ICML 2026 Papers — Page 14
International Conference on Machine Learning · 6554 papers
Decomposed On-Policy Distillation for Vision-Language Reasoning: Steering Gradients for Visual Grounding
Hee Suk Yoon (Korea Advanced Institute of Science and Technology), Chang D. Yoo (Korea Advanced Institute of Science and Technology)
Knowledge DistillationRepresentation LearningTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: This paper proposes a decomposed on-policy distillation method for visual language reasoning and designs a Visual Gradient Guidance (VGS) to enhance the model's visual perception ability.
Decomposing Out-of-Distribution Error in Conditional Flow Matching via Wasserstein Geometry
Long HC Pham (University of Engineering and Technology, Vietnam National University)
GenerationDomain AdaptationExplainability and InterpretabilityFlow-based ModelImage
🎯 What it does: In the conditional flow matching framework, the OOS error is decomposed into three parts: interpolation sparsity, geometric distortion, and training fitting, and the corresponding theoretical upper bound in the Wasserstein space is provided.
Decomposing Query-Key Feature Interactions Using Contrastive Covariances
Andrew Lee (Harvard University), Martin Wattenberg (Harvard University)
Explainability and InterpretabilityRepresentation LearningTransformerContrastive LearningText
🎯 What it does: This paper proposes a contrastive covariance decomposition of the query-key (QK) space in attention heads, decomposing it into interpretable low-rank subspaces, and verifies the effectiveness of this method in toy models and large language models.
Decomposing the Basic Abilities of Large Language Models: Mitigating Cross-Task Interference in Multi-Task Instruct-Tuning
Bing Wang (Jilin University), Masashi Sugiyama (RIKEN Center for Advanced Intelligence Project)
OptimizationComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningMixture of ExpertsText
🎯 What it does: Proposes a new multi-task instruction fine-tuning method called BADIT, which alleviates cross-task interference by decomposing LLM parameters into orthogonal basic capabilities (LoRA experts) and dynamically maintaining their orthogonality during training.
Decomposition-Based Modular Conformal Prediction for Two-Stage Modeling
William Zhang (Massachusetts Institute of Technology), Georgia Perakis (Massachusetts Institute of Technology)
Anomaly DetectionFederated LearningExplainability and InterpretabilityComputational EfficiencyData-Centric LearningTabularTime SeriesSequentialFinance Related
🎯 What it does: Proposes a synthetic prediction framework for two-stage modular models, achieving uncertainty attribution at each pipeline stage through stage-wise decomposition of residuals, and determining scaling parameters via FWER-controlled risk calibration methods;
DECOR: Learning to Decompose and Collaborate in Deep Search via Multi-Agent Reinforcement Learning
Ruiqing Chen (Ant Group), Lin Zhou (Ant Group)
OptimizationFederated LearningExplainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelReinforcement LearningAgentic AITextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the DECOR framework, which decomposes the deep search task into a multi-agent system consisting of three roles: Planner, Filter, and Answerer;
Decouple and Cache: KV Cache Construction for Streaming Video Understanding
Zhanzhong Pang (National University of Singapore), Angela Yao (National University of Singapore)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelVision Language ModelVideoTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Designed and implemented a training-free KV cache construction framework called DSCache, which is used in resource-constrained streaming video understanding tasks to real-time construct and maintain accumulated cache and instant cache, thereby achieving efficient inference for infinitely long video streams.
Decouple Searching from Training: Scaling Data Mixing via Model Merging for Large Language Model Pre-training
Shengrui Li (Xiaohongshu Inc.), Shaosheng Cao (Xiaohongshu Inc.)
Computational EfficiencyKnowledge DistillationData-Centric LearningTransformerLarge Language ModelMixture of ExpertsContrastive LearningTextMultimodality
🎯 What it does: Decouple the data mixing ratio search from model training by using model fusion (weighted model merging) to build a training-free proxy model;
Decoupled Low-Rank Adaptation for Robust Federated Fine-Tuning
Xiuwen Fang (Wuhan University), Mang Ye (Wuhan University)
ClassificationFederated LearningSupervised Fine-TuningContrastive LearningImageText
🎯 What it does: Proposes a robust low-rank adaptation method for federated fine-tuning with noisy labels.
Decoupled Training with Local Reinforcement Fine-Tuning in Federated Learning
Yuting Ma (University of Science and Technology of China), Xiaohua Xu (Hefei University of Technology)
ClassificationDomain AdaptationFederated LearningRepresentation LearningTransformerSupervised Fine-TuningReinforcement LearningPrompt EngineeringContrastive LearningImageTextMultimodality
🎯 What it does: Propose the FedDTL framework, which balances global task adaptation and generalization in federated learning through decoupled training with local image encoders and server-side text encoders, along with two-phase (supervised fine-tuning + reinforcement learning) local fine-tuning.
Decoupling Reasoning and Confidence: Resurrecting Calibration in Reinforcement Learning from Verifiable Rewards
Zhengzhao Ma (Chinese Academy of Sciences), Le Sun (Chinese Academy of Sciences)
OptimizationExplainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringContrastive LearningText
🎯 What it does: Proposed a Decoupled Calibration Policy Optimization (DCPO) framework, aiming to decouple the reasoning accuracy and confidence calibration of large language models, addressing the overconfidence problem generated during RLVR training.
Decoupling Regularization and Privacy in Differentially Private Ridge Regression and ERM
Wanjie Wang (National University of Singapore), Tathagata Banerjee (National University of Singapore)
OptimizationSafty and PrivacyTabular
🎯 What it does: This paper proposes a new framework for differential privacy ridge regression and ridge-regularized empirical risk minimization (ERM), achieving decoupling and optimization of statistical error and privacy noise by splitting the regularization parameter into statistical regularization α and privacy stabilization parameter c.
