EMNLP 2025 Papers — Page 4
Conference on Empirical Methods in Natural Language Processing · 1809 papers
Collab-Overcooked: Benchmarking and Evaluating Large Language Models as Collaborative Agents
Haochen Sun (Beijing University of Posts and Telecommunications), Xiaojie Wang (Beijing University of Posts and Telecommunications)
Large Language ModelAgentic AIBenchmark
🎯 What it does: Proposed the Collab-Overcooked benchmark to evaluate the collaborative capabilities of large language model multi-agent systems in cooking tasks with resource isolation and asynchronous knowledge sharing, achieving multi-agent collaboration through natural language communication.
Collaborative Beam Search: Enhancing LLM Reasoning via Collective Consensus
Yangyifan Xu (University of Chinese Academy of Sciences), Jiajun Zhang (University of Chinese Academy of Sciences)
TransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: Designed and implemented Collaborative Beam Search (CBS), which enhances LLM multi-step reasoning performance by generating candidate steps through multi-model collaboration, validating them using perplexity consensus, and dynamically allocating generation quotas.
Collaborative Rational Speech Act: Pragmatic Reasoning for Multi-Turn Dialog
Lautaro Estienne (International Laboratory on Learning Systems), Pablo Piantanida (International Laboratory on Learning Systems)
TransformerLarge Language ModelTextSequential
🎯 What it does: Proposed the Collaborative Rational Speech Act (CRSA) framework, which formally models multi-turn collaborative dialogues using information-theoretic methods.
COM-BOM: Bayesian Exemplar Search for Efficiently Exploring the Accuracy-Calibration Pareto Frontier
Gaoxiang Luo (University of Minnesota), Aryan Deshwal (University of Minnesota)
OptimizationText
🎯 What it does: Propose reformulating context example selection as a multi-objective optimization problem, with objectives to simultaneously maximize prediction accuracy and minimize expected calibration error, and based on this, propose the COM-BOM algorithm to achieve efficient compositional Bayesian optimization.
CoMAT: Chain of Mathematically Annotated Thought Improves Mathematical Reasoning
Joshua Ong Jun Leang (University of Edinburgh), Shay B. Cohen (University of Edinburgh)
Explainability and InterpretabilityLarge Language ModelTextChain-of-Thought
🎯 What it does: Propose the CoMAT framework, decomposing mathematical reasoning into two stages: symbolic transformation and reasoning execution, enabling mathematical reasoning with pure LLM without external solvers.
Combining Constrained and Unconstrained Decoding via Boosting: BoostCD and Its Application to Information Extraction
Marija Sakota, Robert West (EPFL)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringText
🎯 What it does: Propose a two-stage BoostCD framework that fuses weak predictions generated by autoregressive language models under constrained and unconstrained decoding modes, ultimately achieving more accurate structured outputs.
ComicScene154: A Scene Dataset for Comic Analysis
Sandro Paval (Technical University of Applied Sciences Würzburg Schweinfurt), Ivan P. Yamshchikov (Technical University of Applied Sciences Würzburg Schweinfurt)
SegmentationTransformerLarge Language ModelMultimodalityBenchmark
🎯 What it does: Created the ComicScene154 dataset, manually segmenting public-domain comics into scenes and labeling whether each panel marks the start of a new scene;
CoMMIT: Coordinated Multimodal Instruction Tuning
Xintong Li (University of California San Diego), Jingbo Shang (University of California San Diego)
OptimizationRepresentation LearningLarge Language ModelSupervised Fine-TuningVision Language ModelImageMultimodalityAudio
🎯 What it does: Propose CoMMIT, an adaptive learning rate scheduling scheme for multi-modal instruction fine-tuning, designed to balance the learning progress of feature encoders and large language models (LLMs).
Comparing human and LLM politeness strategies in free production
Haoran Zhao (University of Washington), Robert D. Hawkins (Stanford University)
GenerationTransformerLarge Language ModelText
🎯 What it does: Compare the generation of politeness strategies between humans and large language models (LLMs) in various scenarios, including constrained multiple-choice and open-ended generation experiments; quantitatively and qualitatively analyze the language strategies (e.g., negation, supplementation, irony, polite tone) produced by humans and LLMs to evaluate the models' performance in contextual sensitivity and politeness dimensions.
Comparing Specialised Small and General Large Language Models on Text Classification: 100 Labelled Samples to Achieve Break-Even Performance
Branislav Pecher (Brno University of Technology), Maria Bielikova (Kempelen Institute of Intelligent Technologies)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This paper studies the performance comparison between specialized small language models and general large language models in text classification tasks under the condition of limited labeled samples;
CompassVerifier: A Unified and Robust Verifier for LLMs Evaluation and Outcome Reward
Shudong Liu (Shanghai AI Laboratory), Kai Chen (Shanghai AI Laboratory)
Adversarial AttackTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Proposed and implemented a unified and robust answer verifier called CompassVerifier, and created a verification benchmark named VerifierBench covering multi-domain and multi-answer types.
CompKBQA: Component-wise Task Decomposition for Knowledge Base Question Answering
Yuhang Tian (Beijing Institute of Technology), Luan Zhang (Beijing Institute of Technology)
TransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the CompKBQA framework, decomposing the logical form generation of knowledge base question answering into four subtasks: skeleton generation, topic entity generation, relation retrieval (R3), and final logical form generation, leveraging LLM's chained fine-tuning to enhance generation quality.
