ACL 2026 Papers — Page 6
Annual Meeting of the Association for Computational Linguistics · 2296 papers
Decoding Scientific Experimental Images: The SPUR Benchmark for Perception, Understanding, and Reasoning
Junpeng Ding (Beijing University of Posts and Telecommunications), Siying Lin (Beijing University of Posts and Telecommunications)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Constructed the SPUR benchmark for scientific experiment images to evaluate the capabilities of multimodal large language models in image perception, cross-panel understanding, and expert-level reasoning.
Decoding-Unlearning: Fact Forgetting via Entropy-Guided Inference
Jingwen Pu (University of Electronic Science and Technology of China), Kun She (University of Electronic Science and Technology of China)
Safty and PrivacyExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Proposes a learning-free framework called SEGUE based on inference, which uses a single-channel detector (Single2Dual Probe) to determine whether a query involves forgettable concepts, and locates factual units by utilizing target token entropy during the decoding process. A hybrid decoding strategy combining hard masking and entropy-guided smooth decay is adopted to achieve learning-free forgetting of sensitive knowledge.
Decoupling Generalization and Adaptation in Meta-Learning for Large Language Models
Nitin Vetcha (National University of Singapore), Dianbo Liu (National University of Singapore)
Meta LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: Proposed a meta-learning framework called DeGAML-LLM, which separates the generalization of large language models and task-specific adaptation into two independent modules: a task-conditioned parameter generator (Gφ) responsible for learning cross-task general representations, and a reinforcement learning-based adaptation policy (πψ) responsible for closed-loop iterative optimization on each new task.
Decoupling Task-Solving and Output Formatting in LLM Generation
Haikang Deng (University of California, Los Angeles), Nanyun Peng (University of California, Los Angeles)
GenerationKnowledge DistillationData-Centric LearningTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmark
🎯 What it does: Propose a decoupled decoding framework called DECO-G that separates task solving from output formatting, allowing LLMs to focus on reasoning without being distracted by formatting constraints.
Deep Kernel Fusion for Transformers
Zixi Zhang (Imperial College London), Robert D. Mullins (University of Cambridge)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Proposed and implemented DeepFusionKernel—a deeply fused CUDA kernel specifically designed for the SwiGLU MLP block in Transformers, eliminating intermediate activation buffers, reducing memory transfer, and integrating it into the SGLang inference framework, with a lightweight scheduler to dynamically select the optimal kernel.
Deep Research with Open-Domain Evaluation and Multi-Stage Guardrails for Safety
Wei-Chieh Huang (University of Illinois Chicago), Philip S. Yu (University of Illinois Chicago)
Safty and PrivacyTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes the DEEPRESEARCHGUARD framework, providing multi-stage security protection and open-domain evaluation for four stages of deep research (open-domain research): input, planning, research, and output; simultaneously constructs the DRSAFEBENCH benchmark for comprehensive testing.
Deep Supervised Contrastive Learning of Pitch Contours for Robust Pitch Accent Classification in Seoul Korean
Hyunjung Joo (Rutgers University), GyeongTaek Lee (Gachon University)
ClassificationConvolutional Neural NetworkRecurrent Neural NetworkTransformerSupervised Fine-TuningContrastive LearningAudio
🎯 What it does: Propose Dual-Glob, which uses a dual-perspective supervised contrastive learning framework for fine-grained pitch accent classification of F0 contours in Seoul Korean.
Deep-Reporter: Deep Research for Grounded Multimodal Long-Form Generation
Fangda Ye (National University Of Singapore), Shuicheng Yan (National University Of Singapore)
GenerationData SynthesisRetrievalTransformerLarge Language ModelSupervised Fine-TuningAgentic AIPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed the DEEP-REPORTER framework to achieve automated generation of multi-modal long-text reports;
DeepFact: Co-Evolving Benchmarks and Agents for Deep Research Factuality
Yukun Huang (Duke University), Venkatesh Saligrama (Amazon AGI)
Explainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Designed an evolvable evaluation framework called DeepFact, which uses the Audit-then-Score mechanism to verify the truthfulness of claims in deep research reports (DRR), and released the corresponding benchmark and evaluator DeepFact-Eval.
DeepGuard: Secure Code Generation via Multi-Layer Semantic Aggregation
Li Huang (Chongqing University), Meng Yan (Chongqing University)
Safty and PrivacyAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringContrastive LearningTextSequential
🎯 What it does: Proposes the DEEPGUARD framework, which utilizes multi-layer attention to aggregate mid-to-high-level hidden states of the Transformer, enhancing the security of code generation models while maintaining functional correctness.
DeepPlanning: Benchmarking Long-Horizon Agentic Planning with Verifiable Constraints
Yinger Zhang (Alibaba Group), Junyang Lin (Alibaba Group)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Created the DEEPPLANNING benchmark, using multi-day travel and multi-product shopping tasks to evaluate the ability of LLM agents in proactive information acquisition, local constraint reasoning, and global constraint optimization.
Defense Against Knowledge Poisoning Attack on GraphRAG
Havva Alizadeh Noughabi (University of Guelph), Ali Dehghantanha (University of Guelph)
RetrievalAnomaly DetectionExplainability and InterpretabilityComputational EfficiencyAdversarial AttackGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextGraphRetrieval-Augmented Generation
🎯 What it does: Propose and implement HoGGRAG, a defensive layer located between the GraphRAG retriever and generator, which can detect and repair structural inconsistencies caused by knowledge poisoning after subgraph retrieval, thereby improving the accuracy of answers in multi-hop question answering.
