ACL 2026 Papers — Page 5
Annual Meeting of the Association for Computational Linguistics · 2296 papers
COMPASS: A Framework for Evaluating Organization-Specific Policy Alignment in LLMs
Dasol Choi (AIM Intelligence), Minsuk Kahng
Safty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the COMPASS framework to systematically evaluate the alignment of large language models (LLMs) under organizational-specific allow and prohibit list policies, automatically generating basic and edge queries and using LLM judges to assess the rejection and compliance of responses.
COMPASS: Enhancing Agent Long-Horizon Reasoning with Evolving Context
Guangya Wan (Google Cloud AI), Zizhao Zhang (Google Cloud AI)
AI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AITextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the COMPASS framework, combining a main execution Agent, Meta-Thinker, and Context Manager to enhance reasoning and tool invocation stability in long-term tasks
Compatibility-Aware Dynamic Fine-Tuning for Large Language Models
Yucheng Zhou (University of Macau), Jianbing Shen (University of Macau)
OptimizationRepresentation LearningData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningDiffusion modelTextMultimodalityBenchmark
🎯 What it does: Proposes a compatibility-aware dynamic fine-tuning (CADFT) method, improving traditional supervised fine-tuning (SFT) and dynamic fine-tuning (DFT) by reducing optimization variance through sample-level compatibility weighting and delaying the rewrite of incompatible examples when necessary.
Compete to Complete: Co-opetition Adversarial Learning for Retrieval-Augmented Generation
Xin Liu (Institute of Computing Technology, Chinese Academy of Sciences), Xueqi Cheng (Institute of Computing Technology, Chinese Academy of Sciences)
GenerationData SynthesisRetrievalTransformerLarge Language ModelGenerative Adversarial NetworkTextRetrieval-Augmented Generation
🎯 What it does: Proposes the CARL framework, which trains the retriever and generator through collaborative competitive adversarial learning to enhance the overall performance of retrieval-augmented generation.
Compiling Activation Steering into Weights via Null-Space Constraints for Stealthy Backdoors
Rui Yin (Zhejiang University), Shouling Ji (Zhejiang University)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Designed and implemented a backdoor technique called STEEREDIT, which embeds activation steering into the weights of the model, enabling it to continuously generate malicious outputs when receiving specific triggers, while maintaining safety and original performance when no trigger is present.
Compositional Steering of Large Language Models with Steering Tokens
Gorjan Radevski (Independent), Goran Glavaš (University of Würzburg)
GenerationData SynthesisComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes a method based on 'steering tokens' to achieve multi-behavior combination control at the input level of large language models (LLMs), enabling the generation of outputs that meet various verifiable constraints (such as length, language, format, etc.) simultaneously, while keeping the model parameters frozen.
Compressing LLM Knowledge into Graph Representations for Text-attributed Graphs Learning
Runhuai Chen (Southeast University), Beilun Wang (Southeast University)
ClassificationComputational EfficiencyKnowledge DistillationRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextGraph
🎯 What it does: This paper proposes a text attribute graph (TAG) learning framework that utilizes the knowledge of large language models (LLMs) during training, compressing it into graph neural networks (GNNs) so that no reliance on LLMs is required during inference. The framework achieves knowledge compression and instance refinement through a hierarchical Proxy-Purifier module.
Compressing then Matching: An Efficient Pre-training Paradigm for Multimodal Embedding
Da Li (State Key Laboratory of AI Safety Institute of Computing Technology Chinese Academy of Sciences), Guorui Zhou (Kuaishou Technology)
Data SynthesisComputational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringVision Language ModelAuto EncoderContrastive LearningImageTextMultimodality
🎯 What it does: Propose a two-stage pre-training framework called CoMa, which first uses compression pre-training to enable multi-modal large models to generate comprehensive compressed tokens, and then uses contrastive learning to map these tokens into a unified cross-modal embedding space.
CompTab: A Comprehensive Benchmark for Real-World TableQA with Complex Reasoning and Irregular Tables
Zhen Yang (Anhui University), Shu Zhao (Anhui University)
TransformerLarge Language ModelSupervised Fine-TuningAgentic AIPrompt EngineeringTextTabularBenchmark
🎯 What it does: Proposes the CompTab benchmark and the CompTabLLM two-stage training framework to evaluate the performance of large language models on complex reasoning and irregular tables.
Conditional Memory via Scalable Lookup: A New Axis of Sparsity for Large Language Models
Xin Cheng (Peking University), Wenfeng Liang (DeepSeek-AI National Engineering Research Center of New Electronic Publishing Technologies)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose a depth sparse embedding (DSE) based on hash N-gram retrieval as a conditional memory module for LLMs, significantly improving knowledge recall and reasoning efficiency;
Confidence Should Be Calibrated More Than One Turn Deep
Zhaohan Zhang (Queen Mary University of London), Ioannis Patras (Queen Mary University of London)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes a multi-round calibration task and introduces two methods, MTCal and ConfChat, to dynamically calibrate model confidence in multi-round dialogues.
Confident, Calibrated, or Complicit: Safety Alignment and Ideological Bias in LLM Hate Speech Detection
Sanjeevan Selvaganapathy (University of Western Australia), Mehwish Nasim (University of Western Australia)
ClassificationSafty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper conducts experiments on the performance of LLMs under different levels of deployed censorship alignment (censorship-as-deployed) in detecting implicit and explicit hate speech, and introduces political persona prompts to examine the model's directional bias and fairness issues.