Complex Numerical Reasoning with Numerical Semantic Pre-training Framework
Jun Zhang (Beijing University of Posts and Telecommunications), Haoran Luo (Beijing University of Posts and Telecommunications)
Representation LearningGraph Neural NetworkGraph
🎯 What it does: This paper proposes the CNR-NST framework, which utilizes a pre-trained multi-relational link predictor to perform complex multi-hop reasoning and arithmetic operations on numerical knowledge graphs.
ComplexTempQA: A 100m Dataset for Complex Temporal Question Answering
Raphael Gruber (University of Innsbruck), Adam Jatowt (University of Innsbruck)
Data SynthesisLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Generated and made publicly available a complex temporal question dataset named COMPLEXTEMPQA containing over 100 million question-answer pairs, covering multi-domain events, entities, and time periods from 1987 to 2023;
Composable Cross-prompt Essay Scoring by Merging Models
Sanwoo Lee (Peking University), Yunfang Wu (Peking University)
ClassificationHyperparameter SearchTransformerLarge Language ModelText
🎯 What it does: Under no-source data conditions, the study proposes achieving cross-prompt automatic essay scoring through model fusion, avoiding re-access to source data and large-scale joint training;
Compositional Generalisation for Explainable Hate Speech Detection
Agostina Calabrese (University of Edinburgh), Mirella Lapata (University of Edinburgh)
ClassificationData SynthesisExplainability and InterpretabilityLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: This paper studies the compositional generalization problem in hate speech detection, constructing a balanced synthetic dataset U-PLEAD and a test benchmark TARGET, and enhancing model generalization and interpretability by adding U-PLEAD to real data;
Compound AI Systems Optimization: A Survey of Methods, Challenges, and Future Directions
Yu-Ang Lee (National Taiwan University), Yun-Nung Chen (National Taiwan University)
OptimizationMeta LearningSupervised Fine-TuningReinforcement LearningAgentic AITextReview/Survey Paper
🎯 What it does: Reviews optimization methods, challenges, and future directions for composite AI systems (composed of components such as multimodal, tools, and agents), proposes a unified graph structure and conditional edge formulation, and systematically summarizes 26 representative works by constructing a 2×2 framework based on structural flexibility and learning signals.
Comprehensive and Efficient Distillation for Lightweight Sentiment Analysis Models
Guangyu Xie (Harbin Institute of Technology), Ruifeng Xu (Harbin Institute of Technology)
ClassificationKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose a lightweight sentiment analysis model distillation framework named COMPEFFDIST, which is based on attribute-driven automatic instruction construction and difficulty-based data filtering.
Computational Analysis of Character Development in Holocaust Testimonies
Esther Shizgal (Hebrew University of Jerusalem), Omri Abend (Hebrew University of Jerusalem)
ClassificationTransformerLarge Language ModelText
🎯 What it does: Segmented 1000 Holocaust survivor interview texts, filtered religious content, determined belief and practice tendencies using LLMs, constructed religious trajectories, and clustered analysis to identify multiple common trajectory patterns.
Computational Analysis of Conversation Dynamics through Participant Responsivity
Margaret Hughes (Massachusetts Institute of Technology), Jad Kabbara (Massachusetts Institute of Technology)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper defines and quantifies responsivity in dialogues—whether a speaker's utterance responds to the previous one—and develops a method for evaluating dialogue quality;
Conan-Embedding-v2: Training an LLM from Scratch for Text Embeddings
Shiyu Li (Basic Algorithm Center, PCG, Tencent), Xi Chen (Basic Algorithm Center, PCG, Tencent)
RetrievalRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
🎯 What it does: Proposed and trained a new 1.4B-parameter LLM called Conan-embedding-v2, and fine-tuned it from scratch as a text embedding model
Concept-pedia: a Wide-coverage Semantically-annotated Multimodal Dataset
Karim Ghonim (Sapienza University of Rome), Roberto Navigli (Sapienza University of Rome)
RecognitionRetrievalVision Language ModelContrastive LearningImageMultimodalityBenchmark
🎯 What it does: Proposed an automatic annotation method based on Wikipedia hyperlinks, generating a multimodal dataset called Concept-edia covering 165,000 concepts, and built a manually verified benchmark for visual concept recognition named Concept-10k with 10,000 concepts.
ConCISE: Confidence-guided Compression in Step-by-step Efficient Reasoning
Ziqing Qiao (Tsinghua University), Yaoxue Zhang (Tsinghua University)
CompressionComputational EfficiencySupervised Fine-TuningTextBenchmarkChain-of-Thought
🎯 What it does: Proposes the confidence-based reflection suppression framework CONCISE, generating more concise reasoning chains;
CondAmbigQA: A Benchmark and Dataset for Conditional Ambiguous Question Answering
Zongxi Li (Lingnan University), S. Joe Qin (Lingnan University)
Explainability and InterpretabilityLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Create the CondAmbigQA benchmark, focusing on retrieval-supported conditional identification and conditional answer generation.
CondenseLM: LLMs-driven Text Dataset Condensation via Reward Matching
Cheng Shen (Nanyang Technological University), Joey Tianyi Zhou (Agency for Science, Technology and Research (A*STAR))
CompressionTransformerLarge Language ModelText
🎯 What it does: Proposed the CondenseLM framework, which effectively compresses text datasets by combining large language models (LLM) with reward matching.
Conditional [MASK] Discrete Diffusion Language Model
Hyukhun Koh (Seoul National University), Kyomin Jung (Seoul National University)
GenerationLarge Language ModelDiffusion modelText
🎯 What it does: Propose the Diffusion-EAGS framework, which integrates conditional masked language models with discrete diffusion models, leveraging conditional Markov random fields to achieve iterative text generation.