Decoupling Skeleton and Flesh: Efficient Multimodal Table Reasoning with Disentangled Alignment and Structure-aware Guidance
Yingjie Zhu (Harbin Institute of Technology), Min Zhang (Harbin Institute of Technology)
Computational EfficiencyRepresentation LearningData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelImageTextMultimodalityTabularChain-of-Thought
🎯 What it does: This paper proposes an efficient multi-modal table reasoning framework. It first achieves structure-content separation alignment through DISCO, and then utilizes Table-GLS to guide reasoning with global-local structural guidance during inference.
Decoupling The "What" and "Where" With Polar Coordinate Positional Embedding
Anand Gopalakrishnan (Swiss AI Lab IDSIA USI SUPSI), Michael Curtis Mozer (University of Colorado)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextBiomedical DataAudio
🎯 What it does: Proposed a new relative position encoding method called PoPE, aimed at separating the mixed information of 'content-what' and 'position-where' in Transformers, addressing the difficulties encountered by RoPE during independent matching;
Decoupling Universal Laws and Environmental Heterogeneity: A Physics-Inspired Framework for Robust Spatio-Temporal Forecasting
Aoyu Liu (Tongji University), Yaying Zhang (Tongji University)
Autonomous DrivingOptimizationGraph Neural NetworkTransformerDiffusion modelContrastive LearningGraphTabularTime SeriesBenchmarkPhysics RelatedStochastic Differential Equation
🎯 What it does: Propose the STPDE framework, modeling spatiotemporal prediction as the evolution of a nonhomogeneous partial differential equation governed by an invariant diffusion operator and an environmental substrate.
Decoupling Variance and Scale-Invariant Updates in Adaptive Gradient Descent for Unified Vector and Matrix Optimization
Zitao Song (Purdue University), David F. Gleich (Purdue University)
OptimizationImageText
🎯 What it does: Proposed the DeVA framework, decoupling the variance adaptation of AdaGrad from scale-invariant updates, supporting vector and matrix optimization.
DecoVer: A Decompose-and-Verify Neuro-Symbolic Framework for Embodied Task Planning with BC+
YiXiang Jiang, Xiangbo Shu (Nanjing University of Science and Technology)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringTextMultimodalityRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the DecoVer framework, which transforms natural language tasks into executable BC+ formal programs through decomposition and verification, achieving embedded task planning.
Deep Coupling Learning for Solving PDEs
Lingshi MENG (Chinese University of Hong Kong), Sinno Jialin Pan (Chinese University of Hong Kong)
OptimizationAuto EncoderContrastive LearningPhysics RelatedStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Proposed the CoupledNet structure, enabling deep PINN to maintain numerical stability and achieve higher accuracy when solving PDEs.
Deep Discriminative Structure Proxy Hashing for Cross-modal Retrieval
Kun Cheng (Qufu Normal University), Lei Huang (Ocean University of China)
RetrievalGraph Neural NetworkTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: DDSPH is proposed by constructing sample-specific proxy relationship graphs for each sample, enabling proxies to compete with each other and explicitly contrast positive and negative responses, thus achieving more discriminative cross-modal hashing.
Deep Ensemble Clustering for Visual Representation Learning
Yuwei Wang (Nanjing University of Science and Technology), Wenguan Wang (Zhejiang University)
ClassificationObject DetectionSegmentationRepresentation LearningTransformerDiffusion modelAuto EncoderContrastive LearningImage
🎯 What it does: Propose ENFORMER, a visual backbone that integrates clustering into the visual feature extraction process, leveraging multiple differentiable clustering methods to learn richer and more interpretable visual representations.
Deep Flow Networks
Ozan Candogan (University of Chicago), Ayoub Foussoul (University of Chicago)
OptimizationFlow-based ModelOptical FlowTabular
🎯 What it does: Designed and verified a discrete function approximator based on the minimum cost flow value function — Deep Flow Networks (DFN).
Deep Forcing: Training-Free Long Video Generation with Deep Sink and Participative Compression
Jung Yi (KAIST AI), Seungryong Kim (KAIST AI)
GenerationCompressionTransformerPrompt EngineeringDiffusion modelAuto EncoderVideo
🎯 What it does: Proposes an untrained long video generation framework called Deep Forcing, which reduces error accumulation during long-term inference through two mechanisms: Deep Sink and Participative Compression.
Deep Incentive Design with Differentiable Equilibrium Blocks
Vinzenz Thoma (ETH Zurich), Luke Marris (Google DeepMind)
OptimizationGraph Neural NetworkTransformerReinforcement LearningDiffusion modelScore-based ModelFlow-based ModelRectified FlowAuto EncoderGenerative Adversarial NetworkContrastive LearningGraphTabularBenchmark
🎯 What it does: Propose the Deep Incentive Design (DID) framework, which utilizes a differentiable equilibrium module (DEB) training mechanism to generate a mechanism generator that addresses multi-agent incentive design problems in a unified manner.
Deep Learning for BioImaging: What Are We Really Learning?
Ivan Svatko (Universit´ e Paris Cit´ e), Auguste Genovesio (Universit´ e PSL)
ClassificationRetrievalRepresentation LearningConvolutional Neural NetworkGraph Neural NetworkTransformerAuto EncoderContrastive LearningImagePoint CloudGraphBiomedical DataReview/Survey PaperBenchmark
🎯 What it does: A systematic evaluation of representation learning on microscopy images was conducted, proposing and utilizing simple baselines (pixel statistics, untrained networks, and structural image representations), and comparing the performance of multiple baselines, pre-trained and non-pre-trained models on publicly available benchmarks at the cell culture and tissue levels.
Deep Learning of Compositional Targets with Hierarchical Spectral Methods
Hugo Tabanelli (Ecole polytechnique federale de lausanne), Florent Krzakala (Ecole polytechnique federale de lausanne)
OptimizationRepresentation LearningDiffusion modelScore-based ModelContrastive LearningGaussian SplattingTabularBenchmark
🎯 What it does: This paper studies the sample complexity of deep networks in learning high-order polynomial target functions under the control environment of high-dimensional Gaussian inputs, and proposes a hierarchical spectral estimation method to achieve this learning.