Conflict-Aware Memory for Embodied Agents: Enhancing Vector Data Quality via Detection Rules
Kexin Ma (National University of Defense Technology), Ruochun Jin (National University of Defense Technology)
Data-Centric LearningRobotic IntelligenceTransformerLarge Language ModelPrompt EngineeringImageTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Propose Conflict Detection Rules (CDRs) to detect and correct conflicting data in vector knowledge bases, improving memory retrieval quality and planning performance of LLM-based embodied agents.
ConfSpec: Efficient Step-Level Speculative Reasoning via Confidence-Gated Verification
Siran Liu (Peking University), Yongchao He (ScitiX AI)
Computational EfficiencyAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsTextBenchmarkChain-of-Thought
🎯 What it does: Propose ConfSpec, a confidence-gated cascade verification framework that achieves efficient acceleration in step-level reasoning;
Conjunctive Prompt Attacks in Multi-Agent LLM Systems
Nokimul Hasan Arif (University of Central Florida), Mengxin Zheng (University of Central Florida)
OptimizationFederated LearningSafty and PrivacyAdversarial AttackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Studied synthetic trigger prompt attacks in multi-agent LLM systems, and proposed an optimization framework based on routing and prompt-level strategies;
ConlangCrafter: Constructing Languages with a Multi-Hop LLM Pipeline
Morris Alper (Tel Aviv University), Gasper Begus (UC Berkeley)
GenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Built ConlangCrafter, a pipeline based on multi-hop LLMs, used to generate constructed languages from scratch with complete phonology, morphology, syntax, vocabulary, and translatable texts.
ConsistRM: Improving Generative Reward Models via Consistency-Aware Self-Training
Yu Liang (Baidu Inc.), Daiting Shi (Baidu Inc.)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringContrastive LearningTextBenchmark
🎯 What it does: Propose ConsistRM, a self-supervised generative reward model (GRM) training framework that does not require manual labels, enhancing the preference alignment capability of dialogue models through a consistency-aware reward mechanism.
Consolidation or Adaptation? PRISM: Disentangling SFT and RL Data via Gradient Concentration
Yang Zhao (Harbin Institute of Technology), Ting Liu (Harbin Institute of Technology)
Computational EfficiencyData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AITextBenchmark
🎯 What it does: Propose a data allocation framework called PRISM, which uses gradient concentration to distinguish between SFT and RL training data, in order to achieve efficient training of LLM agents.
Constructing coherent spatial memory in LLM agents through graph rectification
Puzhen Zhang (Technical University of Munich), Liqiu Meng (Technical University of Munich)
Autonomous DrivingOptimizationReinforcement Learning from Human FeedbackGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringSimultaneous Localization and MappingWorld ModelTextGraphBenchmarkRetrieval-Augmented Generation
🎯 What it does: Construct an LLM-driven navigation graph and provide a graph-based spatial memory framework for repair
Constructing Interpretable Features from Compositional Neuron Groups
Or David Shafran (Tel Aviv University), Mor Geva (Tel Aviv University)
Explainability and InterpretabilityTransformerLarge Language ModelAuto EncoderText
🎯 What it does: Perform semi-non-negative matrix decomposition on the activation of the MLP layer in large language models to extract sparse and interpretable neuron combination features.
ContextLens: Modeling Imperfect Privacy and Safety Context for Legal Compliance
Haoran Li (Beihang University), Yangqiu Song (HKUST)
Safty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes ContextLens, a semi-rule-based framework that leverages LLMs to evaluate privacy and security compliance by chunking regulations and answering questions, while explicitly identifying missing and ambiguous contextual factors.
Contextual Relevance and Adaptive Sampling for LLM-Based Document Reranking
Jerry Huang (University of Illinois at Urbana-Champaign), Tong Zhang (University of Illinois at Urbana-Champaign)
RetrievalComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextRetrieval-Augmented Generation
🎯 What it does: Propose the concept of 'contextual relevance' and develop the TS-SetRank two-stage Bayesian sampling algorithm, which enhances document re-ranking quality by leveraging contextual averaging in the batch relevance judgment generated by LLMs.
Continuous Interpretive Steering for Scalar Diversity
Ye-eun Cho (Sungkyunkwan University)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Studied a mechanism for fine-tuning reasoning in large language models that probes scalar diversity by continuously regulating the mediation direction in the activation space.
ContrastKV: Robust KV Cache Eviction via Contrastive Signal Fusion for Multi-Query Generalization
Xingchi Chen (Pengcheng Laboratory), Hui Li (Tsinghua University)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelContrastive LearningTextBenchmark
🎯 What it does: Propose ContrastKV, a query-free KV cache eviction framework based on contrast signal fusion;
ControlAudio: Tackling Text-Guided, Timing-Indicated and Intelligible Audio Generation via Progressive Diffusion Modeling
Yuxuan Jiang (Tsinghua University), Jun Zhu (Tsinghua University)
GenerationData SynthesisTransformerPrompt EngineeringDiffusion modelTextMultimodalityChain-of-ThoughtAudio
🎯 What it does: Proposed ControlAudio, which utilizes progressive diffusion modeling to achieve controllable text-to-audio generation, supporting precise temporal control and intelligible speech content.
Controllable Contamination Detection for Reliable LLM Evaluation with Statistical Guarantees
Zheng Zhang (University of Science and Technology of China), Dacheng Tao (Nanyang Technological University)
Anomaly DetectionData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
🎯 What it does: Proposes a training data detection framework called FTD based on multiple hypothesis testing, used to identify and filter out samples contaminated by pre-training data in large language model evaluations.