Confidence-guided Refinement Reasoning for Zero-shot Question Answering
Youwon Jang (Seoul National University), Byoung-Tak Zhang (Seoul National University)
Explainability and InterpretabilityImageVideoTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Propose an untrained, confidence-based multi-step reasoning framework called C2R to enhance zero-shot performance in cross-modal tasks such as text, image, and video question answering.
Conflict-Aware Soft Prompting for Retrieval-Augmented Generation
Eunseong Choi (Sungkyunkwan University), Jongwuk Lee (Sungkyunkwan University)
GenerationRetrievalKnowledge DistillationTransformerPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Proposed the CARE framework, which introduces a context evaluator into RAG systems. By utilizing soft context embeddings, the framework identifies and resolves conflicts between retrieved context and LLM's internal knowledge, thereby enhancing the robustness of retrieval-augmented generation.
Conflicting Needles in a Haystack: How LLMs behave when faced with contradictory information
Murathan Kurfali, Robert Östling (Stockholm University)
Explainability and InterpretabilityTransformerLarge Language ModelTextBenchmark
🎯 What it does: Constructed a multi-needle (multi-conflict information) version of the Needle-in-a-Haystack evaluation framework, systematically injecting three mutually exclusive 'needles' into long texts and exploring LLM decision-making under conflicting information by varying factors such as position, repetition, layout, and semantic relevance;
Confounding Factors in Relating Model Performance to Morphology
Wessel Poelman (KU Leuven), Miryam de Lhoneux (KU Leuven)
Explainability and InterpretabilityData-Centric LearningLarge Language ModelText
🎯 What it does: This paper investigates the relationship between morphological complexity and language model performance, and identifies confounding factors in experimental design.
Connecting the Knowledge Dots: Retrieval-augmented Knowledge Connection for Commonsense Reasoning
Junho Kim (Korea University), SangKeun Lee (Korea University)
Large Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose RECONNECT, a method that transforms indirect documents into direct explanations for specific questions by using explanations to guide retrieval and knowledge connection, thereby enhancing the performance of large language models in commonsense reasoning tasks.
ConsistentChat: Building Skeleton-Guided Consistent Multi-Turn Dialogues for Large Language Models from Scratch
Jiawei Chen (Chinese Academy of Sciences), Xianpei Han (Chinese Academy of Sciences)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Proposed a multi-turn dialogue instruction generation framework based on a dialogue intent skeleton, and generated the ConsistentChat dataset from scratch, aiming to enhance the consistency and coherence of large language models in multi-turn dialogues.
Constrained Non-negative Matrix Factorization for Guided Topic Modeling of Minority Topics
Seyedeh Fatemeh Ebrahimi (Tampere University), Jaakko Peltonen (Tampere University)
OptimizationRepresentation LearningText
🎯 What it does: Propose a guided topic model based on Constrained Non-negative Matrix Factorization (Constrained NMF), which utilizes a seed word list and soft confidence constraints to uncover low-frequency but domain-critical minority topics.
ConstraintLLM: A Neuro-Symbolic Framework for Industrial-Level Constraint Programming
Weichun Shi (University of Chinese Academy of Sciences), Jian Zhang (University of Chinese Academy of Sciences)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Constructed and trained a neural-symbolic framework called ConstraintLLM specifically for industrial constraint planning, utilizing LLM to automatically generate and correct CP models.
Constructions are Revealed in Word Distributions
Joshua Rozner (Stanford University), Cory Shain (University of Texas at Austin)
Representation LearningTransformerLarge Language ModelText
🎯 What it does: Utilizes the pre-trained language model RoBERTa through masking and intervention to calculate global affinity (probability of a word and its context) and local affinity (distributional differences between word pairs), thereby identifying syntactic constructions and their internal interactions.
Context and POS in Action: A Comparative Study of Chinese Homonym Disambiguation in Human and Language Models
Xie Chenwei, William Shiyuan Wang (Hong Kong Polytechnic University)
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: Studying how humans and large language models (LLMs) resolve ambiguities in Chinese homographs under different contextual and part-of-speech conditions, and comparing the behaviors of both with internal model metrics.
Context is Gold to find the Gold Passage: Evaluating and Training Contextual Document Embeddings
Max Conti (Iluin Technology), Pierre Colombo (Equall.ai)
RetrievalRepresentation LearningTransformerSupervised Fine-TuningContrastive LearningTextBenchmark
🎯 What it does: This paper proposes a dedicated benchmark (ConTEB) for evaluating the ability of retrieval models to utilize full-text context, and designs a lightweight post-training method called InSeNT (In-sequence Negative Training). By combining Late Chunking with contrastive learning, it significantly enhances context-aware text embeddings without compromising the original model performance.
Context Reasoner: Incentivizing Reasoning Capability for Contextualized Privacy and Safety Compliance via Reinforcement Learning
Wenbin Hu (Hong Kong University of Science and Technology), Yangqiu Song (Hong Kong University of Science and Technology)
Safty and PrivacyTransformerLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: This paper constructs a legal framework based on Contextual Integrity (CI) and employs reinforcement learning to enhance the situational reasoning capabilities of large language models in terms of security and privacy compliance.
Context-aware Biases for Length Extrapolation
Ali Veisi (Algonet), Amir M. Mansourian (Algonet)
RetrievalComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Proposed a context-aware relative position encoding called CABLE to enhance the length extrapolation capability of Transformers.