Deep Multi-view Graph Clustering via Attribute-aware Bidirectional Structural Refinement and Pseudo-label Guided Multi-level Fusion
Youqing Wang (Beijing University of Chemical Technology), Jipeng Guo (Beijing University of Chemical Technology)
ClassificationRepresentation LearningGraph Neural NetworkDiffusion modelScore-based ModelFlow-based ModelRectified FlowNeural Radiance FieldAuto EncoderContrastive LearningGaussian SplattingGraph
🎯 What it does: Propose a deep multi-view graph clustering framework called APGC, which integrates attribute-aware bidirectional structural refinement and pseudo-label guided multi-level fusion.
Deep Networks Learn Deep Hierarchical Models
Amit Daniely (Google Research and Hebrew University)
ClassificationOptimizationRepresentation LearningConvolutional Neural NetworkContrastive LearningTabular
🎯 What it does: Prove that using hierarchical stochastic gradient descent (SGD) on residual networks can learn multi-label classification models with label hierarchical structures in polynomial time.
Deep networks learn to parse uniform-depth context-free languages from local statistics
Jack T. Parley (Ecole Polytechnique Federale De Lausanne), Matthieu Wyart (Johns Hopkins University)
Computational EfficiencyRepresentation LearningData-Centric LearningConvolutional Neural NetworkTransformerContrastive LearningText
🎯 What it does: This paper studies how deep networks can learn to parse a unified context-free language based solely on local statistics, proposing a tunable stochastic hierarchical model and providing a provable learning algorithm.
Deep Neural Network Regression with Functional Covariates
Hang Zhou (University of North Carolina at Chapel Hill), Jane-Ling Wang (University of California, Davis)
Data-Centric LearningTabularTime SeriesElectronic Health Records
🎯 What it does: This paper studies the theoretical performance of deep neural networks in functional covariate regression problems, proving that ReLU networks can achieve optimal polynomial convergence rates under functional linear and functional generalized linear models; it also provides a lower bound in the fully nonparametric case and points out that polynomial convergence rates cannot be achieved.
Deep neural networks divide and conquer dihedral multiplication
Sihui Wei (McGill University), Jonathan Love (Leiden University)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerGraph
🎯 What it does: Study the learning process of deep networks on dihedral group multiplication tasks, revealing a universal divide-and-conquer algorithm;
Deep Pre-Alignment for VLMs
Tianyu Yu (Tsinghua University), Yuan Yao (Tsinghua University)
Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Proposes the Deep Pre-Alignment (DPA) architecture, replacing the standard ViT encoder with a small Perceiver VLM as the perceiver to achieve deep alignment of visual features before they enter the large language model, thereby reducing the alignment burden on the LLM and improving performance on multi-modal tasks.
Deep Reinforcement Learning Finds Bayes-Nash Equilibrium in Competitive Newsvendor Problems
Kassian Köck, Martin Bichler (Technical University of Munich)
OptimizationReinforcement LearningTabular
🎯 What it does: This paper studies the learning dynamics of competitive fresh product suppliers, proving that the game is strictly monotonic under uniform demand, thereby ensuring the uniqueness of the Nash equilibrium, and utilizing deep reinforcement learning (PPO) to converge to this equilibrium through self-play; it also proves that under the Bayesian setting, it converges to the Bayesian Nash equilibrium.
Deep Residual Injection for Full-Spectrum Forensic Signal Perception in Multimodal Large Language Models
Kaiqing Lin (Shenzhen University), Shouhong Ding (Tencent Youtu Lab)
Anomaly DetectionExplainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelAuto EncoderContrastive LearningImageTextMultimodality
🎯 What it does: Propose an architecture named Deep Visual Residual MLLM (Deep-VRM), which introduces low-level generator trace features into the deep layers of a multimodal large language model (MLLM) using a residual injection method, while keeping the semantic representations of early layers intact, achieving full-spectrum detection of AI-generated images.
Deep Scientific Reasoning under Physical Constraints: Structure-Aware Spectrum Prediction
Yingheng Wang (Cornell University), Carla P Gomes
Graph Neural NetworkTransformerContrastive LearningGraphTabularPhysics Related
🎯 What it does: Proposed a general framework called DeepSciReasoner for structure-aware scientific spectral prediction under hard physical constraints.
Deep sequence models tend to memorize geometrically; it is unclear why
Shahriar Noroozizadeh (Carnegie Mellon University), Sanjiv Kumar (Google Research)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningGraph Neural NetworkTransformerSupervised Fine-TuningContrastive LearningGraph
🎯 What it does: Investigate the implicit reasoning capabilities of deep sequence models (such as Transformer and Mamba) in graph structure memory tasks, discovering that models form global geometric memory rather than traditional associative memory.
Deep Single-Index Fréchet Regression
Muqing Cui (University of California), Hans-Georg Müller
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningRecurrent Neural NetworkTransformerDiffusion modelScore-based ModelFlow-based ModelRectified FlowAuto EncoderGenerative Adversarial NetworkContrastive LearningGraphTabularTime SeriesBiomedical DataBenchmark
🎯 What it does: Propose a deep single-index Fréchet regression (DeSI) framework, which combines deep neural networks to learn a one-dimensional projection of the input and performs local Fréchet regression on this projection, enabling high-dimensional prediction of non-Euclidean outputs.
Deep Trajectory Supervision: Deep Supervision Strikes Back
Han Wang, Nicu Sebe
ClassificationComputational EfficiencyKnowledge DistillationConvolutional Neural NetworkTransformerSupervised Fine-TuningFlow-based ModelContrastive LearningImage
🎯 What it does: Propose a new intermediate layer supervision method called Deep Trajectory Supervision (DTS), which guides the model's semantic trajectory by using soft targets corresponding to layer depth at every layer of the network.
