ICML 2025 Papers — Page 7
International Conference on Machine Learning · 3257 papers
CoPINN: Cognitive Physics-Informed Neural Networks
Siyuan Duan (Sichuan University), Yuan Sun (Sichuan University)
OptimizationComputational EfficiencyTabularPhysics Related
🎯 What it does: This paper proposes Cognitive Physics-Informed Neural Networks (CoPINN), which addresses the Unbalanced Prediction Problem in traditional PINNs through an adaptive training strategy that progresses from easy to difficult.
Core Context Aware Transformers for Long Context Language Modeling
Yaofo Chen (South China University of Technology), Mingkui Tan (South China University of Technology)
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: This paper proposes a Core Context-Aware Attention (CCA-Attention) for efficient long-text language modeling.
Core Knowledge Deficits in Multi-Modal Language Models
Yijiang Li (University of California San Diego), Hokin Deng (Carnegie Mellon University)
Large Language ModelPrompt EngineeringTextMultimodalityBenchmark
🎯 What it does: This study investigates the deficiencies of multimodal large language models in core knowledge (basic cognitive abilities) and proposes the CoreCognition benchmark and Concept Hacking controlled experiments for evaluation.
CoreMatching: A Co-adaptive Sparse Inference Framework with Token and Neuron Pruning for Comprehensive Acceleration of Vision-Language Models
Qinsi Wang (Duke University), Yiran Chen (Duke University)
OptimizationComputational EfficiencyTransformerVision Language ModelMultimodality
🎯 What it does: This paper proposes a framework called CoreMatching for joint adaptive sparse inference, which accelerates the inference of large-scale vision-language models by simultaneously sparsifying visual tokens and neurons.
Correlated Errors in Large Language Models
Elliot Myunghoon Kim (Cornell University), Nikhil Garg (Cornell University)
TransformerLarge Language ModelText
🎯 What it does: This paper systematically evaluates the error correlation of over 350 large language models (LLMs) and explores the implications of this correlation for LLM evaluation, recruitment, and multi-agent systems.
Correlation Clustering Beyond the Pivot Algorithm
Soheil Behnezhad (Northeastern University), Weiyun ma
OptimizationGraph Neural NetworkGraph
🎯 What it does: The MODIFIEDPIVOT algorithm is proposed, which improves clustering based on PIVOT through local vertex movement and random sampling, and proves that it can maintain a 3-ε approximation in a fully dynamic environment in polynomial logarithmic time.
COSDA: Counterfactual-based Susceptibility Risk Framework for Open-Set Domain Adaptation
Wenxu Wang (Ocean University of China), Nevin L. Zhang (Hong Kong University of Science and Technology)
Domain AdaptationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: A susceptibility risk framework based on causal counterfactuals, COSDA, is proposed to address the issues of domain shift and unknown category recognition in open set domain adaptation (OSDA).
CoSER: Coordinating LLM-Based Persona Simulation of Established Roles
Xintao Wang (Fudan University), Yanghua Xiao (Fudan University)
TransformerLarge Language ModelSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: This paper presents the COSER dataset, the corresponding LLaMA-3.1 base models (COSER-8B and COSER-70B), and a training and evaluation framework based on Given Contextual Abduction (GCA) to achieve realistic role-playing language agents for existing literary characters.
Cost-efficient Collaboration between On-device and Cloud Language Models
Avanika Narayan (Stanford University), Christopher Re
OptimizationComputational EfficiencyTransformerLarge Language ModelTextFinance Related
🎯 What it does: Designed and evaluated two local-cloud language model collaboration protocols (MINION and MINION S), enabling large models in the cloud and small models locally to collaboratively complete long text reasoning tasks.
CostFilter-AD: Enhancing Anomaly Detection through Matching Cost Filtering
Zhe Zhang (Northeastern University), Xiatian Zhu (University of Surrey)
Anomaly DetectionConvolutional Neural NetworkImage
🎯 What it does: A general matching cost filtering plugin, CostFilter-AD, is proposed to enhance the detection and localization performance of unsupervised anomaly detection (UAD) models in both multi-class and single-class scenarios.
Counterfactual Contrastive Learning with Normalizing Flows for Robust Treatment Effect Estimation
Jiaxuan Zhang (Shanxi University), Jiye Liang (Shanxi University)
OptimizationRepresentation LearningFlow-based ModelContrastive LearningTabular
🎯 What it does: This paper proposes a differential isomorphic counterfactual adversarial learning framework (FCCL) based on flow models, aimed at accurately estimating individual treatment effects from observational data.
Counterfactual Effect Decomposition in Multi-Agent Sequential Decision Making
Stelios Triantafyllou (Max Planck Institute for Software Systems), Goran Radanovic (Max Planck Institute for Software Systems)
Explainability and InterpretabilityReinforcement LearningAgentic AISequential
🎯 What it does: A method for counterfactual effect decomposition in multi-agent Markov decision processes is proposed, which can break down the impact of a single agent's action on the final outcome into contributions to subsequent agent behaviors and contributions to environmental state transitions.