Defenses Against Prompt Attacks Learn Surface Heuristics
Li Li, Yue Zhao (Washington University in St. Louis)
Safty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Systematically evaluate prompt-injection defense methods based on supervised fine-tuning and prompt guards, revealing their 'shortcut' behaviors beyond security tasks.
DefGen-Bench: A Benchmark for Chinese Criminal Defence Opinion Generation in LegalAI
Senbo Zhang (Nankai University), Qian Liu (University of Auckland)
GenerationTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the task of generating criminal defense opinions and the benchmark DefGen-Bench, and study how to enable large language models to generate structured and legally compliant defense opinions from a dual perspective (indictment and defendant's statement).
DEFT: Demystifying VLN Failures via a Unified Dual-View Explainability Framework for LLM-based Agents
Yawen Wang (Chinese Academy of Sciences), Qing Wang (Chinese Academy of Sciences)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelMixture of ExpertsVision Language ModelVision-Language-Action ModelImageTextMultimodality
🎯 What it does: Propose the DEFT framework for joint temporal and visual interpretation of LLM-driven vision-language navigation (VLN) agents;
Deliberative Searcher: Improving LLM Reliability via Reinforcement Learning with Constraints
Zhenyun Yin (Fudan University), Yingchun Wang (Shanghai Artificial Intelligence Laboratory)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelReinforcement LearningTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the Deliberative Searcher framework, combining chain-of-thought reasoning, retrieval, and self-confidence assessment, and achieving dual optimization of reliability and accuracy through constrained reinforcement learning.
Demystifying Data Organization for Enhanced LLM Training
Yalun Dai (Nanyang Technological University), Scarlett Li (Microsoft Research)
Data-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper systematically explores the impact of data organization on LLM training, proposing four guiding principles and implementing two sorting strategies, STR and SAW, based on precomputed sample-level scores;
Demystifying Uncertainty in LLMs: Active Calibration between Concepts and Human Evaluations
Pengqi Li (Beijing Institute of Technology), Guoren Wang (Beijing Institute of Technology)
Explainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelAgentic AIPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: This paper reveals that calibration error in large language models (LLMs) cannot disappear under non-interactive settings, through theoretical proof and experimental validation. It then proposes an interactive learning strategy (ILS) based on calibration error, which significantly reduces hallucinations and uncertainty in LLMs by actively asking humans questions and using calibration error to guide queries.
DeReA: Improving Idiom Translation with Detect-Retrieve-Arbitrate Reasoning
Rongqing Jiang (Harbin Institute of Technology), Min Zhang (Harbin Institute of Technology)
GenerationRetrievalTransformerLarge Language ModelReinforcement LearningTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed a three-stage (Detect‑Retrieve‑Arbitrate) framework named DEREA, simulating the human translation process. It first uses a preference-aligned detector to identify idioms in sentences, then employs a refined retriever to obtain corresponding explanations from a knowledge base, and finally selects the optimal result between direct translation and retrieval-enhanced translation through a dual-path arbitration mechanism. Based on this, a new high-difficulty idiom translation benchmark, LoMI, and an Emerging Slang dataset were constructed to verify the model's performance on novel and rapidly evolving idioms.
Deriving Character Logic from Storyline as Codified Decision Trees
Letian Peng (University of California, San Diego), Jingbo Shang (University of California, San Diego)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes the Codified Decision Tree (CDT) framework, which automatically generates executable, context-aware character behavior decision trees from large-scale narrative text, aiming to enhance the consistency and interpretability of role-playing agents.
Detecting RAG Extraction Attack via Dual-Path Runtime Integrity Game
Yuanbo Xie (Institute of Information Engineering, Chinese Academy of Sciences), Tingwen Liu (Institute of Information Engineering, Chinese Academy of Sciences)
RetrievalSafty and PrivacyTransformerPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: To address knowledge base leakage attacks in Retrieval-Augmented Generation (RAG) systems, the CanaryRAG runtime defense framework is proposed. It embeds special semantically meaningless canary markers into retrieval results and simultaneously monitors the presence or absence of these canaries on both the target generation path and the Oracle detection path to detect and intercept leakage behaviors in real time.
Detecting What Queries Seek: Steering LLM Safety with FFN Output Activation Monitoring
Xiaohao Luo (Northeastern University), Rui Zhao (Northeastern University)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelText
🎯 What it does: Propose a conditional activation regulation method called FGAS based on the output activation of FFN, used to detect and defend against adversarial breaking attacks during LLM inference.
DetectRL-X: Towards Reliable Multilingual and Real-World LLM-Generated Text Detection
Junchao Wu (University of Macau), Derek F. Wong (University of Macau)
ClassificationAnomaly DetectionTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextMultimodalityBenchmark
🎯 What it does: Constructed the DetectRL-X multilingual LGT detection benchmark, covering 8 languages, 6 domains, 4 generators, 8 attacks, 4 text lengths, 3 refinement operations, and proposed binary and ternary classification tasks.