Controllable LLM Reasoning via Sparse Autoencoder-Based Steering
Yi Fang (University of Science and Technology of China), Fuli Feng (University of Science and Technology of China)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringAuto EncoderTextChain-of-Thought
🎯 What it does: Propose to use sparse autoencoders (SAE) to decompose the hidden states of large reasoning models (LRM), thereby achieving fine-grained control over reasoning strategies; develop a two-stage SAE-Steering method to efficiently identify and select strategy-related features; use the identified features for positive and negative control, and verify its effectiveness in error correction.
Controllable Memory Usage: Balancing Anchoring and Innovation in Long-Term Human–Agent Interaction
Zisu Huang (Fudan University), Xiaoqing Zheng (Fudan University)
Data SynthesisReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextSequentialRetrieval-Augmented Generation
🎯 What it does: Propose and implement the controllable memory usage framework SteeM, investigate users' preferences for memory dependency in long-term human-computer interaction, and enable real-time adjustment of memory usage during generation.
Controlling Distributional Bias in Multi-Round LLM Generation via KL-Optimized Fine-Tuning
Yanbei Jiang (University of Melbourne), Fajri Koto (MBZUAI)
GenerationOptimizationData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This paper proposes a distribution alignment method for multi-round LLM generation, utilizing Steering Token Calibration and Kahneman-Tversky Optimization (KTO) to finely control the output attribute distribution.
Controlling Multimodal Conversational Agents with Coverage-Enhanced Latent Actions
Yongqi Li (Wuhan University), Yongbin Li (Zhongguancun Academy)
Computational EfficiencyRepresentation LearningTransformerReinforcement LearningVision Language ModelAuto EncoderContrastive LearningWorld ModelImageTextMultimodality
🎯 What it does: This paper proposes using a compressed latent action space in reinforcement learning fine-tuning for multi-modal dialogue agents, significantly reducing the exploration space and improving training efficiency.
Coordinating Search-Informed Reasoning and Reasoning-Guided Search in Claim Verification
Qisheng Hu (Nanyang Technological University), Wenya Wang (Nanyang Technological University)
RetrievalExplainability and InterpretabilityComputational EfficiencyRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningAgentic AITextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Designed and trained a dual-agent framework called HARIS for multi-hop argument verification, significantly improving the collaborative efficiency of reasoning and retrieval.
CoQuIR: A Comprehensive Benchmark for Code Quality-Aware Information Retrieval
Jiahui Geng (Linköping University), Fakhri Karray (MBZUAI)
RetrievalTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextBenchmark
🎯 What it does: This paper proposes the CoQuIR benchmark to evaluate the performance of code retrieval systems across four quality dimensions (correctness, efficiency, safety, maintainability), and constructs a large multilingual dataset through fine-grained annotation.
CoreCodeBench: Decoupling Code Intelligence via Fine-Grained Repository-Level Tasks
Lingyue Fu (Shanghai Jiao Tong University), Yong Yu (Shanghai Jiao Tong University)
AI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose a configurable warehouse-level evaluation benchmark called CoreCodeBench, and decompose public Python repositories into adjustable difficulty and fine-grained single-function and multi-function tasks through the automated COREPIPE framework, thereby splitting different cognitive dimensions of code intelligence.
CoreGaze: Core Subgraph-Driven Visual Gaze Diffusion for Training-Free Referring Multimodal Large Language Models
Xiaoyang Yi (Nankai University), Jian Zhang (Nankai University)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningGraph Neural NetworkTransformerPrompt EngineeringVision Language ModelDiffusion modelContrastive LearningImageTextMultimodality
🎯 What it does: A training-free CoreGaze framework was developed, which utilizes core subgraphs to drive visual attention diffusion and dynamically prunes visual tokens, thereby enhancing the fine-grained understanding and generation of multimodal large language models in pointing dialogues.
Corpus-Dependent Subcharacter Encoding via HMM-Guided Code Assignment
Tatsuya Hiraoka (Mohamed bin Zayed University of Artificial Intelligence)
ClassificationCompressionData-Centric LearningRecurrent Neural NetworkPrompt EngineeringAuto EncoderContrastive LearningTextReview/Survey Paper
🎯 What it does: This paper proposes a corpus-based sub-character encoding framework called Latom, which models the context of character occurrences using HMM and maps characters to fixed-length, reversible atomic codes (atom code) via the Hungarian algorithm;
Correct When Paired, Wrong When Split: Decoupling and Editing Modality-Specific Neurons in MLLMs
Tingchao Fu (Yunnan University), Wei Zhou (Yunnan University)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Through empirical research, this paper finds that multimodal large language models experience 'editing decoupling failure' during knowledge editing, meaning that knowledge updated under multimodal triggers cannot remain consistent under unimodal inputs, and thus proposes the DECODE method to address this issue;
COSMOS: Connectivity-Oriented Submodular Maximization for Optimal Subgraph Retrieval
Boci Peng (Peking University), Yan Zhang (Peking University)
RetrievalOptimizationKnowledge DistillationRepresentation LearningGraph Neural NetworkTransformerContrastive LearningTextGraphBenchmarkChain-of-Thought
🎯 What it does: Proposes the COSMOS framework, modeling subgraph retrieval in knowledge graph question answering as a connected constrained submodular maximization problem, and provides approximation guarantees;
CostBench: Evaluating Multi-Turn Cost-Optimal Planning and Adaptation in Dynamic Environments for LLM Tool-Use Agents
Jiayu Liu (Hong Kong University of Science and Technology), Yi R. Fung (Hong Kong University of Science and Technology)
OptimizationTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose CostBench benchmark to evaluate the cost-optimal planning and adaptability of LLMs in dynamic environments.