Counterfactual Graphical Models: Constraints and Inference
Juan D. Correa (Universidad Autonoma de Manizales), Elias Bareinboim (Columbia University)
Graph Neural NetworkGraph
🎯 What it does: This paper addresses the counterfactual model in causal graphs, proposing an efficient graphical construction (Ancestor Multi-World Network, AMWN) and a complete counterfactual reasoning framework (ctf-calculus), achieving a complete algorithm for reading counterfactual independence from causal graphs.
Counterfactual Voting Adjustment for Quality Assessment and Fairer Voting in Online Platforms with Helpfulness Evaluation
Chang Liu (University of Illinois Chicago), Moontae Lee (University of Illinois Chicago)
Recommendation SystemReinforcement LearningTabular
🎯 What it does: The Counterfactual Voting Adjustment (CVA) framework is proposed, utilizing causal inference and voting trajectories to correct voting biases on online platforms, thereby more fairly assessing and ranking content quality.
Counting atoms faster: policy-based nuclear magnetic resonance pulse sequencing for atomic abundance measurement
Rohan Shenoy (Massachusetts Institute of Technology), Elsa Olivetti
OptimizationReinforcement LearningTime SeriesMagnetic Resonance ImagingPhysics Related
🎯 What it does: Using reinforcement learning to learn pulse sequences on low-field NMR equipment for rapid atomic abundance measurement.
Counting in Small Transformers: The Delicate Interplay between Attention and Feed-Forward Layers
Freya Behrens (Ecole Polytechnique Federale de Lausanne), Lenka Zdeborova (Ecole Polytechnique Federale de Lausanne)
TransformerSequential
🎯 What it does: This study investigates the implementation mechanism of a single-layer Transformer model in counting tasks, comparing the roles of attention and feedforward layers under different architectures and hyperparameters.
Cover learning for large-scale topology representation
Luis Scoccola (Centre de Recherches Mathématiques et Institut des sciences mathématiques), Heather A. Harrington (Max Planck Institute of Molecular Cell Biology and Genetics)
OptimizationRepresentation LearningGraph Neural NetworkMesh
🎯 What it does: Proposed and implemented a topology-friendly rational complex (ShapeDiscover) by optimizing the coverage of geometric datasets.
Covered Forest: Fine-grained generalization analysis of graph neural networks
Antonis Vasileiou (RWTH Aachen University), Christopher Morris
Graph Neural NetworkGraphBiomedical Data
🎯 What it does: This study investigates the generalization performance of Message Passing Neural Networks (MPNN) and proposes a fine-grained generalization analysis framework based on pseudo-metrics such as tree distance and forest distance, along with corresponding theoretical upper bounds.
Cowpox: Towards the Immunity of VLM-based Multi-Agent Systems
YUTONG WU, Tianwei Zhang
OptimizationAdversarial AttackTransformerVision Language ModelMultimodalityRetrieval-Augmented Generation
🎯 What it does: A distributed immune mechanism called COWPOX has been designed and validated, capable of resisting infection-based jailbreak attacks and recovering infected agents in multi-agent visual language model systems.
CPCF: A Cross-Prompt Contrastive Framework for Referring Multimodal Large Language Models
Lanyun Zhu (Singapore University of Technology and Design), Jun Liu (Lancaster University)
RetrievalComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringContrastive LearningImageTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: A CPCF framework is proposed, which enhances the accuracy of responses from directional multimodal LLMs by comparing input visual prompts with automatically generated contrast prompts.
Cradle: Empowering Foundation Agents towards General Computer Control
Weihao Tan (Nanyang Technological University), Zongqing Lu (Peking University)
Robotic IntelligenceTransformerLarge Language ModelAgentic AIVideoMultimodality
🎯 What it does: The CRADLE framework has been developed, enabling foundational agents based on Large Multimodal Models (LMM) to complete interactions and long-term tasks in any software or game using only screenshot inputs and keyboard/mouse outputs.
Craftium: Bridging Flexibility and Efficiency for Rich 3D Single- and Multi-Agent Environments
Mikel Malagón (University of the Basque Country), Jose A. Lozano (Basque Center for Applied Mathematics)
Computational EfficiencyRobotic IntelligenceReinforcement Learning from Human Feedback
🎯 What it does: This paper proposes and implements Craftium, a 3D environment framework based on the open-source voxel engine Luanti, supporting single/multi-agent, customization, procedural generation, and large-scale open worlds, while being compatible with Gymnasium and PettingZoo interfaces.
CRANE: Reasoning with constrained LLM generation
Debangshu Banerjee (University of Illinois), Gagandeep Singh (University of Illinois)
GenerationOptimizationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes a reasoning-enhanced decoding method based on constrained generation called CRANE. Theoretical analysis shows that strict constraints weaken the reasoning ability of LLMs, and augmented grammar rules are introduced to restore this reasoning capability. An adaptive switching mechanism between constrained and unconstrained generation is then implemented to improve grammatical and semantic correctness.
Critical Tokens Matter: Token-Level Contrastive Estimation Enhances LLM’s Reasoning Capability
Zicheng Lin (Tsinghua University), Zhaopeng Tu (Tencent)
TransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextChain-of-Thought
🎯 What it does: Proposes the concept of 'keywords' and identifies key erroneous words in mathematical reasoning processes using rollout sampling and contrastive estimation, enhancing reasoning accuracy by penalizing these words in DPO.