Detoxification for LLM: From Dataset Itself
Wei Shao (State Key Laboratory of AI Safety), Xueqi Cheng (State Key Laboratory of AI Safety)
Safty and PrivacyTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
🎯 What it does: This study proposes a method for removing toxicity directly from the original corpus of large language models during the pre-training phase, named the HSPD pipeline;
DFAMS: Dynamic-flow guided Federated Alignment based Multi-prototype Search
Zhibang Yang (Peking University), Yasha Wang (Peking University)
RetrievalFederated LearningExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringFlow-based ModelContrastive LearningTextBiomedical DataRetrieval-Augmented Generation
🎯 What it does: Propose a federated retrieval framework called DFAMS based on Dynamic-Information-Flow (DIF), which can identify query intent and build semantically aligned knowledge partitions in distributed knowledge bases through DIF;
DIA-HARM: Dialectal Disparities in Harmful Content Detection Across 50 English Dialects
Jason S Lucas, Dongwon Lee (Pennsylvania State University)
ClassificationData-Centric LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextBenchmark
🎯 What it does: Constructed the DIA-HARM benchmark, which includes 195K pieces of misleading content across 50 English dialects (D-CUBE), as well as validation and evaluation tools (D-PURIFY)
Diagnosing and Remedying Representation Deficiencies for Deterministic Reasoning in KGQA
Gewen Liang (Harbin Institute of Technology), Min Zhang (Harbin Institute of Technology)
Explainability and InterpretabilityKnowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringContrastive LearningTextGraph
🎯 What it does: Proposed the DR2 method to diagnose and repair the non-deterministic reasoning problems caused by insufficient semantic representation in large language models during knowledge graph question answering (KGQA) tasks.
Diagnosing Hidden Instabilities in Model Editing via Uncertainty Quantification
Zihan Gu (Institute of Information Engineering Chinese Academy of Sciences), Yue Hu (Institute of Information Engineering Chinese Academy of Sciences)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Explore the robustness of single-shot model editing and propose a method for uncertainty diagnosis based on structured semantic perturbation and SPUQ
Diagnosing Spatial Consistency across Perspectives and Viewpoints in Large Vision-Language Models
Yoonji Kim (Yonsei University), Sung-Bae Cho (Yonsei University)
Explainability and InterpretabilityTransformerPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes the SCOPE benchmark to systematically evaluate the consistency, integration, and updating capabilities of large vision-language models in different perspectives regarding three-dimensional spatial relationships.
DialDefer: A Framework for Detecting and Mitigating LLM Dialogic Deference
Parisa Rabbani (University of Illinois Urbana-Champaign), Dilek Hakkani-Tür (University of Illinois Urbana-Champaign)
Federated LearningExplainability and InterpretabilityComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringText
🎯 What it does: Studied the 'dialogic deference' phenomenon in third-party judgments by large language models, and proposed a detection and mitigation framework called DIALDEFER;
Dialogue is the Plan: From Interface to Joint Action in Agentic AI
Mert Inan (Northeastern University), Anthony Sicilia (West Virginia University)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningAgentic AITextReview/Survey Paper
🎯 What it does: Propose viewing the dialogue of large language model agents as a planning process of joint actions, rather than merely an instruction interface, highlighting the central role of dialogue in shared goals, common ground, clarification, and repair.
Diff4TST: Masked Diffusion Language Model for Text Style Transfer
Xinchen Ma (East China Normal University), Weining Qian (East China Normal University)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningDiffusion modelText
🎯 What it does: Propose Diff4TST, a text style transfer framework based on masked diffusion language models, which directly transforms source text into target styles step by step;
Difference in Task Performance on Sparse Speech Representations
Wenjie Peng (University of Sheffield), Thomas Hain (University of Sheffield)
Computational EfficiencyRepresentation LearningTransformerAuto EncoderContrastive LearningAudio
🎯 What it does: Investigated the performance differences of sparse speech representations across six SUPERB tasks and explained the sparsity preferences of different tasks using the information bottleneck framework.
Different types of syntactic agreement recruit the same units within large language models
Daria Kryvosheieva (Massachusetts Institute of Technology), Greta Tuckute (Kempner Institute at Harvard University)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Localizing functional units within large language models (LLMs) to investigate whether different syntactic phenomena share the same set of units, with a focus on the shared usage of three types of agreement (subject-verb, pronoun, and determiner-noun); simultaneously examining whether agreement units across languages (English, Russian, Chinese) overlap and how this relates to linguistic similarity.
Difficulty-Controllable Cloze Question Distractor Generation
Seokhoon Kang (POSTECH), Gary Lee (POSTECH)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Propose a controllable difficulty multiple-choice fill-in-the-blank question distractor generation framework, which constructs annotated data by utilizing bidirectional candidate generation, filtering, and difficulty clustering, and trains the model using multi-task learning on this data.
Diffuse Thinking: Exploring Diffusion Language Models as Efficient Thought Proposers for Reasoning
Chenyang Shao (Tsinghua University), Yong Li (Tsinghua University)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringDiffusion modelTextBenchmarkChain-of-Thought
🎯 What it does: Proposes the Diffuse Thinking framework, which generates multiple ideas through diffusion language models (DLM) and evaluates and selects them using autoregressive large language models (LLM) to achieve efficient reasoning.
Diffusion-CAM: Faithful Visual Explanations for dMLLMs
Haomin Zuo (Sun Yat-sen University), Xiaofeng Zhang (Northwestern University)
Explainability and InterpretabilityTransformerVision Language ModelDiffusion modelContrastive LearningImageTextMultimodality
🎯 What it does: Propose Diffusion-CAM, a gradient visualization explanation framework specifically designed for diffusion-based multimodal large language models, which can accurately locate the regions of interest during the global parallel generation process.
Digitizing Nepal’s Written Heritage: A Comprehensive HTR Pipeline for Old Nepali Manuscripts
Anjali Sarawgi (LMU Munich), Christof Zotter (Heidelberg Academy of Sciences and Humanities)
RecognitionData SynthesisTransformerSupervised Fine-TuningVision Language ModelContrastive LearningImageText
🎯 What it does: Propose an end-to-end pipeline for Old Nepali Handwritten Text Recognition (HTR), achieving a complete workflow from image preprocessing, line-level segmentation, three-stage transfer learning, model training, decoding to evaluation.