CoSToM: Causal-oriented Steering for Intrinsic Theory-of-Mind Alignment in Large Language Models
Mengfan Li (Huazhong University of Science and Technology), Yang Deng (Singapore Management University)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Propose the COSTOM framework, which first locates the layers containing Theory of Mind (ToM) information in LLMs through causal tracing, and then performs lightweight alignment with activation-oriented adjustments on these layers;
Counteracting the Matthew Effect in Self-Improvement of LVLMs through Head-Tail Re-balancing
Xin Guo (Fudan University), Xuanjing Huang (Fudan University)
OptimizationData-Centric LearningTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelImageTextMultimodality
🎯 What it does: To address the Matthew effect problem that occurs during the self-improvement process of large vision-language models (LVLMs), where simple samples dominate and difficult samples are marginalized, this paper proposes the Head-Tail Rebalance framework, which includes four strategies: threshold clipping, repetition filling, adaptive weighted resampling, and guided sampling, enabling the model to learn different difficulty reasoning trajectories more evenly in each iteration;
CoVerRL: Breaking the Consensus Trap in Label-Free Reasoning via Generator-Verifier Co-Evolution
Teng Pan (Zhejiang University), Yongliang Shen (Zhejiang University)
Data-Centric LearningTransformerLarge Language ModelReinforcement LearningContrastive LearningTextBenchmarkChain-of-Thought
🎯 What it does: Propose the CoVerRL framework, enabling the same model to alternate between generator and validator during training, achieving self-evolving unlabelled reinforcement learning.
CPR-RAG: Clinical Prior-Regularized Retrieval for Anatomy-Aware 3D CT Report Generation
Sungkyu Yang (GIST), Mansu Kim (GIST)
GenerationRetrievalRepresentation LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningImageTextBiomedical DataComputed TomographyRetrieval-Augmented Generation
🎯 What it does: Proposes the CPR-RAG framework, improving 3D CT report generation through organ-level retrieval and clinical comorbidity prior regularization;
CPT-Agent: A Cognitive Process Theory-driven Framework for Student Simulation in Writing Development
Yuhan Chen (Capital Normal University), Wei Song (Capital Normal University)
GenerationData SynthesisExplainability and InterpretabilityKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Propose the CPT-Agent framework, which utilizes cognitive process theory to simulate student writing, generating essays that align with academic reality for evaluating the quality of teacher feedback and teacher training.
CRAFT: Training-Free Cascaded Retrieval for Tabular QA
Adarsh Singh (Arizona State University), Vivek Gupta (Arizona State University)
RetrievalTransformerPrompt EngineeringTextTabularRetrieval-Augmented Generation
🎯 What it does: Proposes CRAFT, an unsupervised cascaded retrieval framework for open-domain table question answering.
Credal Concept Bottleneck Models for Epistemic–Aleatoric Uncertainty Decomposition
Tanmoy Mukherjee (University Artois), Zied Bouraoui (University Artois)
ClassificationExplainability and InterpretabilityTransformerLarge Language ModelMixture of ExpertsContrastive LearningTextMultimodality
🎯 What it does: Propose the CREDENCE framework, which decomposes uncertainty in concept bottleneck models into interpretable epistemic and aleatoric components;
CreditDecoding: Accelerating Parallel Decoding in Diffusion Large Language Models with Trace Credit
Kangyu Wang (Shanghai Jiao Tong University), Weiyao Lin (Shanghai Jiao Tong University)
Computational EfficiencyAI Code AssistantTransformerLarge Language ModelDiffusion modelScore-based ModelTextSequentialBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed and implemented a training-agnostic parallel decoding method called CreditDecoding, aimed at accelerating the inference process of diffusion-based large language models (dLLM).
CRISP: Persistent Concept Unlearning via Sparse Autoencoders
Tomer Ashuach (Technion Israel Institute of Technology), Yonatan Belinkov (Technion Israel Institute of Technology)
Safty and PrivacyExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningAuto EncoderContrastive LearningTextBenchmark
🎯 What it does: Proposed the CRISP method, which utilizes sparse autoencoders (SAE) to automatically identify and suppress target concept features, thereby achieving persistent concept unlearning in large language models (LLMs).
Critical-CoT: A Robust Defense Framework against Reasoning-Level Backdoor Attacks in Large Language Models
Vu Tuan Truong, Long Bao Le (University of Quebec)
Federated LearningSafty and PrivacyExplainability and InterpretabilityAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
🎯 What it does: Proposes Critical-CoT, a defense framework against backdoor attacks on the logical reasoning level of LLMs, and through two-stage fine-tuning (SFT + DPO), enables the model to develop critical thinking, actively identifying and rejecting malicious reasoning steps.
CriticLean: Critic-Guided Reinforcement Learning for Mathematical Formalization
Zhongyuan Peng (Fudan University), Ge Zhang (ByteDance Inc)
OptimizationAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextSequentialBenchmark
🎯 What it does: Propose the CriticLean framework, which utilizes CriticLeanGPT to critique and iteratively optimize Lean4 automation formalization, and construct the CriticLeanBench and FineLeanCorpus datasets.
CrochetBench: Can Vision-Language Models Move from Describing to Doing in Crochet Domain?
Peiyu Li (University of Notre Dame), Nitesh V Chawla
TransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Designed and released CrochetBench, which includes four progressively difficult tasks (knitting pattern recognition, instruction matching, natural language instruction generation, DSL translation), and introduced an executable CrochetPARADE DSL for automated verification.
Cross-Examination Framework: A Task-Agnostic Diagnostic for Information Fidelity in Text-to-Text Generation
Tathagata Raha (M42 Health), Praveenkumar Kanithi
GenerationExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextBiomedical DataBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose a Cross-Examination Framework (CEF) that evaluates the semantic faithfulness of text generation by generating verifiable questions and performing cross-examination.