Context-Aware Hierarchical Taxonomy Generation for Scientific Papers via LLM-Guided Multi-Aspect Clustering
Kun Zhu (Harbin Institute Of Technology), Bing Qin (Harbin Institute Of Technology)
ClassificationTransformerLarge Language ModelTextBenchmark
🎯 What it does: Generate multi-dimensional semantic dimensions of papers dynamically using LLM, encode summaries based on these dimensions, independently cluster each dimension, and utilize dynamic search to determine the optimal clustering scheme, thereby constructing a context-aware hierarchical scientific paper classification system from top to bottom.
Context-Aware Membership Inference Attacks against Pre-trained Large Language Models
Hongyan Chang (Mohamed bin Zayed University of Artificial Intelligence), Reza Shokri (National University of Singapore)
Adversarial AttackTransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose a context-aware membership inference attack (CAMIA) for pre-trained large language models, determining whether a data point was present in the training set by analyzing the perplexity dynamics of each token.
Continuous-Time Attention: PDE-Guided Mechanisms for Long-Sequence Transformers
Yukun Zhang (Chinese University of Hong Kong), Xueqing Zhou (Fudan University)
ClassificationGenerationTransformerText
🎯 What it does: Propose a Continuous-Time Attention framework, integrating partial differential equations (PDEs) (diffusion, wave, reaction-diffusion) driven continuous-time dynamics into the self-attention mechanism of Transformers, enabling attention weights to evolve over pseudo-time.
Continuously Steering LLMs Sensitivity to Contextual Knowledge with Proxy Models
Yilin Wang (Xi'an Jiaotong University), Minnan Luo (Xi'an Jiaotong University)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose the CSKS framework, which continuously adjusts the sensitivity of large LLMs to contextual knowledge by leveraging the output differences between two small proxy models, without modifying the large model's weights.
Contra4: Evaluating Contrastive Cross-Modal Reasoning in Audio, Video, Image, and 3D
Artemis Panagopoulou (Salesforce AI Research), Juan Carlos Niebles (University of Pennsylvania)
RetrievalRepresentation LearningLarge Language ModelSupervised Fine-TuningVision Language ModelContrastive LearningImageVideoMultimodalityPoint CloudBenchmarkAudio
🎯 What it does: Propose the Contra4 dataset to evaluate models' contrastive cross-modal reasoning capabilities across four modalities (image, audio, video, 3D) and conduct benchmark testing on this task.
Controllable Memorization in LLMs via Weight Pruning
Chenjie Ni (Northwestern University), Yanfu Zhang (William and Mary)
CompressionComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Proposed a gradient-based weight pruning framework for controllably adjusting the memory rate of large language models, capable of both suppressing and amplifying the model's memory of training data.
Controlled Generation for Private Synthetic Text
Zihao Zhao (Johns Hopkins University), Anjalie Field (Johns Hopkins University)
GenerationData SynthesisSafty and PrivacyTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextBiomedical DataElectronic Health Records
🎯 What it does: Proposed a privacy-preserving synthetic text generation method based on entity-aware control codes, combining in-context learning and prefix tuning variants.
Convergence and Divergence of Language Models under Different Random Seeds
Finlay Fehlauer (ETH Zürich), Tiago Pimentel (ETH Zürich)
TransformerLarge Language ModelText
🎯 What it does: Study the convergence and divergence behaviors of language model pre-training under different random seeds, quantify four convergence stages of pre-training, and further analyze convergence differences through conditions such as word frequency and part-of-speech.
ConvSearch-R1: Enhancing Query Reformulation for Conversational Search with Reasoning via Reinforcement Learning
Changtai Zhu (Fudan University), Xipeng Qiu (Fudan University)
RetrievalTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: Propose ConvSearch-R1, a completely unsupervised conversational query rewriting framework that utilizes reinforcement learning to self-explore and optimize rewriting effectiveness through retrieval signals;
CoPL: Collaborative Preference Learning for Personalizing LLMs
Youngbin Choi (POSTECH), Dongwoo Kim (POSTECH)
Recommendation SystemGraph Neural NetworkLarge Language ModelMixture of ExpertsText
🎯 What it does: Propose CoPL, which combines graph-structured collaborative filtering (GCF) and MoLE expert models to personalize LLMs, achieving user preference learning under sparse annotations.
CopySpec: Accelerating LLMs with Speculative Copy-and-Paste
Razvan-Gabriel Dumitru (University of Arizona), Mihai Surdeanu (University of Arizona)
Computational EfficiencyLarge Language ModelText
🎯 What it does: This paper proposes a method called CopySpec, which can automatically identify and copy repeated token sequences in the context during LLM inference, thereby reducing unnecessary computations.
Correct-Detect: Balancing Performance and Ambiguity Through the Lens of Coreference Resolution in LLMs
Amber Shore (Portland State University), So Young Lee (Miami University)
TransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: This paper investigates large language models (LLMs) in the coreference resolution task, aiming to accurately select referential relationships while identifying when unresolved ambiguity exists, and proposes CORRECT-DETECT to measure the trade-off between these two conflicting capabilities.
Correlation-Aware Example Selection for In-Context Learning with Nonsymmetric Determinantal Point Processes
Qiunan Du (National University of Defense Technology), Dongsheng Li (National University of Defense Technology)
Data-Centric LearningMeta LearningTransformerLarge Language ModelText
🎯 What it does: This paper proposes a framework based on Asymmetric Discriminative Point Processes (NDPP) for example selection, aiming to enhance the in-context learning (ICL) performance of large language models. By constructing a pseudo-annotated dataset, the framework uses LLM feedback to score example subsets, trains the NDPP model with Kernel Decomposition Maximum Likelihood Estimation (KD-MLE), and during inference employs query-aware kernel adaptation and MAP inference to achieve customized example set selection.