Cross-City Latent Space Alignment for Consistency Region Embedding
Meng Chen (Shandong University), Weiming Huang (Lund University)
Domain AdaptationRepresentation LearningGraph Neural NetworkGraphTabular
🎯 What it does: This paper proposes the CoRE method, which achieves self-supervised learning of regional embeddings and latent space alignment across cities, enabling direct transfer of predictors in the absence of target city labels.
Cross-environment Cooperation Enables Zero-shot Multi-agent Coordination
Kunal Jha (University of Washington), Natasha Jaques (University of Washington)
Robotic IntelligenceReinforcement Learning
🎯 What it does: This paper studies a Cross-Environment Cooperation (CEC) training framework that enables agents to self-play in a large number of procedurally generated diverse environments, achieving zero-shot multi-agent collaboration.
Cross-Modal Alignment via Variational Copula Modelling
Feng Wu (University of Hong Kong), Lequan Yu (University of Hong Kong)
Data SynthesisAnomaly DetectionOptimizationRecurrent Neural NetworkMultimodalityBiomedical DataElectronic Health Records
🎯 What it does: A cross-modal alignment framework (CM2) based on a variational Copula model is proposed, which learns the marginal distributions of each modality through Gaussian mixture models, and then uses Copula to capture complex dependencies between modalities, achieving joint distribution modeling and missing modality completion.
Cross-regularization: Adaptive Model Complexity through Validation Gradients
Carlos Stein Brito (NightCity Labs)
OptimizationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a framework called Cross-regularization, which continuously updates the regularization parameters (such as regularization strength, noise scale, data augmentation magnitude, etc.) during the training process using validation gradients, achieving adaptive control of model complexity.
CROW: Eliminating Backdoors from Large Language Models via Internal Consistency Regularization
Nay Myat Min (Singapore Management University), Jun Sun (Singapore Management University)
OptimizationAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: CROW is proposed, a method to eliminate LLM backdoors by applying consistency regularization to the hidden states of internal layers during the fine-tuning process.
CSG-ODE: ControlSynth Graph ODE For Modeling Complex Evolution of Dynamic Graphs
Zhiqiang Wang (Shanxi University), Jianqing Liang (Shanxi University)
Graph Neural NetworkAuto EncoderGraphOrdinary Differential Equation
🎯 What it does: A ControlSynth Graph ODE (CSG-ODE) framework based on VAE+GNN+ODE is proposed, which integrates information propagation weights to fuse latent space correlations, capturing time-varying relationships and nonlinear state evolution among dynamic graph nodes, and further introduces a stable version of SCSG-ODE with antisymmetric weights;
CSTrack: Enhancing RGB-X Tracking via Compact Spatiotemporal Features
Xiaokun Feng (University of Chinese Academy of Sciences), Kaiqi Huang (University of Chinese Academy of Sciences)
Object TrackingTransformerMultimodality
🎯 What it does: A RGB-X visual tracker named CSTrack is proposed, which simplifies the network structure and enhances tracking robustness by merging RGB and auxiliary modalities (depth, thermal, event) into compact spatiotemporal features.
CSV-Occ: Fusing Multi-frame Alignment for Occupancy Prediction with Temporal Cross State Space Model and Central Voting Mechanism
Ziming Zhu (East China University of Science and Technology), Lihua Sun (East China University of Science and Technology)
SegmentationAutonomous DrivingPoint Cloud
🎯 What it does: The CSV-Occ framework is proposed, which is based on camera 3D semantic occupancy prediction. It utilizes a cross-state space model to fuse multi-frame information and enhances internal occupancy accuracy through a center voting mechanism.
CTBench: A Library and Benchmark for Certified Training
Yuhao Mao (ETH Zürich), Martin Vechev (ETH Zürich)
OptimizationAdversarial AttackConvolutional Neural NetworkImageBenchmark
🎯 What it does: CTBENCH is proposed - a unified library and high-quality benchmark for fair and systematic evaluation and comparison of deterministic provably robust training methods under the L∞ norm.
CtrlSynth: Controllable Image Text Synthesis for Data-Efficient Multimodal Learning
Qingqing Cao (Apple), Sachin Mehta (Apple)
GenerationData SynthesisRetrievalTransformerLarge Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: A controllable image-text synthesis framework CtrlSynth is proposed, which generates diverse image-text data using visual tags and foundational models.
CUPS: Improving Human Pose-Shape Estimators with Conformalized Deep Uncertainty
Harry Zhang (Massachusetts Institute of Technology), Luca Carlone (Massachusetts Institute of Technology)
Pose EstimationTransformerVideo
🎯 What it does: An end-to-end method named CUPS is proposed, capable of recovering 3D human pose and shape from monocular RGB videos, while learning a deep uncertainty function to quantify the uncertainty of predictions and achieve reliable confidence sets.
Curriculum Learning for Biological Sequence Prediction: The Case of De Novo Peptide Sequencing
Xiang Zhang (Fudan University), Siqi Sun (Fudan University)
GenerationProtein Structure PredictionTransformerSequentialBiomedical Data
🎯 What it does: A de novo peptide sequencing model called RefineNovo based on a non-autoregressive Transformer has been constructed. It utilizes CTC loss along with curriculum learning and iterative refinement techniques, allowing the model to learn progressively from simple to complex and to self-improve during inference.