Cross-Modal Coreference Alignment: Enabling Reliable Information Transfer in Omni-LLMs
Hongcheng Liu (Shanghai Jiao Tong University), Yu Wang (Shanghai Jiao Tong University)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVideoTextMultimodalityRetrieval-Augmented GenerationChain-of-ThoughtAudio
🎯 What it does: This study investigates the impact of cross-modal coreference alignment on Omni-LLM inference, and significantly enhances cross-modal reasoning capabilities by constructing the CROSSOMNI dataset and employing two training methods.
Cross-Modal Masked Compositional Concept Modeling for Enhancing Visio-Linguistic Compositionality
Wei Li (University of Science and Technology of China), Xinmei Tian (University of Science and Technology of China)
Explainability and InterpretabilityRepresentation LearningTransformerVision Language ModelAuto EncoderContrastive LearningImageTextMultimodality
🎯 What it does: Propose the MACCO framework, which enhances the compositional understanding ability of VLM by masking compositional concepts in one modality and reconstructing them using the complete context from another modality.
Cross-Modal Taxonomic Generalization in (Vision-) Language Models
Tianyang Xu (Toyota Technological Institute at Chicago), Kanishka Misra (University of Texas at Austin)
ClassificationExplainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: In vision-language models, this study investigates whether language models can infer the presence of superclasses (e.g., 'animal') through cross-modal inference using visual information, without explicit superclass labels during training.
Cross-Prompt Automated Essay Scoring of Multiple Traits: Making Sense of the State of the Art
Shengjie Li (University of Texas at Dallas), Vincent Ng (University of Texas at Dallas)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextReview/Survey Paper
🎯 What it does: Investigated the interactions among four key components in Cross-Prompt AES (feature sets, model architecture, feature normalization, and development set evaluation metrics), and for the first time explored the feasibility of improving cross-prompt scorers using transductive learning in classroom scenarios.
Crosscoding Through Time: Tracking Emergence & Consolidation Of Linguistic Representations Throughout LLM Pretraining
Deniz Bayazit (École Polytechnique Fédérale de Lausanne), Antoine Bosselut (École Polytechnique Fédérale de Lausanne)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelContrastive LearningTextBenchmark
🎯 What it does: This paper utilizes a sparse crosscoder to construct a shared feature space across checkpoints, and combines the relative indirect effect (RELIE) metric to track the emergence, persistence, and disappearance of language features during the pre-training of LLMs.
CrossGuard: Safeguarding MLLMs against Joint-Modal Implicit Malicious Attacks
Xu Zhang (City University of Hong Kong), Zhichao Lu (Washington University in St. Louis)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringVision Language ModelImageTextMultimodality
🎯 What it does: Proposed two modules, ImpForge and CrossGuard, for automatically generating implicit multimodal attack samples and building a security protection model capable of simultaneously defending against explicit and implicit attacks.
Crossing the Reward Bridge: Expanding Reinforcement Learning with Verifiable Rewards Across Diverse Domains
Yi Su (Soochow University), Dong Yu (Tencent)
OptimizationData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBenchmark
🎯 What it does: Propose an RLVR framework based on a generative discriminator, using soft probability rewards instead of traditional rule-based binary rewards, extending RLVR to multi-domain free-text reasoning tasks.
CSPO: Alleviating Reward Ambiguity for Structured Table-to-LaTeX Generation
Yunfan Yang (Beijing Jiaotong University), Yan Lu (Microsoft Research Asia)
GenerationOptimizationExplainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringVision Language ModelImageTextMultimodalityRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose a table image-to-LaTeX code generation framework called CSPO based on reinforcement learning, aiming to eliminate the reward ambiguity problem in traditional RL.
CSRP: Chain-of-Thought Reasoning for Chinese Text Correction via Reinforcement Learning with Efficiency-Aware Rewards
Wei Tian (East China Normal University), Man Lan (East China Normal University)
Computational EfficiencyData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningMixture of ExpertsTextChain-of-Thought
🎯 What it does: Propose a three-stage CSRP framework: first, obtain Chinese grammatical knowledge through Continuous Pre-training (CPT); then, enhance diagnostic thinking via Chain-of-Thought (CoT) Supervised Fine-tuning (SFT); finally, optimize the policy using Efficiency-Aware Reward (EAR) and Group Relative Policy Optimization (GRPO), significantly improving the performance of Chinese grammar error correction and spelling error correction.
CT-FineBench: A Diagnostic Fidelity Benchmark for Fine-Grained Evaluation of CT Report Generation
Ruifeng Yuan (DAMO Academy, Alibaba Group), Jianpeng Zhang (DAMO Academy, Alibaba Group)
GenerationTransformerLarge Language ModelPrompt EngineeringTextBiomedical DataComputed TomographyBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed and implemented a question-answering based evaluation benchmark called CT-FineBench for fine-grained assessment of factual consistency in CT report generation models.
CTRAP: Embedding Collapse Trap to Safeguard Large Language Models from Harmful Fine-Tuning
Biao Yi, Li Shen (Nankai University)
Safty and PrivacyExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
🎯 What it does: Proposes the CTRAP mechanism, embedding collapse traps during the alignment phase, causing LLMs to gradually lose their general language capabilities when subjected to harmful fine-tuning, thereby preventing the model from being maliciously used.
CTTA-T: Continual Test-Time Adaptation for Text Understanding via Teacher-Student with a Domain-aware and Generalized Teacher
Tianlun Liu, Dongsheng Li (National University Of Defense Technology)
Domain AdaptationKnowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextBenchmark
🎯 What it does: In text understanding tasks, a Continuous Test-Time Adaptation (CTTA) framework called CTTA‑T is proposed to online adapt in continuously changing target domains.