Corrupted but Not Broken: Understanding and Mitigating the Negative Impacts of Corrupted Data in Visual Instruction Tuning
Yunhao Gou (Southern University of Science and Technology), Yu Zhang (Southern University of Science and Technology)
Data-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelMultimodality
🎯 What it does: Investigated the negative impact of corrupted data in visual instruction tuning (VIT) and proposed a corruption-robust training scheme that leverages the self-validation capability of MLLM for sample screening and retraining.
Cost-Optimal Grouped-Query Attention for Long-Context Modeling
Yingfa Chen (Tsinghua University), Maosong Sun (Tsinghua University)
OptimizationComputational EfficiencyTransformerText
🎯 What it does: Studied how to decouple the number of attention heads from the hidden dimension and jointly optimize model scale with GQA configuration to find the most cost-effective group query attention (GQA) scheme for long-context reasoning, thereby reducing inference memory and FLOPs.
COUNTDOWN: Contextually Sparse Activation Filtering Out Unnecessary Weights in Down Projection
Jaewon Cheon (Korea University), Pilsung Kang (Seoul National University)
Computational EfficiencyTransformerText
🎯 What it does: This paper proposes the COUNTDOWN framework, which sparsely activates the down-projection matrix of Gated-MLP from the perspective of global weight-weighted summation, and provides two implementation methods: M-COUNTDOWN and D-COUNTDOWN, significantly reducing FLOP and memory usage during inference.
CourtReasoner: Can LLM Agents Reason Like Judges?
Sophia Simeng Han (Yale University), Arman Cohan (Yale University)
TransformerAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed the COURTREASONER benchmark, evaluating LLMs' ability in complete judicial reasoning using expert-annotated US court opinions;
CoVoGER: A Multilingual Multitask Benchmark for Speech-to-text Generative Error Correction with Large Language Models
Zhengdong Yang (Kyoto University), Chenhui Chu (Kyoto University)
RecognitionGenerationTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkAudio
🎯 What it does: Constructed a multilingual, multitask speech-to-text generative error correction benchmark named CoVoGER, and conducted a systematic evaluation of large language models on this benchmark.
CR4-NarrEmote: An Open Vocabulary Dataset of Narrative Emotions Derived Using Citizen Science
Andrew Piper (McGill University), Robert Budac (University of Alberta)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Collected and constructed the CR4-NarrEmote dataset via the citizen science platform Zooniverse, containing 43,713 sentences and 207,721 open-vocabulary emotion labels;
Creativity in LLM-based Multi-Agent Systems: A Survey
Yi-Cheng Lin (National Taiwan University), Yun-Nung Chen (National Taiwan University)
GenerationTransformerLarge Language ModelAgentic AIPrompt EngineeringImageTextReview/Survey Paper
🎯 What it does: This paper provides a systematic review of creativity in large language model (LLM)-driven multi-agent systems, proposing dimensions such as agent proactivity, personality configuration, creative generation techniques, and evaluation methods, and offering a unified framework and future research directions.
CREPE: Rapid Chest X-ray Report Evaluation by Predicting Multi-category Error Counts
Gihun Cho (Seoul National University), Chang Min Park (Seoul National University)
Data SynthesisExplainability and InterpretabilityComputational EfficiencyHyperparameter SearchTransformerSupervised Fine-TuningTextBiomedical DataBenchmark
🎯 What it does: Propose a fast and interpretable chest X-ray report evaluation metric called CREPE, which directly predicts the count of six categories of clinical errors and sums them to obtain an overall score.
Crisp: Cognitive Restructuring of Negative Thoughts through Multi-turn Supportive Dialogues
Jinfeng Zhou (Tsinghua University), Minlie Huang (Tsinghua University)
Data SynthesisTransformerLarge Language ModelSupervised Fine-TuningTextSequential
🎯 What it does: Proposed the CRDIAL framework, decomposing cognitive restructuring (CR) into two-stage multi-round dialogues, integrating emotional support with multi-channel loop mechanisms, and generating a high-quality bilingual dialogue dataset CRISP based on GPT-4o. Subsequently, trained the Qwen-2.5-7B/14B models CRISPERS to achieve human-computer interactive psychotherapy.
CRITICTOOL: Evaluating Self-Critique Capabilities of Large Language Models in Tool-Calling Error Scenarios
Shiting Huang (University of Science and Technology of China), Feng Zhao (University of Science and Technology of China)
TransformerLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: Designed and constructed the CRITICTOOL evaluation benchmark to fine-grainedly assess large language models' self-criticism capabilities in tool call error scenarios (identifying errors, analyzing errors, correcting errors, retrying/skipping/terminating).