DeepAnalyze: Agentic Large Language Models for Autonomous Data Science
Shaolei Zhang (Renmin University of China), Xiaoyong Du (Renmin University of China)
OptimizationData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringMixture of ExpertsTextTabularBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes DeepAnalyze, an agent-based large language model capable of autonomously completing the full workflow of data science from raw data to research reports.
DeepBlip: Estimating Conditional Average Treatment Effects Over Time
Haorui Ma (LMU), Stefan Feuerriegel (LMU)
OptimizationFederated LearningExplainability and InterpretabilityComputational EfficiencyRepresentation LearningRecurrent Neural NetworkTransformerAuto EncoderContrastive LearningTabularTime SeriesBiomedical DataElectronic Health Records
🎯 What it does: Propose DeepBlip, the first neural framework based on SNMM, for estimating time-varying conditional average treatment effects (CATE) and achieving end-to-end training;
DeepHA: Scaling Action Chains Elicits Deep Hierarchical Agents
Zihao Wang (Peking University), Yitao Liang (Peking University)
Robotic IntelligenceTransformerSupervised Fine-TuningReinforcement LearningMixture of ExpertsVision Language ModelVision-Language-Action ModelImageTextMultimodalityChain-of-Thought
🎯 What it does: Propose DeepHA, a unified hierarchical agent capable of dynamically generating and executing multi-level, multi-modal actions during inference;
DeepImageSearch: Benchmarking Multimodal Agents for Context-Aware Image Retrieval in Visual Histories
Chenlong Deng (Renmin University of China), Zhicheng Dou (Renmin University of China)
RetrievalTransformerLarge Language ModelAgentic AIVision Language ModelContrastive LearningImageTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed the DeepImageSearch retrieval paradigm, transforming image retrieval into a multi-step exploration task requiring cross-time and cross-event context reasoning, and constructed the DISBench benchmark and the ImageSeeker baseline framework;
DeepSight: Long-Horizon World Modeling via Latent States Prediction for End-to-End Autonomous Driving
Lingjun Zhang (Tsinghua University), Hong Wang (Tsinghua University)
Autonomous DrivingTransformerVision Language ModelContrastive LearningWorld ModelImageVideoMultimodalityChain-of-Thought
🎯 What it does: Propose a long-term spatiotemporal bird's-eye-view world model called DeepSight based on implicit semantic feature prediction, achieving end-to-end autonomous driving.
DEER: A Benchmark for Evaluating Deep Research Agents on Expert Report Generation
Janghoon Han (LG AI Research), Honglak Lee (LG AI Research)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the DEER benchmark, which evaluates the quality of expert-level reports generated by deep research agents through a fine-grained evaluation system formulated by experts and expert guidance, and introduces a full-report verification module.
DeFacto: Counterfactual Thinking with Images for Enforcing Evidence-Grounded and Faithful Reasoning
Tianrun Xu (Tsinghua University), Feng Chen (Tsinghua University)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerReinforcement LearningPrompt EngineeringVision Language ModelDiffusion modelContrastive LearningImageTextMultimodalityChain-of-Thought
🎯 What it does: Designed and implemented the DeFacto framework, combining adversarial samples and reinforcement learning to achieve multi-step reasoning based on visual evidence and interpretable answers.
Deformba: Vision State Space Model with Adaptive State Fusion
Hongyu Ke (Georgia State University), Haoxin Wang (Georgia State University)
ClassificationObject DetectionSegmentationAutonomous DrivingTransformerMixture of ExpertsContrastive LearningImagePoint Cloud
🎯 What it does: Propose a visual state space model called Deformba, which achieves adaptive spatial fusion by decoupling writing and reading in SSM.
DEGAP: Dynamic Entropy-Guided Attention Perturbation for Contrastive Decoding in Large Vision-Language Models
Hyein Seo (University of Hanyang), Yong Suk Choi (University of Hanyang)
Explainability and InterpretabilityComputational EfficiencyTransformerPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: DEGAP proposes a contrastive decoding scheme that suppresses hallucinations generated by large-scale vision-language models by dynamically masking visual attention and utilizing partial reasoning to generate contrastive logits.
Degradation-Aware Metric Prompting for Hyperspectral Image Restoration
Binfeng Wang (Beijing Institute of Technology), Jing Zhang (Wuhan University)
RestorationExplainability and InterpretabilityComputational EfficiencyConvolutional Neural NetworkTransformerPrompt EngineeringMixture of ExpertsContrastive LearningImage
🎯 What it does: A unified hyperspectral image restoration framework named DAMP is proposed, which utilizes interpretable spatial-spectral metrics as degradation hints and dynamically routes experts through degradation-adaptive Mixture-of-Experts to achieve unified restoration for various degradations.
Delayed Momentum Aggregation: Communication-efficient Byzantine-robust Federated Learning with Partial Participation
Kaoru Otsuka (Okinawa Institute of Science and Technology), Makoto Yamada (Okinawa Institute of Science and Technology)
OptimizationFederated LearningSafty and PrivacyConvolutional Neural NetworkAgentic AIContrastive LearningImage
🎯 What it does: Propose a delayed momentum aggregation principle, construct the DeMoA algorithm, and achieve Byzantine-robust federated learning under partial participation.
Delegation and Verification under AI
Lingxiao Huang (Nanjing University), Nisheeth K. Vishnoi (Yale University)
OptimizationFederated LearningExplainability and InterpretabilityComputational EfficiencyBiomedical Data
🎯 What it does: A static optimization framework was constructed to study how human workers choose between delegating to AI and performing verification within institutional workflows, and how this choice affects the institutional-level assessment of worker quality.