CUB: Benchmarking Context Utilisation Techniques for Language Models
Lovisa Hagström (Chalmers University of Technology), Isabelle Augenstein
Explainability and InterpretabilityComputational EfficiencyHyperparameter SearchData-Centric LearningTransformerSupervised Fine-TuningAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Developed the CUB benchmark to systematically evaluate the performance of context utilization techniques (CMT) in retrieval-augmented generation across different noisy contexts, models, and datasets.
CuBridge: An LLM-Based Framework for Understanding and Reconstructing High-Performance Attention Kernels
Xing Ma (Shanghai Jiao Tong University), Jin Song Dong (National University of Singapore)
OptimizationComputational EfficiencyAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: Proposes CuBridge—a framework that semantically adapts and restructures expert-written CUDA attention kernels by leveraging large language models and an executable intermediate representation called CuIR.
Culinary Crossroads: A RAG Framework for Enhancing Diversity in Cross-Cultural Recipe Adaptation
Tianyi Hu (Aarhus University), Daniel Hershcovich (University of Copenhagen)
GenerationDomain AdaptationRecommendation SystemTransformerPrompt EngineeringContrastive LearningTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes a retrieval-augmented generation (RAG) framework named CARRIAGE, which is used to generate recipes that are both culturally appropriate and diverse in cross-cultural recipe adaptation tasks.
Cultivating Forensic Reasoning for Generalizable Multimodal Manipulation Detection
Yuchen Zhang (Xi'an Jiaotong University), Zhedong Zheng (University of Macau)
Anomaly DetectionFederated LearningExplainability and InterpretabilityComputational EfficiencyRepresentation LearningReinforcement Learning from Human FeedbackTransformerReinforcement LearningPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodalityChain-of-Thought
🎯 What it does: Developed a reasoning-driven multimodal tampering detection framework called REFORM, and constructed a large-scale dataset named ROM with 704k samples, 9 types of tampering, and 5 news domains, annotated with reasoning.
Cultural Benchmarking of LLMs in Standard and Dialectal Arabic Dialogues
Muhammad Dehan Al Kautsar (Mohamed bin Zayed University of Artificial Intelligence), Fajri Koto (Mohamed bin Zayed University of Artificial Intelligence)
Large Language ModelTextBenchmark
🎯 What it does: This paper proposes the ArabCulture-Dialogue benchmark for Arabic cultural dialogue, covering MSA and local dialects across 13 Arab countries, and conducts three evaluations: multi-choice cultural reasoning, dialect-to-dialect translation, and dialect-driven generation.
Culture-Aware Machine Translation in Large Language Models: Benchmarking and Investigation
Zekun Yuan (Harbin Institute of Technology), Bing Qin (Huawei Technologies Co Ltd)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Constructed the cross-12-language-direction cultural-aware machine translation benchmark CanMT, and systematically evaluated the performance of large language models in culturally adaptive translation through a five-dimensional theoretical evaluation framework.
CuMA: Aligning LLMs with Sparse Cultural Values via Demographic-Aware Mixture of Adapters
Ao Sun (Southeast University), Shu Su (Southeast University)
Federated LearningExplainability and InterpretabilityComputational EfficiencyRepresentation LearningData-Centric LearningTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsContrastive LearningText
🎯 What it does: Proposes the CUMA framework, achieving multi-cultural value alignment through demographically aware Mixture-of-Adapters.
CURA: Clinical Uncertainty Risk Alignment for Language Model–Based Risk Prediction
Sizhe Wang (Washington University in St. Louis), Chenyang Lu (Washington University in St. Louis)
ClassificationExplainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningTextBiomedical DataElectronic Health Records
🎯 What it does: Propose a dual-layer uncertainty calibration framework called CURA for risk prediction based on clinical language models;
CURE-Med: Curriculum-Informed Reinforcement Learning for Multilingual Medical Reasoning
Eric Onyame (University of Virginia), Chirag Agarwal (University of Virginia)
Drug DiscoveryReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBiomedical DataBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed an end-to-end training framework called CURE-MED for multilingual medical reasoning, and created an open-source medical reasoning benchmark called CUREMED-BENCH containing 13 languages.
CURE: Critique-Driven Unified Reinforcement Learning for Test-Time Self-Improvement
Guirong Chen (Renmin University of China), Yankai Lin (Tencent Inc)
AI Code AssistantTransformerLarge Language ModelReinforcement LearningPrompt EngineeringMixture of ExpertsText
🎯 What it does: Propose the CURE framework, which jointly trains the problem-solving, critical thinking, and re-investigation capabilities of LLMs, enabling the model to self-verify and iteratively improve during reasoning.
Curing Miracle Steps in LLM Mathematical Reasoning with Rubric Rewards
Youliang Yuan (Chinese University of Hong Kong), Pinjia He (Chinese University of Hong Kong)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Studied and addressed the 'false positives' phenomenon in large language models during mathematical reasoning tasks, where the model provides the correct answer but the reasoning process is unreliable, and proposed a Rubric Reward Model based on problem-specific evaluation criteria to improve model reasoning quality through reinforcement learning.
Current Agents Fail to Leverage World Model as Tool for Foresight
Cheng Qian (University of Illinois Urbana Champaign), Heng Ji (Johns Hopkins University)
Autonomous DrivingExplainability and InterpretabilityRobotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIVision Language ModelWorld ModelImageVideoTextMultimodality
🎯 What it does: This paper constructs an evaluation framework to systematically assess the willingness, effectiveness, and errors of visual language model agents when optionally invoking world models, and explores their performance in agent decision-making and visual question answering tasks;
Cut Your Losses! Learning to Prune Paths Early for Efficient Parallel Reasoning
Jiaxi Bi (Chinese University of Hong Kong Shenzhen), Benyou Wang (Chinese University of Hong Kong Shenzhen)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose a path pruning method called STOP based on internal learnable signals to address the problem of excessively high parallel inference costs.