CROP: Contextual Region-Oriented Visual Token Pruning
Jiawei Guo (Institute of Automation, Chinese Academy of Sciences), Yu Zhou (Institute of Automation, Chinese Academy of Sciences)
Computational EfficiencyTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: In the visual question answering task, a context-region-based visual token pruning framework called CROP is proposed. It first uses a lightweight VLM to locate continuous visual regions related to the question, then significantly reduces the number of tokens through two strategies—pre-LLM compression (PLC) and intra-LLM pruning (ILP)—while maintaining high-quality answers;
Cross-Document Cross-Lingual NLI via RST-Enhanced Graph Fusion and Interpretability Prediction
Mengying Yuan (Wuhan University), Donghong Ji (Wuhan University)
ClassificationExplainability and InterpretabilityGraph Neural NetworkTransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposed the cross-document cross-lingual reasoning (CDCL-NLI) task and constructed the corresponding dataset, designing a model based on RST-enhanced graph fusion and EDU-level explanations;
Cross-domain Rumor Detection via Test-Time Adaptation and Large Language Models
Yuxia Gong (Inner Mongolia University), Huaiwen Zhang (Inner Mongolia University)
ClassificationDomain AdaptationGraph Neural NetworkLarge Language ModelSupervised Fine-TuningContrastive LearningTextGraph
🎯 What it does: Proposed a test-time adaptation framework for cross-domain rumor detection, TARD2, which includes self-supervised contrastive learning on graph structure and features, as well as model adaptation using pseudo-labels generated by large language models (LLMs).
Cross-MoE: An Efficient Temporal Prediction Framework Integrating Textual Modality
Ruizheng Huang (University of Electronic Science and Technology of China), Yong Wang (University of Electronic Science and Technology of China)
Computational EfficiencyMixture of ExpertsTextMultimodalityTime Series
🎯 What it does: This paper proposes the Cross-MoE framework, which is compatible with any time series model, achieving efficient fusion of text and time series through MoE and Cross-Ranker.
CrystalICL: Enabling In-Context Learning for Crystal Generation
Ruobing Wang (Jilin University), Xin Wang (New York University Shanghai)
GenerationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTabularPhysics Related
🎯 What it does: Proposed the CrystalICL model for crystal generation, fully leveraging the few-shot context learning (ICL) of large language models (LLMs), and enhancing performance through space-group-based crystal tokenization, condition-structure hybrid instruction tuning, and multi-task instruction tuning.
CTCC: A Robust and Stealthy Fingerprinting Framework for Large Language Models via Cross-Turn Contextual Correlation Backdoor
Zhenhua Xu (Zhejiang University), Meng Han (Zhejiang University)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Designed and implemented a backdoor fingerprint framework triggered by cross-turn dialogue context relevance for verifying the ownership of large language models in black-box environments.
Culture Cartography: Mapping the Landscape of Cultural Knowledge
Caleb Ziems (Stanford University), Diyi Yang (Stanford University)
Data-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIPrompt EngineeringText
🎯 What it does: By integrating a hybrid active human-AI collaboration method (CULTURE CARTOGRAPHY) and the tool CULTURE EXPLORER, systematically collect culturally missing knowledge from diverse cultures and build a tree-structured QA knowledge base for large language models (LLMs).
Current Semantic-change Quantification Methods Struggle with Discovery in the Wild
Khonzoda Umarova (Cornell University), Laerdon Kim (Cornell University)
TransformerLarge Language ModelText
🎯 What it does: This paper evaluates the performance of word sense change detection methods in the 'wild discovery' scenario and proposes a top-k ranking-based evaluation framework. It further expands the annotated data of the SemEval-EN and LiverpoolFC corpora by 85% and 90%, respectively, to verify the actual semantic changes of high-scoring words.
CYCLE-INSTRUCT: Fully Seed-Free Instruction Tuning via Dual Self-Training and Cycle Consistency
Zhanming Shen (Zhejiang University), Junbo Zhao (Zhejiang University)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Built a seed-free and external teacher model-free instruction tuning framework called CYCLE-INSTRUCT, which automatically generates question-answer pairs from unannotated raw text using dual-model self-training and cyclic consistency;
D-CoDe: Scaling Image-Pretrained VLMs to Video via Dynamic Compression and Question Decomposition
Yiyang Huang (Northeastern University), Yun Fu (Northeastern University)
Computational EfficiencyConvolutional Neural NetworkTransformerLarge Language ModelVision Language ModelVideoTextMultimodalityChain-of-Thought
🎯 What it does: Propose the D-CoDe framework, leveraging dynamic compression and problem decomposition to transfer image-pretrained VLM without training to video tasks;
D-RAG: Differentiable Retrieval-Augmented Generation for Knowledge Graph Question Answering
Guangze Gao (Chinese Academy of Sciences), Weiming Hu (Chinese Academy of Sciences)
GenerationRetrievalGraph Neural NetworkTransformerPrompt EngineeringTextGraphRetrieval-Augmented Generation
🎯 What it does: Designed a differentiable retrieval-augmented generation framework, D-RAG, for knowledge graph question answering, enabling end-to-end gradient passing between the retriever and generator.
DA-Pred: Performance Prediction for Text Summarization under Domain-Shift and Instruct-Tuning
Anum Afzal (Technical University of Munich), Alexander Fabbri
Domain AdaptationTextBenchmark
🎯 What it does: Propose a prediction framework called DA-Pred that leverages known high-resource domains and model performance to predict text summarization performance in low-resource domains.