Deliberate Evolution: Agentic Reasoning for Sample-Efficient Symbolic Regression with LLMs
Xinyu Pang (Tsinghua University), Changshui Zhang (Tsinghua University)
OptimizationTransformerLarge Language ModelAgentic AIPrompt EngineeringTextTabularBenchmarkPhysics RelatedRetrieval-Augmented Generation
🎯 What it does: Propose the Deliberate Evolution framework, decoupling the symbolic expression generation of LLMs from search control, utilizing directional operators, diagnostic tools, and reflective memory to improve the efficiency of symbolic regression.
DeltaEvolve: Accelerating Scientific Discovery through Momentum-Driven Evolution
Jiachen Jiang (Ohio State University), Zhihui Zhu (Ohio State University)
OptimizationComputational EfficiencyAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: In large language model-driven program evolution frameworks, replacing complete code with 'semantic increments' and constructing hierarchical databases and progressive disclosure samplers to achieve more efficient evolutionary search;
Delving into Muon and Beyond: Deep Analysis and Extensions
Xianbiao Qi (Intellifusion Inc), Rong Xiao (Intellifusion Inc)
OptimizationTransformerLarge Language ModelTextPhysics Related
🎯 What it does: This paper studies the Muon optimizer and its variants through a spectral transformation framework, systematically evaluating their stability and performance on matrix parameters;
Delving into Non-Exchangeability for Conformal Prediction in Graph-Structured Multivariate Time Series
Ruichao Guo (Shanghai Jiao Tong University), Hesheng Wang (Shanghai Jiao Tong University)
Anomaly DetectionExplainability and InterpretabilityComputational EfficiencyRepresentation LearningGraph Neural NetworkTransformerDiffusion modelScore-based ModelAuto EncoderGenerative Adversarial NetworkContrastive LearningGraphTime SeriesBenchmark
🎯 What it does: Studied confidence interval prediction for graph-structured multivariate time series, proposing conditional exchangeability in the frequency domain and the corresponding SCALE method to achieve reliable conformal prediction.
Demystifying Action Space Design for Robotic Manipulation Policies
Yuchun Feng (Tsinghua University), Xianyuan Zhan (Tsinghua University)
Robotic IntelligenceRecurrent Neural NetworkTransformerReinforcement LearningVision-Language-Action ModelFlow-based ModelImageTime Series
🎯 What it does: Systematically studied the role of action space in robot manipulation policy learning, decomposed along two axes—time (absolute/relative) and space (joint/task)—and conducted large-scale experiments on real robots;
Demystifying Entropy Control in LLM RL Training: Theoretical Analysis and Dynamic Scheduling
Jingchu Gai (Carnegie Mellon University), Aditi Raghunathan (Carnegie Mellon University)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: The role of entropy regularization in LLM reinforcement learning fine-tuning is studied, and a dynamic entropy scheduling method based on entropy discrepancy is proposed.
Demystifying LLM-as-a-Judge: Analytically Tractable Model for Inference-Time Scaling
Indranil Halder (Harvard University), Cengiz Pehlevan (Harvard University)
Explainability and InterpretabilityComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: Propose an analyzable Bayesian linear regression + reward-weighted sampling model to study the impact of sampling scale (k) and temperature (T) on generalization error during inference.
Demystifying Mergeability: Interpretable Properties to Predict Model Merging Success
Luca Zhou (Sapienza University Of Rome), Emanuele Rodolà (Sapienza University Of Rome)
ClassificationExplainability and InterpretabilityTransformerSupervised Fine-TuningContrastive LearningImage
🎯 What it does: Propose an interpretable framework that combines the success rates of 28 interpretable two-model prediction models for metrics using a linear combination, optimized through L1 regularization and leave-one-task-out cross-validation.
Demystifying Multimodal Biomolecular Co-design With Intrinsic Geodesic Coupling
Keyue Qiu (Tsinghua University), Wei-Ying Ma (Tsinghua University)
Drug DiscoveryProtein Structure PredictionTransformerDiffusion modelFlow-based ModelMultimodalityBiomedical Data
🎯 What it does: This paper proposes the GeoCoupling framework, which learns the temporal coupling relationship between two heterogeneous modalities (sequence and structure) during the multi-modal biomolecular co-design process, and treats it as a temporal optimal transport problem;
Demystifying Scientific Problem-Solving in LLMs by Probing Knowledge and Reasoning
Alan Li (Yale University), Arman Cohan (Yale University)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes two unified scientific reasoning benchmarks, SCIREAS and SCIREAS-PRO, and develops the KRUX framework to analyze the interaction between knowledge and reasoning.
Demystifying the Optimal Fair Classifier in Multi-Class Classification
Li Zhang (Zhejiang University), Chaochao Chen (Zhejiang University)
ClassificationOptimizationFederated LearningSafty and PrivacyExplainability and InterpretabilityImageTabular
🎯 What it does: Propose the OptFair framework, theoretically derive the Bayes optimal discriminator for multi-class fair classification, and provide an implementable attribute-blind algorithm to achieve the Pareto frontier between accuracy and fairness.
Demystifying When Pruning Works via Representation Hierarchies
Shwai He (University of Maryland), Ang Li (University of Maryland)
RetrievalExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
🎯 What it does: This paper systematically studies the impact of network pruning on large language models (LLMs), using a representation hierarchy (embedding-log-probability) framework to explain why pruning performs well on non-generation tasks but significantly fails on generation tasks.
Denoising without Diffusion: Fixed-Noise Denoiser Anomaly Detection in Tabular Data
Manuel Hirth (Daimler Truck AG), Enkelejda Kasneci (Technical University of Munich)
Anomaly DetectionTransformerDiffusion modelScore-based ModelContrastive LearningTabularBenchmarkFinance Related
🎯 What it does: Proposes a single-step fixed-noise denoising anomaly detection method for tabular data called DenoiserAD, which utilizes a self-supervised denoiser and calculates stability scores through multiple noise perturbations during inference.