DimABSA: Building Multilingual and Multidomain Datasets for Dimensional Aspect-Based Sentiment Analysis
Lung-Hao Lee (National Yang Ming Chiao Tung University), Saif M. Mohammad (National Research Council Canada)
Data-Centric LearningLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmarkFinance Related
🎯 What it does: Proposed the DimABSA multilingual multi-domain dimensional sentiment analysis dataset, and designed three subtasks and a unified evaluation metric cF1.
DiningBench: A Hierarchical Multi-view Benchmark for Perception and Reasoning in the Dietary Domain
Song Jin (Renmin University of China), Rui Yan (Wuhan University)
ClassificationRecognitionRecommendation SystemTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: This paper proposes DiningBench, a multi-modal benchmark in the food domain that includes multi-perspective, multi-task data (fine-grained classification, nutritional estimation, visual question answering).
DiNO: Disinformation Narrative Observer
Witold Sosnowski, Adam Wierzbicki (Polish-Japanese Academy Of Information Technology)
RecognitionData-Centric LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Developed and evaluated the DiNO method, which can extract and summarize misinformation narratives from news articles.
DisCal: Distribution-Aware Calibration for Mathematical Reasoning Under Character-Level Noisy Inputs
Bo Zhang (PLA Rocket Force University of Engineering), Zhong Wang (PLA Rocket Force University of Engineering)
Data-Centric LearningTransformerSupervised Fine-TuningPrompt EngineeringContrastive LearningTextBenchmarkChain-of-Thought
🎯 What it does: Propose DisCal, a distribution-aware confidence calibration framework designed for character-level noise, applied to mathematical reasoning tasks.
DisCo_Speech: Controllable Zero-Shot Speech Generation with A Disentangled Speech Codec
Tao Li (China Mobile Nineverse Artificial Intelligence Technology Co Ltd), Junlan Feng (China Mobile Nineverse Artificial Intelligence Technology Co Ltd)
GenerationCompressionTransformerLarge Language ModelDiffusion modelAuto EncoderGenerative Adversarial NetworkContrastive LearningAudio
🎯 What it does: Propose the DisCo-Speech framework to achieve zero-shot controllable speech generation; DisCodec is used to decouple speech into three factors: content, prosody, and timbre, and a language model is utilized to continue the prosody. After decoupling, the decoder injects the target timbre.
Disco-RAG: Discourse-Aware Retrieval-Augmented Generation
Dongqi Liu (Saarland University), Yabiao Wang (Zhejiang University)
GenerationRetrievalExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Designed and implemented Disco-RAG, a system that introduces discourse structure (RST tree, discourse graph) and generation planning into the retrieval-augmented generation (RAG) framework, to improve the accuracy and coherence of knowledge-intensive question answering and long document summarization.
Discourse Coherence and Response-Guided Context Rewriting for Multi-Party Dialogue Generation
Zhiyu Cao (Soochow University), Qiaoming Zhu (Soochow University)
GenerationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Proposes the DRCR framework, which rewrites multi-party dialogue contexts through dual feedback on discourse coherence and response quality, thereby improving the quality of multi-party dialogue generation.
Discourse Realization of Generics in Human and LLM-generated Texts
Søren Kirkegaard Fomsgaard (University of Caen), Nelleke Oostdijk (Radboud University)
ClassificationExplainability and InterpretabilityRecurrent Neural NetworkTransformerLarge Language ModelPrompt EngineeringTextReview/Survey PaperRetrieval-Augmented Generation
🎯 What it does: This paper constructs sentence-level generalization annotations and designs a text-level generalization score metric to compare the generalization usage of humans and large language models in argumentative texts.
Discover and Prove: An Open-source Agentic Framework for Hard Mode Automated Theorem Proving in Lean 4
Chengwu Liu (Peking University), Ming Zhang (Peking University)
OptimizationExplainability and InterpretabilityComputational EfficiencyAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper designs the Hard Mode ATP environment, releases two re-annotated datasets, MiniF2F-Hard and FIMO-Hard, and proposes the Discover and Prove (DAP) framework, which first uses an LLM to discover answers and then completes formal proofs using an automatic theorem prover.
Discovering a Shared Logical Subspace: Steering LLM Logical Reasoning via Alignment of Natural-Language and Symbolic Views
Feihao Fang (University of Illinois Urbana-Champaign), Yuanyuan Lei (University of Florida)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextBenchmarkChain-of-Thought
🎯 What it does: Discover and leverage the shared logical subspaces inherent in large language models by aligning residual activations of natural language and symbolic reasoning chains, achieving untrained reasoning chain derivation intervention to enhance multi-step logical reasoning performance.
Discovering and Causally Validating Emotion-Sensitive Neurons in Large Audio-Language Models
Xiutian Zhao (Johns Hopkins University), Berrak Sisman (Johns Hopkins University)
RecognitionExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextMultimodalityAudio
🎯 What it does: Investigate the existence and causal role of emotion-sensitive neurons (ESN) in large audio-language models (LALMs), and perform neuron-level interpretation and emotion control through activation logging, selector screening, and intervention during inference (deactivation, amplification).