Curse of High Dimensionality Issue in Transformer for Long Context Modeling
Shuhai Zhang (South China University of Technology), Mingkui Tan (South China University of Technology)
Computational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Proposes Dynamic Group Attention (DGA), which reduces redundant attention computations in long sequences by aggregating unimportant tokens;
CursorCore: Assist Programming through Aligning Anything
Hao Jiang (University of Science and Technology of China), Shijin Wang (iFLYTEK Co., Ltd.)
AI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: A unified programming assistance framework called Assistant-Conversation has been designed, a comprehensive evaluation benchmark APEval covering various information combinations has been proposed, and a Programming-Instruct data generation pipeline that does not require manual annotation has been developed. Using this data, the CursorCore series of models have been trained and released, achieving leading performance in multilingual programming assistance tasks.
Curvature Enhanced Data Augmentation for Regression
Ilya Kaufman (Ben-Gurion University of the Negev), Omri Azencot (Ben-Gurion University of the Negev)
Data SynthesisImageTabularTime Series
🎯 What it does: A regression data augmentation method based on manifold second-order curvature information, CEMS, is proposed to generate new input-output pairs during training.
Curvature-aware Graph Attention for PDEs on Manifolds
Yunfeng Liao (Harbin Institute of Technology), Xiucheng Li (Harbin Institute of Technology)
OptimizationComputational EfficiencyGraph Neural NetworkTransformerGraphTime SeriesOrdinary Differential Equation
🎯 What it does: This paper proposes a curvature-aware graph attention mechanism for efficiently solving time-dependent partial differential equations (such as heat diffusion, p-Laplace diffusion, and wave equations) on discrete manifolds.
CurvGAD: Leveraging Curvature for Enhanced Graph Anomaly Detection
Karish Grover (Carnegie Mellon University), Christos Faloutsos (Carnegie Mellon University)
Anomaly DetectionGraph Neural NetworkAuto EncoderGraph
🎯 What it does: A curvature-based mixed curvature graph autoencoder (CurvGAD) is proposed, which detects graph structure, attributes, and geometric anomalies through two parallel pipelines: curvature equivariant and curvature invariant.
Customizing the Inductive Biases of Softmax Attention using Structured Matrices
Yilun Kuang (New York University), Andrew Gordon Wilson (New York University)
TransformerLarge Language ModelTextTime Series
🎯 What it does: This paper addresses the low-rank bottleneck and the lack of distance-related computational bias by rewriting the Softmax attention through the integration of Block Tensor-Train with Multi-Level Low-Rank matrix embedded attention score functions.
Cut out and Replay: A Simple yet Versatile Strategy for Multi-Label Online Continual Learning
Xinrui Wang (Nanjing University of Aeronautics and Astronautics), Songcan Chen (Nanjing University of Aeronautics and Astronautics)
ClassificationObject DetectionKnowledge DistillationTransformerImage
🎯 What it does: A cutting and experience replay strategy named CUTER is proposed to address catastrophic forgetting, missing labels, and class imbalance issues in multi-label online continual learning.
CVE-Bench: A Benchmark for AI Agents’ Ability to Exploit Real-World Web Application Vulnerabilities
Yuxuan Zhu (University of Illinois), Daniel Kang (University of Illinois)
TransformerLarge Language ModelAgentic AIBenchmark
🎯 What it does: CVE-Bench was constructed, a security benchmark based on 40 high-risk CVEs from real-world web applications, and implemented a sandbox environment, reference exploits, and an automated evaluation system; experiments were conducted on three LLM agents (Cy-Agent, T-Agent, AutoGPT) in zero-day and one-day scenarios, reporting their success rates and costs.
D-Fusion: Direct Preference Optimization for Aligning Diffusion Models with Visually Consistent Samples
Zijing Hu (Zhejiang University), Kun Kuang (Zhejiang University)
GenerationOptimizationReinforcement Learning from Human FeedbackTransformerReinforcement LearningDiffusion modelImage
🎯 What it does: Proposes the D-Fusion method, which utilizes self-attention fusion to generate visually consistent samples that are similar to the original image but well-aligned with the text, and pairs them with the original samples for direct preference optimization (DPO) fine-tuning of the diffusion model, thereby improving the alignment quality between text and images.
DA-KD: Difficulty-Aware Knowledge Distillation for Efficient Large Language Models
Changyi He (Beihang University), Xianglong Liu (Beihang University)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: The DA-KD framework is proposed for efficient distillation of large language models.
DAMA: Data- and Model-aware Alignment of Multi-modal LLMs
Jinda Lu (University of Science and Technology of China), Xiangnan He (University of Science and Technology of China)
OptimizationTransformerLarge Language ModelReinforcement LearningMultimodalityBenchmark
🎯 What it does: The DAMA method is proposed for direct preference optimization of multimodal large language models, dynamically adjusting β to balance the learning of easily distinguishable and difficult-to-distinguish samples.