Dense associative memory for Gaussian distributions
Chandan Tankala (Independent Researcher), Krishna Balasubramanian
RetrievalRepresentation LearningAuto EncoderContrastive LearningImageText
🎯 What it does: Extend dense associative memory (DAM) from the vector space to the Gaussian distribution space, define the LSE energy based on the Bures-Wasserstein distance, and provide the corresponding retrieval dynamics.
DenseMLLM: Standard Multimodal LLMs for Dense Prediction
Yi Li (Tencent), Xiaomeng Li (Tencent)
RecognitionSegmentationDepth EstimationConvolutional Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelImageTextMultimodality
🎯 What it does: Developed a standard multimodal large language model, DenseMLLM, capable of directly generating pixel-level dense predictions (semantic segmentation, depth estimation, referring expression segmentation) from visual tokens, without requiring additional decoders or task-specific heads.
DenseSteer: Steering Small Language Models towards Dense Math Reasoning
Yang Ouyang (North Carolina State University), Jung-Eun Kim (North Carolina State University)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationTransformerPrompt EngineeringContrastive LearningTextBenchmarkChain-of-Thought
🎯 What it does: Propose DenseSteer, a steering framework that does not require training during inference, significantly improving mathematical reasoning performance by rewriting sparse reasoning of small language models into denser reasoning.
Density-Aware Translation of Spurious Correlations in Zero-Shot VLMs
Afsaneh Hasanebrahimi (University of Melbourne), Sarah Monazam Erfani
ClassificationDomain AdaptationExplainability and InterpretabilityRepresentation LearningTransformerPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: Proposes a zero-shot compensation method based on local geometric density called Density-Aware Translation (DAT), aimed at eliminating biases in vision-language models when facing spurious correlations.
Density-Guided Robust Counterfactual Explanations on Tabular Data under Model Multiplicity
Jun Tan (Central South University), Ning Gui (Central South University)
OptimizationExplainability and InterpretabilityRecurrent Neural NetworkTransformerScore-based ModelFlow-based ModelContrastive LearningTabularTime SeriesBenchmarkStochastic Differential Equation
🎯 What it does: Proposes the DensityFlow framework, which generates robust counterfactual explanations (CEs) by combining continuous-time Neural ODE with Noise Contrastive Estimation (NCE), ensuring the validity and robustness of explanations in multi-model (model multiplicity) scenarios.
Dependence-Aware Label Aggregation for LLM-as-a-Judge via Ising Models
Krishna Balasubramanian (Amazon Web Service), Shiva Kasiviswanathan (Amazon Web Service)
ClassificationExplainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelText
🎯 What it does: This paper proposes considering dependencies among judges when aggregating LLM evaluators, and constructs a hierarchical aggregation method using the Ising model.
Depth over Fidelity in Fixed-Budget Noisy Evolution Strategies
Sichen Wang (Shenzhen MSU BIT University), Zhipeng Lu (Shenzhen MSU BIT University)
OptimizationReinforcement LearningTabularTime SeriesBenchmark
🎯 What it does: In fixed-budget noisy evolutionary strategies, the idea of depth over accuracy is proposed, and probability elite members (PEM), residual bootstrap PEM (RB-PEM), and an adaptive probe-switch mechanism are designed to improve the handling of uncertainty during the ranking phase in the presence of noise.
Depth-Breadth Synergy in RLVR: Unlocking LLM Reasoning Gains with Adaptive Exploration
Zhicheng Yang (Hong Kong University of Science and Technology (Guangzhou)), Jing Tang (Hong Kong University of Science and Technology (Guangzhou))
OptimizationComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringMixture of ExpertsTextBenchmark
🎯 What it does: This paper investigates the interaction between depth (problem difficulty) and breadth (number of training instances) in the Reinforcement Learning with Verifiable Rewards (RLVR) framework, and proposes Difficulty-Adaptive Episode Sampling (DARS) as well as the DARS-Breadth method combining DARS with large-batch training, to enhance the reasoning performance of large language models (LLMs).
Depth-Progressive Monotonic Learning without Global Backpropagation
Chenhao Ye (Southern University of Science and Technology), Ming Tang (Southern University of Science and Technology)
ClassificationImage TranslationOptimizationKnowledge DistillationConvolutional Neural NetworkTransformerDiffusion modelScore-based ModelAuto EncoderGenerative Adversarial NetworkContrastive LearningImageTextGraph
🎯 What it does: Designed Depth-Progressive Monotonic Learning (DMoL), training without global backpropagation by setting local cooperative objectives for each layer
Derivative Informed Learning of Exchange-Correlation Functionals
Eike Eberhard, Stephan Günnemann
OptimizationKnowledge DistillationDrug DiscoveryGraph Neural NetworkDiffusion modelScore-based ModelAuto EncoderContrastive LearningGraphTabularPhysics RelatedStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Train machine learning XC functionals by distilling the energy and density information of high-order hybrid functional B3LYP through Grassmannian supervised learning with gradients and Hessians.
Deriving Neural Scaling Laws from the Statistics of Natural Language
Francesco Cagnetta (International School for Advanced Studies), Matthieu Wyart (Johns Hopkins University)
GenerationData-Centric LearningTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Proposed a theoretical framework to derive the exponential learning curve of large language models (LLMs) under data scarcity conditions from natural language statistics, and validated it on two distinct text datasets: TinyStories and WikiText.