CxMP: A Linguistic Minimal-Pair Benchmark for Evaluating Constructional Understanding in Language Models
Miyu Oba (Nara Institute of Science and Technology), Saku Sugawara (National Institute of Informatics)
Large Language ModelTextBenchmark
🎯 What it does: Designed and constructed the CxMP benchmark to evaluate language models' understanding of constructed meaning.
CypherSmith: Transforming Text-to-Cypher Generation for LLMs with Synthetic Data
Zeyu Zhang (Amazon AGI), Thuy Vu (Amazon AGI)
Data SynthesisAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextGraphTabularBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose a method to generate a large-scale, high-quality, reasoning-text-containing text-to-Cypher query dataset called CypherSmith using open-source large language models (LLMs), addressing the issues of small scale, single domain, and low quality in existing datasets.
d-TreeRPO: Towards More Reliable Policy Optimization for Diffusion Language Models
Leyi Pan (Tsinghua University), Lijie Wen (Tsinghua University)
OptimizationExplainability and InterpretabilityKnowledge DistillationTransformerLarge Language ModelReinforcement LearningDiffusion modelTextChain-of-Thought
🎯 What it does: Proposes d-TreeRPO, a reliable reinforcement learning framework for diffusion-based large language models, which utilizes tree-structured rollout to achieve fine-grained verifiable rewards and improves prediction confidence through temporal self-distillation, thereby reducing probability estimation errors.
D^2Plan: Dual-Agent Dynamic Global Planning for Complex Retrieval-Augmented Reasoning
Kangcheng Luo (Peking University), Yansong Feng (Peking University)
Recommendation SystemOptimizationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AITextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes the D2PLAN dual-agent dynamic global planning paradigm, combining Reasoner and Purifier to achieve complex retrieval-enhanced reasoning, and designs a two-stage training approach (SFT cold start + SPLANRL) to enhance multi-hop reasoning capabilities.
DAC-Bench: A Decision-Aware Benchmark for Compositional Mobile GUI Tasks
Yuqing Zhang (Zhejiang University), Fei Wu (Zhejiang University)
Graph Neural NetworkTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextGraphSequentialBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the DAC-Bench benchmark, designing long-term sequential and decision-aware mobile GUI tasks;
Dark & Stormy: Modeling Humor in Sentences from the Bulwer-Lytton Fiction Contest
Venkata S Govindarajan (Ithaca College), Laura Biester (Middlebury College)
ClassificationRecognitionData SynthesisTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextBenchmark
🎯 What it does: This paper constructs and analyzes sentences from the Bulwer-Lytton Fiction Contest (BLFC), aiming to study 'deliberately bad humor,' an underestimated form of textual humor; meanwhile, it compares human-written sentences with those generated by large language models (LLMs), exploring their stylistic differences.
DARM: Distribution-Aware Reward Modeling by Alleviating Biases from Low Preference-Context Dependency Data
Shaofan Liu (Fudan University), Xuanjing Huang (Tsinghua University)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningContrastive LearningText
🎯 What it does: Proposes the DARM method, introducing conditional mutual information regularization in the reward model (RM) training to mitigate context-ignoring bias caused by low-preference-context relevance samples.
DART: Disambiguation-Aware Reasoning for Video-guided Machine Translation
Boyu Guan (Chinese Academy of Sciences), Chengqing Zong (University of Chinese Academy of Sciences)
Computational EfficiencyRepresentation LearningData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringVision Language ModelContrastive LearningVideoTextMultimodality
🎯 What it does: This paper proposes the DART framework, which uses reinforcement learning and the TVRF mechanism to dynamically determine whether video context is needed and perform corresponding reasoning in video-assisted machine translation;
DarwinTOD: LLM-Driven Lifelong Self-evolution for Task-oriented Dialog Systems
Shuyu Zhang (Shanghai Jiao Tong University), Bin Li (SIAT, CAS)
Autonomous DrivingOptimizationFederated LearningMeta LearningTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringMixture of ExpertsText
🎯 What it does: Proposed the DarwinTOD framework to achieve lifelong self-evolution in task-oriented dialogue systems.
Dashboard2Code: Evaluating Multimodal Models on Reconstructing Interactive Dashboards
Tianhao Niu (Harbin Institute of Technology), Wanxiang Che (Harbin Institute of Technology)
AI Code AssistantTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the Dashboard2Code task, which requires the model to actively explore interactive dashboards and generate code that reproduces their appearance and interactive logic, and construct the first DashboardMimic benchmark dataset based on Plotly+Dash;
Data Efficient RLVR via Off-Policy Influence Guidance
Erle Zhu (Tsinghua University), Hongning Wang (Tsinghua University)
Data-Centric LearningTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: This paper proposes a data selection framework for RLVR based on offline gradient estimation using influence functions and sparse random projection (CROPI), aimed at accelerating reinforcement learning in large language models.
Data Mixing Agent: Learning to Re-weight Domains for Continual Pre-training
Kailai Yang (University of Manchester), Mao Yang (Microsoft Research)
Domain AdaptationRepresentation LearningData-Centric LearningTransformerLarge Language ModelReinforcement LearningAgentic AIText
🎯 What it does: Proposes the Data Mixing Agent, which uses reinforcement learning to automatically weight the source and target domains, thereby maintaining the source domain's capabilities and improving the target domain's performance during continual pre-training.