Discovering Properties of Inflectional Morphology in Neural Emergent Communication
Miles Gilberti (University of Michigan), Huteng Dai (University of Michigan)
GenerationReinforcement Learning from Human FeedbackRecurrent Neural NetworkReinforcement LearningContrastive LearningTextReview/Survey Paper
🎯 What it does: Propose an 'Emergent Morphology' framework based on an attribute-value reconstruction game, emphasizing dual structuring (small vocabulary size) and mapping the game to a morphological 'root + inflection' structure.
Discovery and Reinforcement of Tool-Integrated Reasoning Chains via Rollout Trees
Kun Li (The Chinese University of Hong Kong), Bo Zhou (Tencent)
OptimizationAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringDiffusion modelTextSequentialBenchmarkChain-of-Thought
🎯 What it does: Train large reasoning models to spontaneously invoke tools in long Chain-of-Thought reasoning, using reinforcement learning to build a dynamic backtracking tree to discover and reinforce beneficial tool usage paths.
Disentangling Reasoning Logic to Resolve Explicit Knowledge Conflicts
Xianda Zheng (University of Auckland), Kaiqi Zhao (Harbin Institute of Technology)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationReinforcement Learning from Human FeedbackGraph Neural NetworkTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextGraphRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed the KCR framework, which resolves explicit knowledge conflicts by structuring and separating reasoning chains in conflicting contexts.
Dissecting Failure Dynamics in Large Language Model Reasoning
Wei Zhu (Yunnan University), Zhiwen Tang (Yunnan University)
Explainability and InterpretabilityComputational EfficiencyAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: A fine-grained analysis of error trajectories in large language models during the reasoning process reveals that errors often emerge early and are marked by local entropy peaks, which can be corrected through local branches.
Distillation Traps and Guards: A Calibration Knob for LLM Distillability
Weixiao Zhan (Nanyang Technological University), Dacheng Tao (AGH University of Krakow)
Knowledge DistillationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: Investigate the fundamental reasons for the failure of knowledge distillation and propose a teacher calibration method based on reinforcement learning, which can control the distillability of the teacher through a coefficient;
Distilling Large Embeddings via Hyperspherical Householder Quantization
Yihang Wang (State Key Laboratory of AI Safety), Xueqi Cheng (State Key Laboratory of AI Safety)
RetrievalCompressionKnowledge DistillationRepresentation LearningTransformerSupervised Fine-TuningAuto EncoderContrastive LearningText
🎯 What it does: By iteratively applying Householder reflections on the unit sphere, high-dimensional contrastive learning embeddings are compressed into short discrete identifiers for generative retrieval.
Distributional Clarity: The Hidden Driver of RL-Friendliness in Large Language Models
Shaoning Sun (Tsinghua University), Haifeng Wang (Baidu Inc)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Investigated and validated distributed clarity as a core structural feature for the trainability (RL-Friendliness) of large language models (LLMs) in reinforcement learning (RL);
DiVE: Decoupling Intra-layer Visual Evidence for Mitigating Hallucinations in Large Vision-Language Models
Xinwei Li (Southeast University), Hanqian Wu (Southeast University)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: DiVE proposes a training-agnostic decoding framework that alleviates visual hallucinations by decoupling visual information within a single forward pass.
Diversity in Unity, Theory in Practice: Hierarchical Multitask Benchmarks for Chinese Minority Languages
Yijie Li (Minzu University of China), Jia Qing Cai Wang
Large Language ModelSupervised Fine-TuningPrompt EngineeringTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Constructed CMiLBench, a hierarchical multi-task benchmark for Tibetan, Mongolian, and Uyghur languages (with 24,663 instances, 17 tasks, and 5 difficulty levels), and used this benchmark to systematically evaluate 14 mainstream LLMs (including commercial and open-source models).
Diving into the Decoding Space of Non-Autoregressive Models via Lexically Constrained Search
Chenyang Huang (University of Alberta), Osmar Zaiane
GenerationData-Centric LearningTransformerSupervised Fine-TuningText
🎯 What it does: Studied the decoding space of non-autoregressive models (DAT) and proposed a search-based decoding algorithm based on LexMAP to achieve lexical constrained machine translation.
DIXITWORLD: Evaluating Multimodal Abductive Reasoning in Vision-Language Models with Multi-Agent Dixit Gameplay
Yunxiang MO, Yangqiu Song (HKUST)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Designed the DixitWorld evaluation suite, which includes the multi-agent dynamic game DixitArena and the static multi-difficulty QA benchmark DixitBench, for systematic evaluation of the generative and selective capabilities of vision-language models in multi-modal inductive reasoning.
DiZiNER: Disagreement-guided Instruction Refinement via Simulating Pilot Annotation for Zero-shot Named Entity Recognition
Siun Kim (Seltasquare), Hyung-Jin Yoon (Seoul National University)
RecognitionTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed the DiZiNER framework, which simulates the human pilot annotation process, utilizing diverse LLMs as annotators and supervisors to achieve controversy-guided instruction refinement for zero-shot named entity recognition (NER);
DMHM: Density-aware Manifold Learning and Hybrid Mahalanobis Energy for LLMs-generated Text Detection
Tianle Liu (National University of Defense Technology), Dongsheng Li (National University of Defense Technology)
Anomaly DetectionTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Propose a detection framework for LLM-generated text (DMHM) based on density-aware manifold learning and hybrid Mahalanobis distance energy, which constructs a manifold with more distinct density differentiation in the embedding space and clearly separates generated text (LGT) from human-written text (HWT) through energy learning and contrastive learning.
