in a single-edge LLM by reusing KV caches and injecting side-path instructions.
CIKT: A Collaborative and Iterative Knowledge Tracing Framework with Large Language Models
Runze Li, Wei Zhang (East China Normal University)
OptimizationExplainability and InterpretabilityKnowledge DistillationReinforcement Learning from Human FeedbackTransformerLarge Language ModelTabularSequential
🎯 What it does: Propose a collaborative iterative knowledge tracking framework (CIKT), which generates explainable student knowledge profiles through the Analyst of a large language model, predicts learning performance using the profiles, and achieves mutual improvement via Kahneman-Tversky optimization loops.
Circuit Complexity Bounds for RoPE-based Transformer Architecture
Bo Chen, Jiahao Zhang
Computational EfficiencyTransformer
🎯 What it does: Systematically analyze the RoPE (Rotary Position Embedding) enhanced Transformer architecture from the perspective of circuit complexity, proving that it can be uniformly simulated by TC₀ circuits and that it cannot solve arithmetic formula evaluation or Boolean formula value problems with constant layers and polynomial precision (unless TC₀=NC₁).
CiteBART: Learning to Generate Citations for Local Citation Recommendation
Ege Yiğit Çelik (Izmir Institute of Technology), Selma Tekir (Izmir Institute of Technology)
GenerationRecommendation SystemTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Propose the CiteBART model, which utilizes BART pretraining and custom citation masking to end-to-end generate citations in local contexts, subsequently expanding to a global version by incorporating cited paper titles and abstracts.
CityEQA: A Hierarchical LLM Agent on Embodied Question Answering Benchmark in City Space
Yong Zhao (National University of Defense Technology), Jincai Huang (National University of Defense Technology)
Robotic IntelligenceTransformerLarge Language ModelVision Language ModelVision-Language-Action ModelImageMultimodalityBenchmark
🎯 What it does: This study proposes CityEQA-EC, an embodied question answering benchmark in open city environments, and implements an active exploration and natural language question answering system in urban spaces using a hierarchical LLM-based Planner-Manager-Actor (PMA) agent.
ClimateViz: A Benchmark for Statistical Reasoning and Fact Verification on Scientific Charts
Ruiran Su (University of Oxford), Janet B. Pierrehumbert (University of Oxford)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringMultimodalityTabularBenchmark
🎯 What it does: This study constructs the CLIMATEVIZ dataset, providing a fact-checking benchmark based on real scientific charts and conducting a systematic evaluation of multimodal large language models.
CLIP-MoE: Towards Building Mixture of Experts for CLIP with Diversified Multiplet Upcycling
Jihai Zhang (Chinese University of Hong Kong), Yu Cheng (Chinese University of Hong Kong)
ClassificationRetrievalSupervised Fine-TuningMixture of ExpertsVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Fine-tune the existing CLIP through multi-stage contrastive learning by only updating the FFN layer, generating a set of diversified expert models, and integrating them into a sparse activation Mixture of Experts (MoE) framework, thereby constructing CLIP-MoE and significantly enhancing CLIP's representation capability;
CLLMate: A Multimodal Benchmark for Weather and Climate Events Forecasting
Haobo Li (Hong Kong University of Science and Technology), Huamin Qu (Hong Kong University of Science and Technology)
TransformerLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodalityTime SeriesBenchmarkPhysics Related
🎯 What it does: Propose the Weather and Climate Event Forecasting (WCEF) task and construct the first multi-modal dataset CLLMate, utilizing ERA5 meteorological raster data and environmental news text for event prediction;
CLMTracing: Black-box User-level Watermarking for Code Language Model Tracing
Boyu Zhang (Zhejiang University), Jianwei Yin (Zhejiang University)
Safty and PrivacyAdversarial AttackAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose a black-box user-level watermarking framework called CLMTracing, which embeds traceable watermarks into code language models using regularized watermarks and utility-preserving injection methods.
CMedCalc-Bench: A Fine-Grained Benchmark for Chinese Medical Calculations in LLM
Yunyan Zhang (Tencent Jarvis Lab), Xian Wu (Tencent Jarvis Lab)
Large Language ModelPrompt EngineeringTextBiomedical DataElectronic Health RecordsBenchmarkChain-of-Thought
🎯 What it does: This paper constructs CMedCalc-Bench, a Chinese medical calculation benchmark covering 69 medical calculators and containing 1143 real clinical cases, along with a fine-grained evaluation framework.