DANCE: Dual Unbiased Expansion with Group-acquired Alignment for Out-of-distribution Graph Fairness Learning
Yifan Wang (University of International Business and Economics), Xiao Luo (University of California)
Domain AdaptationGraph Neural NetworkContrastive LearningGraphFinance Related
🎯 What it does: This paper proposes the DANCE framework, specifically designed to address the fairness issues of graph neural networks in out-of-distribution scenarios;
Data Mixing Optimization for Supervised Fine-Tuning of Large Language Models
Yuan Li (Carnegie Mellon University), Eric P. Xing
OptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A data mixing optimization method for supervised fine-tuning (SFT) of large language models is proposed, treating data mixing as an optimization problem and parameterizing the validation loss to find optimal domain weights.
Data-driven Design of Randomized Control Trials with Guaranteed Treatment Effects
Santiago Cortes-Gomez (Carnegie Mellon University), Bryan Wilder (Carnegie Mellon University)
🎯 What it does: A two-stage randomized controlled trial (RCT) design is proposed, along with a top-k strategy based on sample splitting, achieving a high confidence lower bound for at least one treatment option while maintaining low adaptability.
Data-Driven Selection of Instrumental Variables for Additive Nonlinear, Constant Effects Models
Xichen Guo (Beijing Technology and Business University), Zhi Geng (Beijing Technology and Business University)
Tabular
🎯 What it does: A method for selecting an effective set of instrumental variables (IV Set) in the Additive Nonlinear Constant Effects Model (ANICE) is proposed, along with testable CAT conditions and corresponding algorithms.
Data-Juicer Sandbox: A Feedback-Driven Suite for Multimodal Data-Model Co-development
Daoyuan Chen (Alibaba Group), Jingren Zhou (Alibaba Group)
OptimizationData-Centric LearningTransformerLarge Language ModelVision Language ModelImageVideoTextMultimodality
🎯 What it does: Constructed Data-Juicer Sandbox, an open-source experimental platform that supports data-model collaborative optimization.
DataDecide: How to Predict Best Pretraining Data with Small Experiments
Ian Magnusson (Allen Institute for AI), Jesse Dodge (Allen Institute for AI)
Large Language ModelText
🎯 What it does: The DATADECIDE suite has been constructed and made public, including 25 types of pre-trained data, 14 model sizes, and 3 random seeds, to comprehensively evaluate the predictive ability of small-scale experiments in large-scale model data selection.
Dataflow-Guided Neuro-Symbolic Language Models for Type Inference
Gen Li (Huazhong University of Science and Technology), Zheng Wang (University of Leeds)
Explainability and InterpretabilityAI Code AssistantLarge Language ModelText
🎯 What it does: The NESTER framework is proposed, which combines neural networks with symbolic reasoning for type inference in dynamic languages.
David and Goliath: Small One-step Model Beats Large Diffusion with Score Post-training
Weijian Luo (Xiaohongshu Inc), Zhengyang Geng (Carnegie Mellon University)
GenerationComputational EfficiencyReinforcement Learning from Human FeedbackReinforcement LearningDiffusion modelScore-based ModelImageText
🎯 What it does: A post-training method named Diff-Instruct* (DI*) is proposed to align single-step text-to-image generation models with human preferences through reinforcement learning from human feedback (RLHF), while maintaining the diversity and realism of high-resolution images.
DCBM: Data-Efficient Visual Concept Bottleneck Models
Katharina Prasse (University of Mannheim), Margret Keuper (Max Planck Institute for Informatics)
Object DetectionSegmentationExplainability and InterpretabilityData-Centric LearningContrastive LearningImage
🎯 What it does: This paper proposes Data-Efficient Visual Concept Bottleneck Models (DCBM), which extracts visual concepts from images using segmentation/detection base models, enabling the construction of interpretable CBMs with very few training samples.
DCTdiff: Intriguing Properties of Image Generative Modeling in the DCT Space
Mang Ning (Utrecht University), Itir Onal Ertugrul
GenerationCompressionComputational EfficiencyTransformerDiffusion modelImage
🎯 What it does: This paper proposes DCTdiff, which transfers diffusion models to the DCT frequency domain for image generation, improving quality and training efficiency.
De-AntiFake: Rethinking the Protective Perturbations Against Voice Cloning Attacks
Wei Fan (University of Science and Technology of China), Nenghai Yu (University of Science and Technology of China)
Safty and PrivacyAdversarial AttackDiffusion modelScore-based ModelAudio
🎯 What it does: This paper systematically evaluates the defense effectiveness of existing protective perturbations against zero-shot voice cloning attacks and proposes a two-stage 'Purification-Refinement' adversarial purification method that can effectively disrupt protective perturbations.
De-coupled NeuroGF for Shortest Path Distance Approximations on Large Terrain Graphs
Samantha Chen (University of California - San Diego), Yusu Wang (University of California San Diego)
Graph Neural NetworkGraph
🎯 What it does: A separated training NeuroGF framework is proposed for quickly approximating shortest path distance queries on large-scale terrain maps.
De-mark: Watermark Removal in Large Language Models
Ruibo Chen (University of Maryland), Heng Huang (University of Maryland)
TransformerLarge Language ModelText
🎯 What it does: Proposes the DE-MARK framework, which utilizes random selection probing techniques to reverse engineer, remove, and utilize n-gram watermarks in large language models.