DMN: A Compositional Framework for Jailbreaking Multimodal LLMs with Multi-Image Inputs
Wenzhuo Xu (360 AI Security Lab), Quanchen Zou (360 AI Security Lab)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelPrompt EngineeringVision Language ModelDiffusion modelImageTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the DMN framework, using multi-graph input to achieve anti-piercing of multi-modal large language models (MLLM);
DMRetriever: A Family of Models for Improved Text Retrieval in Disaster Management
Kai Yin (Texas A&M University), James Caverlee (Texas A&M University)
RetrievalTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextBenchmark
🎯 What it does: This paper proposes a series of dense retrieval models specifically designed for disaster management (33M–7.6B) along with a three-stage training framework to enhance retrieval performance in disaster scenarios.
Do LLM Agents Mirror Socio-Cognitive Effects in Power-Asymmetric Conversations?
Anvesh Rao Vijjini (UNC Chapel Hill), Snigdha Chaturvedi (UNC Chapel Hill)
Safty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelAgentic AIPrompt EngineeringText
🎯 What it does: This paper simulates multi-round power-asymmetric dialogues by assigning different roles and personalities to large language models, systematically evaluating their performance in four social cognitive effects: pronoun usage, linguistic coordination, persuasiveness, and harmful compliance.
Do LLMs Capture Embodied Cognition and Cultural Variation? Cross-Linguistic Evidence from Demonstratives
Yu Wang (Hong Kong Polytechnic University), Chu-Ren Huang (Hong Kong Polytechnic University)
RecognitionExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: This paper constructs a bilingual (English, Chinese) demonstrative word (e.g., this/that, this/that) question-answer dataset to investigate whether large language models (LLMs) can capture embodied cognition and cultural differences, and evaluates them by comparing with 6,400 answers from 320 native participants.
Do LLMs Encode Functional Importance of Reasoning Tokens ?
Janvijay Singh (University of Illinois Urbana Champaign), Dilek Hakkani-Tür (University of Illinois Urbana Champaign)
Explainability and InterpretabilityKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This study investigates whether large language models (LLMs) encode token-level functional importance within reasoning chains, and proposes likelihood-based greedy pruning as a diagnostic tool to compress reasoning chains and identify important tokens.
Do LLMs Know Tool Irrelevance? Demystifying Structural Alignment Bias in Tool Invocations
Yilong Liu (Institute of Information Engineering, Chinese Academy of Sciences), Yanan Cao (Institute of Information Engineering, Chinese Academy of Sciences)
Explainability and InterpretabilityTransformerLarge Language ModelContrastive LearningTextBenchmark
🎯 What it does: Studied the structural alignment bias (Structural Alignment Bias) that occurs in large language models when using external tools, constructed the SABEval dataset capable of decoupling structural alignment from semantic relevance, and conducted quantitative and mechanistic analysis of the model's erroneous invocation behavior;
Do LLMs Really Memorize Personally Identifiable Information? Revisiting PII Leakage with a Cue-Controlled Memorization Framework
Xiaoyu Luo (Aalborg University), Johannes Bjerva (Aalborg University)
Safty and PrivacyTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Reevaluate PII leakage in multilingual large-scale LLMs, proposing the Cue-Resistant Memorization (CRM) framework, which distinguishes true memorization from surface-cue-driven reconstruction by controlling the stem overlap between prompts and target PII.
Do LLMs Really Need 10+ Thoughts for “Find the Time 1000 Days Later”? Towards Structural Understanding of LLM Overthinking
Xinliang Frederick Zhang (University of Michigan), Shyam Upadhyay (Google DeepMind)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This study investigates the phenomenon of overreasoning in large language models (LLMs) during long-chain reasoning and proposes the TRACE tool to perform fine-grained decomposition of the thinking process, relationship annotation, process graph construction, and pattern clustering, identifying two main thinking patterns (Explorer and Late Landing), as well as a structured definition and management strategies for overreasoning based on thinking structure.
Do MLLMs Capture How Interfaces Guide User Behavior? A Benchmark for Multimodal UI/UX Design Understanding
Jaehyun Jeon (Yonsei University), Youngjae Yu (Seoul National University)
Recommendation SystemExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Created the WiserUI-Bench benchmark to evaluate the understanding of multi-modal large language models regarding the impact of UI/UX design on user behavior, and proposed two tasks: selection and explanation.
Do Morals Guide How LLMs Think? The Role of Ethical Perspectives in General Problem Solving
Iseo Kim (Hankuk University of Foreign Studies), Juae Kim (Hankuk University of Foreign Studies)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: This paper investigates the impact of moral conditions on the general reasoning performance of large language models (LLMs) in non-ethical tasks by injecting moral perspectives at different levels (based on Kohlberg's three-stage moral theory) into LLMs and controlling the model's moral stance through prompts.
Do Not Step Into the Same River Twice: Learning to Reason from Trial and Error
Chenming Tang (Peking University), Yunfang Wu (Peking University)
OptimizationRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Propose a prompting mechanism called LTE based on the model's own errors, using the model's trial-and-error results to guide subsequent sampling, thereby addressing the exploration stagnation problem in RLVR.
Doc-V^*: Coarse-to-Fine Interactive Visual Reasoning for Multi-Page Document VQA
Yuanlei Zheng (Huazhong University of Science and Technology), Xiang Bai (Huazhong University of Science and Technology)
RetrievalOptimizationTransformerReinforcement LearningAgentic AIPrompt EngineeringVision Language ModelImageTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Proposes the DocV* OCR-free interactive framework, treating multi-page DocVQA as a serialized evidence aggregation process, utilizing global thumbnail previews, semantic retrieval, and page acquisition actions, performing grounded reasoning in a structured working memory;
DocLens: A Tool-Augmented Multi-Agent Framework for Long Visual Document Understanding
Dawei Zhu (Peking University), Jinsung Yoon (Google)
RetrievalExplainability and InterpretabilityComputational EfficiencyTransformerAgentic AIPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Answering questions on long visual documents through a tool-enhanced multi-agent framework for evidence localization and reasoning.