CMHG: A Dataset and Benchmark for Headline Generation of Minority Languages in China
Guixian Xu (Key Laboratory of Ethnic Language Intelligent Analysis and Security Governance of MOE), Yushuang Dong (Key Laboratory of Ethnic Language Intelligent Analysis and Security Governance of MOE)
GenerationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: This paper introduces the CMHG dataset, specifically designed for the title generation task in Chinese minority languages (Tibetan, Uyghur, Mongolian), with scales of 100,000, 50,000, and 50,000 entries respectively; simultaneously, it provides a high-quality test set of approximately 3,000 entries reviewed by native speakers.
Co-Eval: Augmenting LLM-based Evaluation with Machine Metrics
Ling-I Wu (Shanghai Jiao Tong University), Guoqiang Li (Shanghai Second Polytechnic University)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Propose the Co-Eval framework, which leverages machine metrics to enhance the scalability and fairness of LLM evaluation.
Co-Evolving LLMs and Embedding Models via Density-Guided Preference Optimization for Text Clustering
Zetong Li (Sun Yat-sen University), Yin Yang (China Mobile Internet Company Ltd)
OptimizationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringContrastive LearningText
🎯 What it does: Propose a bidirectional training framework that enables large language models (LLM) and embedding models to mutually enhance each other in text clustering tasks;
Coarse-to-Fine Grounded Memory for LLM Agent Planning
Wei Yang (Chinese Academy of Sciences), Bo Xu (Chinese Academy of Sciences)
OptimizationComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes the Coarse-to-Fine Grounded Memory (CFGM) framework, which first uses the internal knowledge of large language models (LLMs) to perform coarse-grained extraction of focal points to guide experience collection, then refines experience prompts through mixed-level hierarchies, and finally conducts fine-grained self-questioning and self-answering during inference to correct errors.
COAS2W: A Chinese Older-Adults Spoken-to-Written Transformation Corpus with Context Awareness
Chun Kang (Fudan University), Yangfan Zhou (Fudan University)
GenerationData SynthesisData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkAudio
🎯 What it does: Constructed the COAS2W corpus (10,004 transcribed elderly speech sentences with 4-sentence context, fine-grained error labels, and written results), and trained and evaluated context-aware speech-to-written models on this corpus.
CoBA: Counterbias Text Augmentation for Mitigating Various Spurious Correlations via Semantic Triples
Kyohoon Jin (DATUMO AITRICS), YoungBin Kim
Domain AdaptationExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelText
🎯 What it does: Propose the COBA framework, which first decomposes text into semantic triplets, then modifies and reconstructs them at the triplet level to generate counterbias data, eliminating multiple biases and enhancing model OOD robustness.
CoBia: Constructed Conversations Can Trigger Otherwise Concealed Societal Biases in LLMs
Nafiseh Nikeghbal (Technical University of Munich), Jana Diesner (Technical University of Munich)
Adversarial AttackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Proposed a lightweight adversarial attack called CoBia that reveals hidden social biases in LLMs by constructing dialogues.
COCO-Tree: Compositional Hierarchical Concept Trees for Enhanced Reasoning in Vision-Language Models
Sanchit Sinha (University of Virginia), Aidong Zhang (University of Virginia)
Explainability and InterpretabilityLarge Language ModelVision Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Enhance the performance of vision-language models on compositional reasoning tasks by constructing hierarchical concept trees and integrating them with LLM reasoning, while providing interpretable reasoning paths.
CoCoA: Confidence- and Context-Aware Adaptive Decoding for Resolving Knowledge Conflicts in Large Language Models
Anant Khandelwal (Microsoft), Puneet Agrawal (Microsoft)
TransformerText
🎯 What it does: Proposed an adaptive decoding algorithm called COCOA based on confidence and context to resolve conflicts between model parameter knowledge and retrieved context during LLM inference, enhancing the authenticity of the generated text.
Code Execution as Grounded Supervision for LLM Reasoning
Dongwon Jung (University of California, Davis), Muhao Chen (University of California, Davis)
Explainability and InterpretabilityData-Centric LearningAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: Proposed a method that generates verifiable chain-of-thought (CoT) data by leveraging the execution trajectory of executable code, and translates it into natural language to supervise the training of large language models.