DAMON: A Dialogue-Aware MCTS Framework for Jailbreaking Large Language Models
Xu Zhang (Wangxuan Institute of Computer Technology Peking University), Xiaojun Wan (Wangxuan Institute of Computer Technology Peking University)
Adversarial AttackTextSequential
🎯 What it does: Propose DAMON, a multi-round dialogic jailbreaking framework based on Monte Carlo Tree Search (MCTS), which automatically generates sub-instruction sequences to bypass the safety alignment of large language models (LLMs);
DART: Distilling Autoregressive Reasoning to Silent Thought
Nan Jiang (Nanjing University), Shaobing Lian (Tencent Inc)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: Propose the DART framework, enabling large language models to perform reasoning tasks in non-autoregressive mode;
DASA-Trans-STM: Adaptive Efficient Transformer for Short Text Matching using Data Augmentation and Semantic Awareness
Jiguo Liu (Chinese Academy of Sciences), Dali Zhu (University of Chinese Academy of Sciences)
RetrievalComputational EfficiencyRepresentation LearningTransformerLarge Language ModelTextGraph
🎯 What it does: Proposed the DASA-Trans-STM framework, which enhances data through LLM-generated synonymous sentences and improves Chinese short text matching using N-HowNet semantic information, employing a word lattice graph and adaptive Transformer encoding to build a short text matching model.
Data Descriptions from Large Language Models with Influence Estimation
Chaeri Kim (Ulsan National Institute of Science and Technology), Taehwan Kim (Ulsan National Institute of Science and Technology)
ClassificationData-Centric LearningTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Use large language models to generate textual descriptions of image categories, and select the most useful descriptions for improving the training and zero-shot reasoning of image classification models through influence estimation and CLIP similarity.
Data Drives Unstable Hierarchical Generalization in LMs
Tian Qin (Harvard University), David Alvarez-Melis (Harvard University)
Data-Centric LearningTransformerLarge Language ModelText
🎯 What it does: Examines how the complexity and diversity of training data influence the hierarchical generalization and preference for linear approximation in large language models during controlled syntactic learning tasks, revealing the decisive role of central nested sentences and syntactic tree diversity in determining model rule learning and training stability.
Data to Defense: The Role of Curation in Aligning Large Language Models Against Safety Compromise
Xiaoqun Liu (Michigan State University), Zhaohan Xi (Binghamton University)
Safty and PrivacyComputational EfficiencyData-Centric LearningLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Proposes the Data to Defense (D2D) framework, which defends against malicious attacks encountered by LLMs during fine-tuning by enhancing perplexity and injecting security-relevant content into any text, applied in the pre-, mid-, and post-stages of the customization lifecycle.
Data-Efficient Hate Speech Detection via Cross-Lingual Nearest Neighbor Retrieval with Limited Labeled Data
Faeze Ghorbanpour (Technische Universitaet Muenchen), Alexander Fraser (Technische Universitaet Muenchen)
ClassificationDomain AdaptationTransformerSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: Propose a cross-lingual nearest neighbor retrieval method that selects the most relevant source language instances from a multilingual retrieval pool using only a small number of target language annotated samples, and then mixes these instances with target data for fine-tuning to improve hate speech detection performance in low-resource languages.
Data-Efficient Selection via Grammatical Complexity in Continual Pre-training of Domain-Specific LLMs
Yizhou Ying (Fudan University), Yanghua Xiao (Fudan University)
Data-Centric LearningTransformerLarge Language ModelTextFinance Related
🎯 What it does: This paper proposes a grammar complexity (GC) evaluation framework based on lexical and syntactic diversity, combined with cumulative distribution function (CDF) for balanced sampling to efficiently select high-quality samples in the financial domain during continuous pre-training.
Database-Augmented Query Representation for Information Retrieval
Soyeong Jeong (KAIST), Jong C. Park (KAIST)
RetrievalRepresentation LearningGraph Neural NetworkTransformerTabular
🎯 What it does: Propose a framework named DAQu that utilizes multi-table metadata from relational databases to generate incremental representations for retrieval queries, addressing the issue of insufficient information in short queries.
DatawiseAgent: A Notebook-Centric LLM Agent Framework for Adaptive and Robust Data Science Automation
Ziming You, Yu Huang (Peking University)
Data-Centric LearningAI Code AssistantTransformerLarge Language ModelAgentic AITextBenchmark
🎯 What it does: Proposes DatawiseAgent, an LLM agent framework based on computational notebooks, which achieves end-to-end, adaptive, and robust data science automation through a four-phase process (DFS-style planning, incremental execution, self-debugging, post-filtering) driven by a finite state machine, under a unified Markdown+code cell interaction representation.
Date Fragments: A Hidden Bottleneck of Tokenization for Temporal Reasoning
Gagan Bhatia (University of Aberdeen), Wei Zhao (University of Aberdeen)
Explainability and InterpretabilityTransformerLarge Language ModelTextBenchmark
🎯 What it does: Investigated the impact of date fragmentation caused by subword tokenization on the performance of large language models in temporal reasoning tasks;
DCIS: Efficient Length Extrapolation of LLMs via Divide-and-Conquer Scaling Factor Search
Lei Yang (Tianjin University), Deyi Xiong (Tianjin University)
Computational EfficiencyHyperparameter SearchTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: This paper proposes a Divide-and-Conquer Incremental Search (DCIS) algorithm based on RoPE to efficiently find suitable frequency scaling factors, thereby expanding the context window of LLMs without significantly increasing fine-tuning costs.
DCP: Dual-Cue Pruning for Efficient Large Vision-Language Models
Lei Jiang (University of Science and Technology of China), Xiaohua Xu (University of Science and Technology of China)
Computational EfficiencyTransformerVision Language ModelMultimodality
🎯 What it does: Propose the Dual-Cue Pruning (DCP) framework, which performs training-free pruning of visual tokens in vision-language models by leveraging dual cues from text and vision;
DCR: Quantifying Data Contamination in LLMs Evaluation
Cheng Xu (University College Dublin), Tahar Kechadi
Explainability and InterpretabilityLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Proposed and validated the Data Corruption Risk (DCR) framework to detect and quantify the degree of data contamination in large language models during benchmark evaluations, and adjusted model accuracy using the DCR Factor.