Does Memory Need Graphs? A Unified Framework and Empirical Analysis for Long-Term Dialog Memory
Sen Hu (Peking University), Lei Zou (Zhejiang University)
RetrievalRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelContrastive LearningTextGraphBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed a unified framework for decomposing and systematically evaluating long-term dialogue memory systems, with comparative experiments conducted under both graph and non-graph structures.
Does RLVR Extend Reasoning Boundaries? Investigating Capability Expansion in Vision-Language Models
Minghe Shen (University College London), Che Liu (University of California, Los Angeles)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelContrastive LearningImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Investigate the extension effects of Reinforcement Learning with Verifiable Rewards (RLVR) on the spatial reasoning capabilities of Vision-Language Models (VLM), using the verifiable maze navigation experimental framework Ariadne to evaluate the reasoning boundary.
Does Self-Consistency Improve the Recall of Encyclopedic Knowledge?
Sho Hoshino (CyberAgent), Peinan Zhang (CyberAgent)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed a subject-level knowledge recall and symbolic reasoning split based on the '=' symbol on the MMLU benchmark, achieving a maximum accuracy of 89% under zero-shot conditions using self-consistency technology.
Domain Generalizable AI Guardrails with Augmented Policy Training
Minqian Liu (Virginia Tech), Mark Dredze (Bloomberg)
ClassificationSafty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: This paper studies the generalization ability of AI guardrail models when facing unseen policies and new domains, and proposes an Augmented Policy Training (APT) training method to enhance generalization by strategically perturbing policy definitions.
Don’t Act Blindly: Robust GUI Automation via Action-Effect Verification and Self-Correction
Yuzhe Zhang (Beijing University of Technology), Haiwei Wang (Baidu Inc.)
Autonomous DrivingRobotic IntelligenceTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelVision-Language-Action ModelDiffusion modelAuto EncoderImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: This paper proposes VeriGUI, a GUI automation agent based on a visual language model, which can verify the effects of its actions after execution and perform self-correction when failures occur.
Don’t Adapt Small Language Models for Tools; Adapt Tool Schemas to the Models
Jonggeun Lee (Seoul National University), Yohan Jo (Seoul National University)
OptimizationComputational EfficiencyRepresentation LearningData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes PA-Tool, an untrained tool architecture optimization method, which reduces pattern mismatch errors during tool usage by remapping tool and parameter names into naming patterns that the model is already familiar with during the pre-training phase.
Don’t Be Misled by Style: A Style-Adaptive Reranker for Capturing Effective Knowledge in Retrieval-Augmented Generation
Ruwen Zhang (Southeast University), Jiuxin Cao (Southeast University)
GenerationData SynthesisRetrievalTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Propose a re-ranking model called SARK that adapts to different writing styles, focusing on capturing core knowledge rather than surface-level style features in retrieval-augmented generation.
Don’t Click That: Teaching Web Agents to Resist Deceptive Interfaces
Yilin Zhang (Renmin University of China), Yueguo Chen (Renmin University of China)
Anomaly DetectionSafty and PrivacyExplainability and InterpretabilityReinforcement LearningPrompt EngineeringVision Language ModelDiffusion modelImageTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed the DUDE framework, which uses a two-stage approach (Hybrid-Reward Learning + Experience Summarization) to enhance the robustness of Vision-Language Web Agents against UI deception.
Don’t Corrupt the Fact: A Trustworthy RAG Watermarking Framework based on Dual Factual Shield
Hao Huang (Chinese Academy of Sciences), Yuling Liu (Chinese Academy of Sciences)
Safty and PrivacyExplainability and InterpretabilityTransformerPrompt EngineeringContrastive LearningTextRetrieval-Augmented Generation
🎯 What it does: Propose the Dual Factual Shield (DFS) framework, which constructs a three-stage post-watermarking process for RAG systems, balancing security and factual fidelity;
DORA: A Dual-Objective Reinforcement Learning Framework for Effective and Efficient Multimodal Agentic Search
Guangming Qin (Beijing Institute of Technology), Lei Cao (University of Arizona)
OptimizationComputational EfficiencyTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringImageTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Propose a dual-objective reinforcement learning framework called DORA for visual question answering, specifically trained to efficiently and accurately invoke search engines and gradually integrate external knowledge in multimodal large models (MLLMs).
Double: Breaking the Acceleration Limit via Double Retrieval Speculative Parallelism
Yuhao Shen (Zhejiang University), Cong Wang (Zhejiang University)
Computational EfficiencyAI Code AssistantTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Propose the DOUBLE framework, combining iterative retrieval-based draft generation with target retrieval verification to achieve parallel inference acceleration.
DPC: Training-Free Text-to-SQL Candidate Selection via Dual-Paradigm Consistency
Boyan Li (Hong Kong University of Science and Technology (Guangzhou)), Yuyu Luo (Hong Kong University of Science and Technology (Guangzhou))
Computational EfficiencyData-Centric LearningAI Code AssistantTransformerLarge Language ModelAgentic AIPrompt EngineeringTextTabularBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes the DPC framework, which selects candidate SQL without training by utilizing adversarial MDD and SQL/Python dual paradigm consistency.