Code to Think, Think to Code: A Survey on Code-Enhanced Reasoning and Reasoning-Driven Code Intelligence in LLMs
Dayu Yang (Meta AI), Julian McAuley (University of California, San Diego)
AI Code AssistantTransformerLarge Language ModelAgentic AIPrompt EngineeringMultimodalityReview/Survey PaperBenchmarkChain-of-Thought
🎯 What it does: Summarized and systematized the mutual reinforcement relationship between code and reasoning in large language models, elucidating how code serves as a structured medium to enhance reasoning capabilities, and conversely, how advances in reasoning drive code intelligence from basic completion toward intelligent code agents.
CodeArena: Evaluating and Aligning CodeLLMs on Human Preference
Jian Yang, Junyang Lin (Alibaba Group)
AI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Propose the CodeArena benchmark to evaluate user preference alignment in LLM code generation.
CodeMixBench: Evaluating Code-Mixing Capabilities of LLMs Across 18 Languages
Yilun Yang, Yekun Chai (ETH Zurich)
Data SynthesisData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: This paper proposes CodeMixBench, a novel evaluation benchmark comprising 8 tasks, 18 languages, 22 synthetic datasets, and 30 public datasets, specifically designed to assess the performance of large language models (LLMs) in code-mixing scenarios.
CodeRAG: Finding Relevant and Necessary Knowledge for Retrieval-Augmented Repository-Level Code Completion
Sheng Zhang (Xiamen University), Hui Li (Xiamen University)
RetrievalKnowledge DistillationAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Propose the CodeRAG framework for warehouse-level code completion, addressing issues such as improper query construction, single-path retrieval, and mismatch between the retriever and code LLM.
CodeSSM: Towards State Space Models for Code Understanding
Shweta Verma (TU Darmstadt), Mira Mezini (TU Darmstadt)
ClassificationRetrievalLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed CodeSSM, an encoder based on the state space model (SSM) for code understanding tasks (retrieval, classification, clone detection, etc.)
CODI: Compressing Chain-of-Thought into Continuous Space via Self-Distillation
Zhenyi Shen (King's College London), Yulan He (King's College London)
Computational EfficiencyKnowledge DistillationTextChain-of-Thought
🎯 What it does: Proposed a self-distillation framework named CODI, which compresses Chain-of-Thought (CoT) reasoning into a continuous latent space, achieving implicit CoT reasoning.
CoEvo: Coevolution of LLM and Retrieval Model for Domain-Specific Information Retrieval
Ang Li, Kun Kuang (Zhejiang University)
RetrievalTransformerLarge Language ModelContrastive LearningTextBiomedical Data
🎯 What it does: Study the collaborative evolution of LLM and retriever in professional domain retrieval, proposing the CoEvo framework to achieve alternating optimization
CogDual: Enhancing Dual Cognition of LLMs via Reinforcement Learning with Implicit Rule-Based Rewards
Cheng Liu (Tencent), Xiaolong Li (Northeastern University)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
🎯 What it does: Propose CogDual, a role-playing language agent that adopts a dual-cognitive framework with cognitive precedence (context-awareness and self-awareness), and further enhances generation quality through reinforcement learning.
Cognitive Linguistic Identity Fusion Score (CLIFS): A Scalable Cognition‐Informed Approach to Quantifying Identity Fusion from Text
Devin R. Wright (Indiana University Bloomington), Yong-Yeol Ahn (Indiana University Bloomington)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed a metaphor detection method based on LLM, constructed an automated and scalable cognitive linguistics identity fusion metric (CLIFS), and significantly improved performance in violence risk prediction.
COLA: Collaborative Multi-Agent Framework with Dynamic Task Scheduling for GUI Automation
Di Zhao (National University of Defense Technology), Zhao Lv (Academy of Military Sciences)
OptimizationExplainability and InterpretabilityRobotic IntelligenceTransformerLarge Language ModelAgentic AIMixture of ExpertsBenchmark
🎯 What it does: Proposes the COLA framework, achieving GUI automation through collaborative multi-agent dynamic task scheduling.
CoLA: Compute-Efficient Pre-Training of LLMs via Low-Rank Activation
Ziyue Liu (University of California at Santa Barbara), Zheng Zhang (University of California at Santa Barbara)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelAuto EncoderText
🎯 What it does: Propose CoLA, which replaces the full-size MLP and projection layers in LLMs with low-rank autoencoders, enforcing activations to remain low-rank, thereby significantly reducing model parameters and FLOPs.