DEALing with Image Reconstruction: Deep Attentive Least Squares
Mehrsa Pourya (École Polytechnique Fédérale de Lausanne), Sebastian Neumayer (Technische Universität Chemnitz)
RestorationSuper ResolutionConvolutional Neural NetworkImageMagnetic Resonance Imaging
🎯 What it does: A deep attention-based least squares reconstruction framework (DEAL) is proposed, which iteratively solves quadratic optimization and uses learnable filters and attention weights for adaptive regularization.
Decision Making under the Exponential Family: Distributionally Robust Optimisation with Bayesian Ambiguity Sets
Charita Dellaporta (University College London), Theodoros Damoulas (University of Warwick)
OptimizationTabularTime SeriesFinance Related
🎯 What it does: This paper proposes a distributionally robust optimization framework based on Bayesian posterior—DRO-BAS, which utilizes two sets of uncertainties formed by posterior inference (BASPP based on posterior prediction and BASPE based on posterior expectation), and provides the corresponding strong dual form under extreme limit theorems, thus transforming the worst-case risk problem into a single-stage stochastic programming problem.
Decision Mixer: Integrating Long-term and Local Dependencies via Dynamic Token Selection for Decision-Making
Hongling Zheng (Wuhan University), Dacheng Tao (Nanyang Technological University)
TransformerReinforcement LearningMixture of ExpertsSequential
🎯 What it does: This paper proposes Decision Mixer (DM), a transformer structure that dynamically selects key tokens in offline reinforcement learning and feeds them into the attention layer to balance long-term dependencies and local Markov properties.
Decision Theoretic Foundations for Conformal Prediction: Optimal Uncertainty Quantification for Risk-Averse Agents
Shayan Kiyani (University of Pennsylvania), Hamed Hassani (University of Pennsylvania)
Recommendation SystemReinforcement LearningImageTabular
🎯 What it does: This paper establishes a decision-theoretic framework based on predictive sets under risk-averse decision-making and proposes the Risk-Averse Calibration (RAC) algorithm.
Decision-aware Training of Spatiotemporal Forecasting Models to Select a Top-K Subset of Sites for Intervention
Kyle Heuton (Tufts University), Michael C Hughes
Recommendation SystemOptimizationReinforcement LearningTime Series
🎯 What it does: A spatiotemporal prediction model training and ranking method based on the Best Possible Rate (BPR) metric is proposed for selecting the optimal K sites for intervention under resource constraints.
Decoding Rewards in Competitive Games: Inverse Game Theory with Entropy Regularization
Junyi Liao (Duke University), Vahid Tarokh (Duke University)
Reinforcement LearningTabular
🎯 What it does: A unified inverse game theory framework is proposed to recover unknown reward functions based on observed strategies and actions in two-player zero-sum matrix games and Markov games (both using entropy regularization).
Decomposition of Graphic Design with Unified Multimodal Model
Hui Nie (University of Chinese Academy of Sciences), Xinglong Wu
RestorationGenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningGenerative Adversarial NetworkImageMultimodality
🎯 What it does: This paper proposes the Layered Deconstruction of Graphic Design (LDGD) task and designs a unified multimodal model DeaM, which can decompose composite graphic designs into ordered RGB-A layers and corresponding metadata.
Decoupled SGDA for Games with Intermittent Strategy Communication
Ali Zindari (CISPA Helmholtz Center for Information Security), Sebastian U Stich
OptimizationFederated LearningReinforcement LearningGenerative Adversarial NetworkImage
🎯 What it does: A game optimization algorithm for intermittent strategy communication, Decoupled SGDA, is proposed, allowing players to perform local updates using only outdated opponent strategies, significantly reducing communication costs.
Deep Bayesian Filter for Bayes-Faithful Data Assimilation
Yuta Tarumi (Preferred Networks), Shin-ichi Maeda (Preferred Networks)
Recurrent Neural NetworkAuto EncoderTime Series
🎯 What it does: A deep Bayesian filtering (DBF) method is proposed for data assimilation in nonlinear state space models.
Deep Electromagnetic Structure Design Under Limited Evaluation Budgets
Shijian Zheng (South China University of Technology), Mingkui Tan (South China University of Technology)
OptimizationConvolutional Neural NetworkReinforcement LearningTabular
🎯 What it does: A quadtree-based evolutionary search method is proposed, which can quickly find high-quality electromagnetic structure (EMS) design solutions under a limited evaluation budget.
Deep Fuzzy Multi-view Learning for Reliable Classification
Siyuan Duan (Sichuan University), Peng Hu (Sichuan University)
ClassificationMultimodality
🎯 What it does: A deep multi-view learning framework based on fuzzy set theory, called FUML, is designed to achieve reliable classification and accurate uncertainty estimation in the presence of conflicting views.
Deep Linear Network Training Dynamics from Random Initialization: Data, Width, Depth, and Hyperparameter Transfer
Blake Bordelon (Harvard University), Cengiz Pehlevan (Harvard University)
Tabular
🎯 What it does: This study investigates the gradient descent dynamics of deep linear networks starting from random initialization, systematically considering factors such as network width, depth, data volume, and feature learning intensity.