DDO: Dual-Decision Optimization for LLM-Based Medical Consultation via Multi-Agent Collaboration
Zhihao Jia (Central South University), Jianxin Wang (Central South University)
OptimizationTransformerLarge Language ModelReinforcement LearningAgentic AIBiomedical DataElectronic Health Records
🎯 What it does: Proposed the DDO (Dual-Decision Optimization) framework, which separates and optimizes two decision-making processes in medical consultations: symptom inquiry and disease diagnosis, through multi-agent collaboration.
Debatable Intelligence: Benchmarking LLM Judges via Debate Speech Evaluation
Noy Sternlicht (Hebrew University of Jerusalem), Noam Slonim (IBM Research)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Proposed and implemented a method using long-form debate speeches as a benchmark for LLM judges (LLMaJ), systematically evaluating LLMs' performance on multi-dimensional debate assessment tasks.
Debate-to-Detect: Reformulating Misinformation Detection as a Real-World Debate with Large Language Models
Chen Han (University of Chinese Academy of Sciences), Xijin Tang (University of Chinese Academy of Sciences)
ClassificationExplainability and InterpretabilityTransformerLarge Language ModelAgentic AIText
🎯 What it does: Propose the Debate-to-Detect (D2D) framework, transforming rumor detection into a multi-agent debate process, leveraging domain knowledge, five-phase debate, and multi-dimensional evaluation to achieve credible and interpretable truthfulness judgments.
DEBATE, TRAIN, EVOLVE: Self‐Evolution of Language Model Reasoning
Gaurav Srivastava (Virginia Tech), Xuan Wang (Virginia Tech)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringText
🎯 What it does: Proposes the DEBATE, TRAIN, EVOLVE (DTE) framework, which employs self-generated reasoning trajectories from multi-agent debates to conduct self-evolution training on a single language model.
Debiasing Multilingual LLMs in Cross-lingual Latent Space
Qiwei Peng (University of Copenhagen), Anders Søgaard (University of Copenhagen)
Representation LearningData-Centric LearningTransformerAuto EncoderText
🎯 What it does: Perform debiasing in the cross-lingual latent space of multilingual LLMs.
Decoding Dense Embeddings: Sparse Autoencoders for Interpreting and Discretizing Dense Retrieval
Seongwan Park (SungKyunKwan University), Youngjoong Ko (SungKyunKwan University)
RetrievalExplainability and InterpretabilityComputational EfficiencyRepresentation LearningLarge Language ModelAuto EncoderText
🎯 What it does: Use a sparse autoencoder (SAE) to decompose the dense vectors of the Dense Passage Retrieval model into interpretable latent concepts, and build a concept-level sparse retrieval framework called CL-SR;
Decoding the Rule Book: Extracting Hidden Moderation Criteria from Reddit Communities
Youngwoo Kim (University of Virginia), Thomas Hartvigsen (University of Virginia)
Explainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelText
🎯 What it does: By using an interpretable Partial Attention Transformer (PAT) to extract vocabulary-level moderation criteria from Reddit subcommunity historical deletion records, and constructing a CriteriaMatrix to compare implicit norms across different subcommunities.
DecoupleSearch: Decouple Planning and Search via Hierarchical Reward Modeling
Hao Sun (Alibaba Group), Yan Zhang (Alibaba Group)
RetrievalOptimizationTransformerLarge Language ModelAgentic AITextRetrieval-Augmented Generation
🎯 What it does: In the Agentic RAG framework, decoupling planning and retrieval is achieved by introducing a dual-value model, and Monte Carlo Tree Search (MCTS) is used to construct a reasoning tree, enabling independent evaluation and optimization of planning and retrieval steps during the reasoning process.
Deep Associations, High Creativity: A Simple yet Effective Metric for Evaluating Large Language Models
Ziliang Qiu (University of Illinois Urbana Champaign), Renfen Hu (Beijing Normal University)
TextBenchmarkChain-of-Thought
🎯 What it does: Proposes PACE (Parallel Association Chain Evaluation) — a novel metric for evaluating the creativity of large language models (LLM) by leveraging parallel association chains.
DeepResearcher: Scaling Deep Research via Reinforcement Learning in Real-world Environments
Yuxiang Zheng (Shanghai Jiao Tong University), Pengfei Liu (Shanghai Jiao Tong University)
Large Language ModelReinforcement LearningAgentic AITextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: DeepResearcher proposes an end-to-end trained LLM deep research agent that performs adaptive retrieval and information integration in real-world web search environments through reinforcement learning.
DeepResonance: Enhancing Multimodal Music Understanding via Music-centric Multi-way Instruction Tuning
Zhuoyuan Mao (Sony Group Corporation), Yuki Mitsufuji (Sony Group Corporation)
TransformerLarge Language ModelVision Language ModelImageVideoTextMultimodalityAudio
🎯 What it does: Developed the multimodal music understanding large language model DeepResonance, achieving fusion of music, text, image, and video through multi-path instruction tuning.
DeepWell-Adol: A Scalable Expert-Based Dialogue Corpus for Adolescent Positive Mental Health and Wellbeing Promotion
Wenyu Qiu (Tsinghua University), Shiguang Ni (Tsinghua University)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Constructed a scalable, expert-driven dialogue corpus named DeepWell-Adol for promoting positive mental health in adolescents