Design Linear Constrained Neural Layers with Implicit Convex Optimization
Junchi Yan (Shanghai Jiao Tong University), Liangliang Shi (Shanghai Institute for Mathematics and Interdisciplinary Sciences)
OptimizationFederated LearningComputational EfficiencyGraphTabularFinance Related
🎯 What it does: Propose a pluggable differentiable linear constraint layer (BLCLayer/GLCLayer) that implicitly realizes hard linear constraints through convex optimization;
Design-Based Anytime-Valid Inference for Randomized Experiments with Delayed Outcomes and Staggered Entry
Michael Lindon (Netflix), Nathan Kallus (Cornell University)
Anomaly DetectionOptimizationFederated LearningExplainability and InterpretabilityComputational EfficiencyData-Centric LearningReinforcement Learning from Human FeedbackLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIPrompt EngineeringMixture of ExpertsDiffusion modelScore-based ModelContrastive LearningGaussian SplattingTabularTime SeriesSequentialReview/Survey PaperBenchmarkFinance RelatedRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes a design benchmark any-time valid inference method for randomized experiments with delayed outcomes and staggered entry, constructing confidence sequences for the cumulative reward process.
Designing Observation and Action Models for Efficient Reinforcement Learning with LLMs
Deunsol Yoon (LG AI Research), Kanghoon Lee (LG AI Research)
Computational EfficiencyRobotic IntelligenceTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: Designing the observation and action spaces in reinforcement learning using large language models (LLMs), generating efficient agent spaces from raw environments;
Desirable Effort Fairness and Optimality Trade-offs in Strategic Learning
Valia Efthymiou (Massachusetts Institute of Technology), Chara Podimata (Massachusetts Institute of Technology)
OptimizationFederated LearningExplainability and InterpretabilityReinforcement Learning from Human FeedbackReinforcement LearningTabularFinance Related
🎯 What it does: This paper proposes a unified framework that simultaneously considers feature desirability, causal propagation, and fair incentives in strategic learning, and seeks the optimal decision strategy under fair constraints within this framework.
Detached Skip-Links and $R$-Probe: Decoupling Feature Aggregation from Gradient Propagation for MLLM OCR
Ziye Yuan (Peking University), Ming Zhang (Peking University)
RecognitionExplainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: In multi-modal large language models, Detached Skip-Links are proposed, which separate feature aggregation from gradient propagation by applying stop-gradient to shallow skips during multi-layer feature fusion. At the same time, R-Probe is introduced, a diagnostic tool that utilizes shallow LLM decoders for pixel-level reconstruction to evaluate the fine-grained information retention and availability of visual features to LLMs.
DetailMaster: Can Your Text-to-Image Model Handle Long Prompts?
Qirui Jiao (Sun YatSen University), Yaliang Li (Alibaba Group)
GenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringVision Language ModelDiffusion modelImageTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Designed and released the DETAILMASTER benchmark to evaluate the generation capabilities of text-to-image models under long, detail-rich prompts, and provided an automated data construction and evaluation process.
Detecting and Filtering Unsafe Training Data via Data Attribution with Denoised Representation
Yijun Pan (Yale University), Jiaqi W. Ma (University of Illinois Urbana-Champaign)
Anomaly DetectionSafty and PrivacyExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelAuto EncoderContrastive LearningText
🎯 What it does: Propose a data attribution method based on denoised representations (DRA) to detect and filter unsafe samples in LLM training data.
Detecting Contextual Hallucinations in Large Language Models with Frequency-Aware Attention
Siya Qi (King's College London), Lin Gui (King's College London)
Anomaly DetectionExplainability and InterpretabilityTransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose a frequency-aware attention analysis method to detect contextual hallucinations in large language models during context-based generation.
Detecting Errors in AI-Generated Annotations: When and Why Semantic Neighbors Help
Na Di (Hong Kong University of Science and Technology), Zhaowei Zhu (BIAI, ZJUT)
Anomaly DetectionData-Centric LearningTransformerLarge Language ModelVision Language ModelContrastive LearningImageTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Propose the SAGE method, which utilizes kNN to retrieve semantically similar samples as reference points for error detection
Detecting Fluent Optimization-Based Adversarial Prompts via Sequential Entropy Changes
Mohammed Alshaalan (University College London), Miguel R. D. Rodrigues (University College London)
Anomaly DetectionAdversarial AttackTransformerLarge Language ModelPrompt EngineeringTextSequential
🎯 What it does: This paper proposes an online change point detection method based on continuous entropy variation (CPD Online), which standardizes each user input token using an entropy benchmark hinted by the system, and captures the persistent entropy shift after an attack through first-order Page-CUSUM recursion, thereby detecting and locating optimized attack suffixes in real-time without requiring additional training, being model-agnostic, and being operational in real-time.
Detecting Perspective Shifts in Multi-Agent Systems
Eric W Bridgeford (Helivan), Hayden Helm (Helivan)
Anomaly DetectionExplainability and InterpretabilityRepresentation LearningContrastive LearningTabularTime Series
🎯 What it does: This study proposes the Temporal Data Kernel Perspective Space (TDKPS) framework for detecting changes in agent perspectives over time in black-box multi-agent systems, and implements hypothesis testing at both the agent layer and group layer;
Detecting the Semantic Fixed Point: A Geometric Framework for Efficient Inference
Jiawei Gu (Great Bay University), Xiao Luo (University of Wisconsin-Madison)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Designed a early stopping mechanism called GCEE based on the geometric convergence of Transformer hidden states, directly monitoring the magnitude and directional stability of hidden state updates to determine when to stop inference.
Deterministic Component Mining for Multi-Framework UI2Code Generation
Zixiong Yang (Peking University), Jiechao Gao (Stanford University)
GenerationAI Code AssistantTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodality
🎯 What it does: Proposed a three-stage Deterministic Component Mining (DCM) pipeline for generating reusable frontend code (React, Vue, Angular, Vanilla) from webpage screenshots.
Deterministic Differentiable Structured Pruning for Large Language Models
Weiyu Huang (Tsinghua University), Jianfei Chen (Tsinghua University)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelMixture of ExpertsRectified FlowText
🎯 What it does: This paper proposes a deterministic differential structural pruning method called DDP, which only optimizes masks, achieving efficient pruning of large language models.