Deep Neural Cellular Potts Models
Koen Minartz (Eindhoven University of Technology), Vlado Menkovski (Eindhoven University of Technology)
GenerationData SynthesisConvolutional Neural NetworkReinforcement LearningImageBiomedical Data
🎯 What it does: Proposes NeuralCPM, which utilizes neural networks to parameterize the energy function of the cellular Potts model, allowing for direct learning of cellular dynamics from observational data.
Deep Principal Support Vector Machines for Nonlinear Sufficient Dimension Reduction
Yinfeng Chen, Rui Qiu (Peking University)
Supervised Fine-TuningTabular
🎯 What it does: A nonlinear sufficient dimension reduction method based on deep principal support vector machines is proposed.
Deep Reinforcement Learning from Hierarchical Preference Design
Alexander Bukharin (NVIDIA), Tuo Zhao (Georgia Institute of Technology)
Robotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningTabular
🎯 What it does: This paper proposes the HERON (Hierarchical Preference-based Reinforcement Learning) framework, which constructs a reward model by comparing hierarchical preferences of trajectories using the hierarchical relationships of feedback signals or sparse reward scenarios, thereby enhancing the performance of reinforcement learning.
Deep Ridgelet Transform and Unified Universality Theorem for Deep and Shallow Joint-Group-Equivariant Machines
Sho Sonoda (RIKEN AIP), Masahiro Ikeda (University of Osaka)
🎯 What it does: This paper presents a constructive universal approximation theorem for deep and shallow networks targeting joint-group-equivariant feature maps, along with the corresponding Ridgelet transform.
Deep Streaming View Clustering
Honglin Yuan (Southwest University of Science and Technology), Zhenwen Ren
Representation LearningAuto EncoderImage
🎯 What it does: A deep learning framework for view stream clustering (SVC) called DSVC is proposed, which can online update the model when new views arrive and overcome the problem of concept drift.
Deep Sturm–Liouville: From Sample-Based to 1D Regularization with Learnable Orthogonal Basis Functions
David Vigouroux (IRT Saint Exupery), Victor Boutin (CerCo CNRS Université de Toulouse)
ImageTabularOrdinary Differential Equation
🎯 What it does: This paper proposes the Deep Sturm-Liouville (DSL) model, which introduces Sturm-Liouville theory into deep learning to achieve one-dimensional continuous regularization along input space trajectories, enhancing the model's generalization ability.
Deep Unsupervised Hashing via External Guidance
Qihong Song (Sichuan University), Peng Hu (Xidian University)
RetrievalContrastive LearningImageText
🎯 What it does: A deep hashing method that utilizes external textual knowledge as semantic guidance in unsupervised hashing is proposed.
DeepCrossAttention: Supercharging Transformer Residual Connections
Mike Heddes (University of California), Vahab Mirrokni (Google Research)
TransformerImageText
🎯 What it does: This paper proposes the DeepCrossAttention (DCA) framework, which introduces learnable, input-dependent residual weights and deep cross attention in the Transformer to dynamically fuse outputs from different layers, enhancing information flow and model performance.
DeepLayout: Learning Neural Representations of Circuit Placement Layout
Yuxiang Zhao (Peking University), Yibo Lin (Peking University)
OptimizationRepresentation LearningGraph Neural NetworkTransformerAuto EncoderGraph
🎯 What it does: Proposes the DeepLayout framework, which utilizes a self-supervised mask autoencoder for representation learning of backend circuit layouts and can be transferred to tasks such as congestion prediction and post-routing wire length estimation.
DEFAME: Dynamic Evidence-based FAct-checking with Multimodal Experts
Tobias Braun (Technical University of Darmstadt), Anna Rohrbach (Technical University of Darmstadt)
RetrievalExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringImageTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Designed DEFAME, a zero-shot dynamic evidence retrieval and reasoning framework based on multimodal LLM, achieving end-to-end multimodal fact-checking.
Defending LVLMs Against Vision Attacks Through Partial-Perception Supervision
Qi Zhou (Zhejiang University), Qing Guo
Adversarial AttackLarge Language ModelVision Language ModelImage
🎯 What it does: A black-box, untrained defense method called DPS is proposed, which utilizes the local perception results of images to supervise the responses of the complete image, thereby resisting visual attacks.
DeFoG: Discrete Flow Matching for Graph Generation
Yiming QIN, Pascal Frossard (École Polytechnique Fédérale de Lausanne)
GenerationData SynthesisGraph Neural NetworkTransformerFlow-based ModelGenerative Adversarial NetworkGraph
🎯 What it does: DeFoG is proposed—a graph generation framework based on Discrete Flow Matching (DFM), achieving decoupling of training and sampling, significantly enhancing sampling flexibility and efficiency;
Delay-DSGN: A Dynamic Spiking Graph Neural Network with Delay Mechanisms for Evolving Graph
Zhiqiang Wang (Shanxi University), Jianqing Liang (Shanxi University)
ClassificationGraph Neural NetworkSpiking Neural NetworkGraph
🎯 What it does: A dynamic pulse graph neural network called Delay-DSGN is proposed, which captures the temporal delays and historical information of dynamic graphs through a learnable delay mechanism.