DPDV: Dual-Pathway and Dual-View Representation Learning for Bridging Information Asymmetry in Text-Video Retrieval
Zequn Xie (Zhejiang University), Tao Jin (Zhejiang University)
RetrievalRepresentation LearningTransformerVision Language ModelContrastive LearningVideoTextMultimodality
🎯 What it does: Propose a dual-channel dual-view model (DPDV) for text-video retrieval, addressing the problem of information asymmetry.
DPEPO: Diverse Parallel Exploration Policy Optimization for LLM-based Agents
JunShuo Zhang (University of Electronic Science and Technology of China), Shen Gao (University of Electronic Science and Technology of China)
Autonomous DrivingOptimizationRobotic IntelligenceTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIPrompt EngineeringText
🎯 What it does: Propose a parallel environment interaction framework and develop the DPEPO algorithm, which uses multi-layer rewards to encourage LLM agents to perform diverse parallel exploration;
DPWriter: Reinforcement Learning with Diverse Planning Branching for Creative Writing
Qian Cao (Renmin University of China), Han Li (Kuaishou Technology)
GenerationData SynthesisReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Propose the DPWriter framework, which introduces an explicit planning phase by utilizing a semi-structured long Chain-of-Thought (CoT), and employs diverse planning branching and group-based diversity rewards in reinforcement learning to enhance the output diversity of creative writing.
DR-Arena: an Automated Evaluation Framework for Deep Research Agents
Yiwen Gao (National University of Singapore), Wenxuan Zhang (Singapore University of Technology and Design)
Large Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose DR-Arena, a fully automated evaluation framework that constructs an information tree by real-time web scraping and uses an automated Examiner with an adaptive evolutionary loop to conduct dual-dimensional tests on Deep Research agents, including depth (multi-hop reasoning) and breadth (information aggregation).
Draft, Verify, Restore: Self-Refining Historical Inscription Restoration with a Unified MLLM
Yuyi Zhang (South China University of Technology), Lianwen Jin (South China University of Technology)
RestorationTransformerLarge Language ModelMixture of ExpertsVision Language ModelRectified FlowImageTextMultimodalityBenchmark
🎯 What it does: Proposed a unified multimodal large language model, UniHIR, for end-to-end restoration of damaged inscriptions' text and appearance.
DREAM-S: Speculative Decoding with Searchable Drafting and Target-Aware Refinement for Multimodal Generation
Zining Liu (New York University), Sai Qian Zhang (New York University)
GenerationComputational EfficiencyKnowledge DistillationNeural Architecture SearchTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes DREAM-S, a speculative decoding framework specifically designed for vision-language models (VLMs), aimed at significantly accelerating multimodal generation.
DREAM: Deep Research Evaluation with Agentic Metrics
Elad Ben Avraham (Aws Agentic Ai), Ron Litman (Aws Agentic Ai)
Explainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed and implemented the DREAM framework, which conducts multi-dimensional, reference-free evaluation of long reports generated by deep research agents by constructing a queryable evaluation protocol and using tool-calling agents.
DRInQ: Evaluating Conversational Implicature with Controlled Context Variation
Hirona Jacqueline Arai (University of Southern California), Xiang Ren (University of Southern California)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes the DRInQ benchmark, which uses a semi-automated pipeline to generate multiple-choice instances with varying contexts that induce different meanings under the same question sentence, and evaluates the performance of LLMs in dialogue meaning reasoning.
Dual Activation-Weight Sparsity: A Training-Free Framework for Efficient Large Language Model Compression
Luoyang Sun (Chinese Academy of Sciences), Jun Wang (University College London)
CompressionComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelText
🎯 What it does: Propose the DAWS framework, a training-free dual activation-weight sparse compression scheme, which utilizes three-layer threshold routing to route high-energy activations through full weights, directly discard low-energy activations, and use sparse weights with magnitude pruning for activations in between, achieving efficient LLM inference.
Dual Alignment Between Language Model Layers and Human Sentence Processing
Tatsuki Kuribayashi (MBZUAI), Ethan Gotlieb Wilcox (Georgetown University)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
🎯 What it does: A systematic investigation into the relationship between the surprisal of internal layers in Transformer models and human sentence processing
DUAL RM: Beyond Rule-based Preference Reward Modeling via Meta-Reward
Xiaobo Liang, Min Zhang (Soochow University)
OptimizationExplainability and InterpretabilityRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Built a dual reward framework (DUAL RM) that integrates discriminative and generative reward models, using non-parametric meta-rewards to guide generative models, solving the recursive validation bottleneck in reward models (Matryoshka Doll Problem);
Dual-Axis Generative Reward Model Toward Semantic and Turn-taking Robustness in Interactive Spoken Dialogue Models
Yifu Chen (Zhejiang University), Zhou Zhao (Zhejiang University)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelTextMultimodalityChain-of-ThoughtAudio
🎯 What it does: Propose a dual-axis generative reward model that evaluates the semantic quality of complete bidirectional dialogues and the temporal/interaction timing, providing binary rewards.
Dual-Cluster Memory Agent: Resolving Multi-Paradigm Ambiguity in Optimization Problem Solving
Xinyu Zhang (Xi'an Jiaotong University), Jun Liu (Xi'an Jiaotong University)
OptimizationAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the Dual-Cluster Memory Agent (DCM-Agent), which addresses the ambiguity problem in optimization problems by categorizing historical solution cases into two clusters: modeling and encoding, and extracting three types of structured knowledge—methods, checklists, and pitfalls—from them;