Deterministic Inference across Tensor Parallel Sizes That Eliminates Training-Inference Mismatch
Ziyang Zhang (Independent Researcher), Zirui Liu (University of Minnesota)
Computational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: The paper proposes a framework called Tree-Based Invariant Kernels (TBIK), which maintains fully deterministic inference across different tensor parallel (TP) scales, addressing the problem of probabilistic inconsistency between training and inference.
DexMachina: Functional Retargeting for Bimanual Dexterous Manipulation
Zhao Mandi (Stanford University), Shuran Song (NVIDIA)
Robotic IntelligenceReinforcement LearningVision-Language-Action ModelAuto EncoderVideoSequential
🎯 What it does: Propose a functional retargeting method, which uses single-segment human hand-object demonstrations to train dual-robot hand policies, enabling them to track object states in the demonstrations.
DF-ExpEnse: Diffusion Filtered Exploration for Sample Efficient Finetuning
Calvin Luo (Stanford University), Shuran Song (Stanford University)
Robotic IntelligenceTransformerReinforcement LearningDiffusion modelMultimodality
🎯 What it does: Propose the DF-ExpEnse technique, which uses diffusion strategies to generate multimodal candidate actions, and evaluates exploration interest through a critic ensemble, thus significantly improving the sample efficiency of fine-tuning pre-trained generative control policies during online experience collection.
DF-LoGiT: Data-Free Logic-Gated Backdoor Attacks in Vision Transformers
Xiaozuo Shen (University of Arizona), Hongyi Wu (University of Arizona)
Adversarial AttackTransformerPrompt EngineeringContrastive LearningImage
🎯 What it does: Proposes a fully data-agnostic logic-gated backdoor attack (DF-LoGiT), which injects backdoors by directly modifying the weights of ViT checkpoints, without requiring training data or fine-tuning.
DFlash: Block Diffusion for Flash Speculative Decoding
Jian Chen (UC San Diego), Zhijian Liu (UC San Diego)
GenerationComputational EfficiencyAI Code AssistantTransformerLarge Language ModelDiffusion modelContrastive LearningTextRetrieval-Augmented Generation
🎯 What it does: Proposes DFlash, a framework that uses a lightweight block diffusion model for speculative decoding, leveraging the hidden features of the target large model for conditioning.
DFSAttn: Dynamic Fine-grained Sparse Attention for Efficient Video Generation
Jie Hu (Peking University), Kun Yuan (Peking University)
GenerationComputational EfficiencyTransformerDiffusion modelScore-based ModelVideoText
🎯 What it does: Designed and implemented DFSAttn, an unsupervised dynamic fine-grained sparse attention framework for accelerating the generation of video diffusion models (DiTs);
DGG-HMR: Multi-Person Human Mesh Recovery with Depth-Guided Geometric Anchoring
Yanjie Li (Shaanxi Normal University), Shigang Liu (Shaanxi Normal University)
Pose EstimationDepth EstimationGraph Neural NetworkTransformerDiffusion modelAuto EncoderContrastive LearningImageMesh
🎯 What it does: Recover 3D meshes of multiple people from a single image, addressing the problem of depth uncertainty.
dgMARK: Decoding-Guided Watermarking for Diffusion Language Models
Pyo Min Hong (Hongik University), Albert No (Yonsei University)
GenerationData SynthesisSafty and PrivacyTransformerLarge Language ModelPrompt EngineeringDiffusion modelScore-based ModelText
🎯 What it does: Propose dgMARK, a method that embeds parity watermarks by guiding the decoding order of discrete diffusion language models;
DGS-Net: Distillation-Guided Gradient Surgery for CLIP Fine-Tuning in AI-Generated Image Detection
Jiazhen Yan (Nanjing University of Information Science and Technology), Zhangjie Fu (Nanjing University of Information Science and Technology)
ClassificationAnomaly DetectionKnowledge DistillationRepresentation LearningTransformerSupervised Fine-TuningVision Language ModelDiffusion modelContrastive LearningImageTextMultimodality
🎯 What it does: Propose a DGS-Net framework based on knowledge distillation guided gradient refinement for fine-tuning AI-generated image detection without losing the pre-trained prior of CLIP.
Di-BiLPS: Denoising induced Bidirectional Latent-PDE-Solver under Sparse Observations
Zhonghao Li (Harbin Institute of Technology), Zhang Qian
OptimizationTransformerDiffusion modelScore-based ModelAuto EncoderContrastive LearningGraphTabularBenchmarkPhysics Related
🎯 What it does: Proposed a unified neural network framework called Di-BiLPS for simultaneously solving forward and inverse problems of PDEs under extremely sparse observations.
Diagnosing and Correcting Concept Omission in Multimodal Diffusion Transformers
Kanghyun Baek (Seoul National University), Sungroh Yoon (Seoul National University)
GenerationData SynthesisTransformerPrompt EngineeringDiffusion modelFlow-based ModelImageTextMultimodality
🎯 What it does: Propose an intervention method called OSI based on concept missing signals, specifically addressing the concept missing problem in text-image generation with the MM-DiT model.
Diagnosing Multi-step Reasoning Failures in Black-box LLMs via Stepwise Confidence Attribution
Xiaoou Liu (Arizona State University), Hua Wei (Arizona State University)
Explainability and InterpretabilityTransformerLarge Language ModelContrastive LearningTextChain-of-Thought
🎯 What it does: Propose a step-wise confidence attribution (SCA) framework for black-box large language models (LLMs) to diagnose erroneous steps in multi-step reasoning processes.
Diagnosing the Reliability of LLM-as-a-Judge via Item Response Theory
Junhyuk Choi (Chung-Ang University), Bugeun Kim (Chung-Ang University)
Explainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelPrompt EngineeringImageText
🎯 What it does: Propose a two-stage diagnostic framework based on IRT to evaluate the reliability of LLM-as-a-Judge.