Data Pollination: An Emergent Ecological Process Driving AI Population Evolution
Shufang Xie (Renmin University of China), Rui Yan (Wuhan University)
OptimizationFederated LearningData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningGaussian SplattingTextBenchmark
🎯 What it does: This paper studies the phenomenon where deployed language models propagate and form an evolutionary ecology through generated synthetic data in global networks, proposes the concept of data pollination, and constructs a population-based evolutionary framework to verify the role of ecological diversity in model robustness.
Data-efficient Targeted Token-level Preference Optimization for LLM-based Text-to-Speech
Rikuto Kotoge (SpiralAI Inc.), Yuichi Sasaki (SpiralAI Inc.)
GenerationData SynthesisOptimizationTransformerLarge Language ModelReinforcement LearningPrompt EngineeringScore-based ModelFlow-based ModelContrastive LearningTextAudio
🎯 What it does: This paper proposes a Token-level Kahneman-Tversky Optimization (TKTO) method for preference optimization in LLM-based TTS, eliminating the need for paired data and directly training at the token level.
De-Anonymization at Scale via Tournament-Style Attribution
Lirui Zhang (Beihang University), Huishuai Zhang (Peking University)
Safty and PrivacyTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Propose a two-stage method for author de-anonymization using large language models on a large-scale anonymous text collection — first using rough retrieval to narrow down the candidate set, then performing elimination-style matching based on LLMs, finally outputting the top-k candidate texts.
DE-CLIP: Few-Shot Anomaly Detection via Difference-Guided Embedding Editing
Yage Zhang (CISPA Helmholtz Center for Information Security), Yang Zhang (CISPA Helmholtz Center for Information Security)
Anomaly DetectionPrompt EngineeringVision Language ModelContrastive LearningImageTextBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Propose a training-free, CLIP-based few-shot anomaly detection and segmentation framework called DE-CLIP, which enhances the semantic separation between positive and negative samples by directly rewriting text embeddings through differential guided embedding editing (DiffEdit).
DEBAR: Mitigating Contextual Bias in Cross-Document Relation Extraction via Dual-Stream Decoupling
Zhixuan Yang (Northeastern University), Jingwei Cheng (Northeastern University)
Graph Neural NetworkTransformerContrastive LearningText
🎯 What it does: Propose the DEBAR framework, which uses dual-stream graph reasoning and bridge-aware input construction to alleviate relation transfer bias in cross-document relation extraction, and introduces a curriculum learning-based ranking optimization loss to enhance discriminativeness.
Debate-of-Thoughts: Resolving Knowledge Conflicts in LLMs Through Internal Deliberation
Guocong Li, Hongxia Xu (State Key Laboratory Of Transvascular Implantation Devices And Tidri)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed and implemented a framework called Debate-of-Thoughts (DoT), which uses internal debate to resolve knowledge conflicts in Retrieval-Augmented Generation (RAG), avoiding simple 'select or suppress' strategies.
Debiased Orthogonal Boundary-Driven Efficient Noise Mitigation
Hao Li (Washington University in St. Louis), Lianli Gao (University of Electronic Science and Technology of China)
ClassificationRetrievalData-Centric LearningTransformerVision Language ModelScore-based ModelContrastive LearningImageTextMultimodality
🎯 What it does: Propose the One-Step Anti-Noise (OSA) method, which identifies and mitigates noisy labels in a single forward pass by leveraging the high-dimensional orthogonality of pre-trained multi-modal models.
Debiasing Reward Models via Causally Motivated Inference-Time Intervention
Kazutoshi Shinoda (Human Informatics Labs., NTT, Inc.), Kyosuke Nishida (Human Informatics Labs., NTT, Inc.)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmark
🎯 What it does: Intervene in the reward model with causal activation during inference to eliminate various format-related biases.
Decentralized Arena: Towards Democratic and Scalable Automatic Evaluation of Language Models
Yanbin Yin (University of California, San Diego), Zhiting Hu (University of California, San Diego)
Federated LearningExplainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose Decentralized Arena (DE-ARENA), using all LLMs to be evaluated as judges, constructing a fully automated and democratic evaluation framework.
DecIF: Improving Instruction-Following through Decomposition
Tingfeng Hui (Beijing University of Posts and Telecommunications), Sen Su (Beijing University of Posts and Telecommunications)
Data SynthesisReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Propose DecIF, a decomposition-based automated data synthesis framework that can generate verifiable, controllable, and diverse instruction-following data from scratch for supervised fine-tuning (SFT) and reinforcement learning (RL) training.
Decisive: Guiding User Decisions with Optimal Preference Elicitation from Unstructured Documents
Akriti Jain (Adobe Research), Aparna Garimella (Adobe Research)
Recommendation SystemOptimizationFederated LearningExplainability and InterpretabilityData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelTextFinance RelatedChain-of-Thought
🎯 What it does: DECISIVE constructs an interactive decision support framework by extracting option scoring matrices from unstructured documents and combining them with Bayesian preference reasoning.
DecoCal: Decoding with Calibration in Diffusion Large Language Models
Fan Xu (Peking University), Xiaojun Wan (Peking University)
Explainability and InterpretabilityComputational EfficiencyAI Code AssistantTransformerLarge Language ModelPrompt EngineeringDiffusion modelContrastive LearningText
🎯 What it does: Proposes the DecoCal decoding framework, which utilizes the historical predictions across diffusion steps through KL-weighted aggregation for token confidence calibration, and determines when to decode or re-mask based on the calibrated confidence, thereby achieving more reliable early decoding and error correction in DLLMs.