Deliberation in Latent Space via Differentiable Cache Augmentation
Luyang Liu (Google DeepMind), Arthur Szlam (Google DeepMind)
TransformerLarge Language ModelText
🎯 What it does: Train an offline coprocessor to generate potential embeddings on the frozen KV cache of the Gemma-2 2B LLM to improve the accuracy of subsequent token generation.
Delta Decompression for MoE-based LLMs Compression
Hao Gu (Hong Kong University of Science and Technology), Yike Guo (Hong Kong University of Science and Technology)
CompressionTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: Designed and implemented the D2-MOE framework, achieving efficient compression without training by splitting the MoE expert weights into shared base weights and expert-specific Delta weights.
Demeaned Sparse: Efficient Anomaly Detection by Residual Estimate
Yifan Fang (Xiamen University), Yue Huang (Xiamen University)
Anomaly DetectionConvolutional Neural NetworkImage
🎯 What it does: A statistical test based on mean-deviation Fourier transform residual projection and factor model is proposed, and a DFS module is designed to achieve unsupervised image anomaly detection.
Demonstration Selection for In-Context Learning via Reinforcement Learning
Xubin Wang (Beijing Normal University), Weijia Jia (Beijing Normal University)
ClassificationTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextChain-of-Thought
🎯 What it does: In the context of few-shot learning (ICL), this paper proposes a reinforcement learning-based demonstration selection framework called RDES, which dynamically balances the relevance and diversity of demonstrations to enhance the generalization performance of large language models in text classification and reasoning tasks.
Demystifying Catastrophic Forgetting in Two-Stage Incremental Object Detector
Qirui Wu (Northwestern Polytechnical University), Yanning Zhang (Northwestern Polytechnical University)
Object DetectionImage
🎯 What it does: This study investigates the catastrophic forgetting mechanism of the two-stage incremental object detector Faster R-CNN and proposes the NSGP-RePRE framework to mitigate forgetting in the RoI Head classifier.
Demystifying Cost-Efficiency in LLM Serving over Heterogeneous GPUs
YOUHE JIANG, Eiko Yoneki (University of Cambridge)
OptimizationComputational EfficiencyLarge Language ModelTextBenchmark
🎯 What it does: This paper first conducts a comprehensive benchmarking of the cost-effectiveness of various GPU types in LLM inference, and based on this, proposes a scheduling framework that simultaneously optimizes GPU combinations, deployment configurations, and workload allocations;
Demystifying Long Chain-of-Thought Reasoning
Shiming Yang (IN.AI), Xiang Yue (Carnegie Mellon University)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
🎯 What it does: This paper systematically studies the mechanism of long chain-of-thought (CoT) reasoning in large language models, exploring the effects of supervised fine-tuning (SFT), reinforcement learning (RL), reward design, and the use of noise-verified data on the model's generation of long CoTs.
Demystifying Singular Defects in Large Language Models
Haoqi Wang (École polytechnique fédérale de Lausanne), Mathieu Salzmann (École polytechnique fédérale de Lausanne)
TransformerLarge Language ModelText
🎯 What it does: This paper conducts a systematic analysis of the high-norm tokens that appear in large language models (LLMs) and proposes a new 'singular defect' theoretical framework, revealing four key stages in the emergence and disappearance of high-norm tokens (development, triggering, explosion, decay).
Demystifying the Paradox of Importance Sampling with an Estimated History-Dependent Behavior Policy in Off-Policy Evaluation
Hongyi Zhou (Tsinghua University), Chengchun Shi (London School of Economics and Political Science)
Reinforcement LearningSequential
🎯 What it does: This paper studies policy evaluation in offline reinforcement learning, proposing to improve estimators such as importance sampling (OIS, SIS, DR, MIS) using history-dependent behavior policy estimates, and provides complete theoretical proofs and error decomposition.
Dendritic Localized Learning: Toward Biologically Plausible Algorithm
Changze Lv (Fudan University), Xuanjing Huang (Fudan University)
ClassificationExplainability and InterpretabilityConvolutional Neural NetworkRecurrent Neural NetworkImageTextTime Series
🎯 What it does: A biologically interpretable learning algorithm called Dendritic Localized Learning (DLL) is proposed for training multilayer neural networks, achieving supervised learning on architectures such as MLP, CNN, and RNN.
Density Ratio Estimation with Conditional Probability Paths
Hanlin Yu (University of Helsinki), Omar Chehab (ENSAE, CREST, IP Paris)
Score-based ModelContrastive LearningImage
🎯 What it does: This paper transforms the high-dimensional density ratio estimation problem into learning time scores by constructing conditional probability paths and proposes a new time score matching method.
Density Ratio Estimation-based Bayesian Optimization with Semi-Supervised Learning
Jungtaek Kim (University of Wisconsin-Madison)
OptimizationTabular
🎯 What it does: A Bayesian optimization framework for density ratio estimation combined with semi-supervised learning (DRE‑BO‑SSL) is proposed, which enhances the search efficiency for the global optimum by utilizing unlabeled samples.
Dequantified Diffusion-Schrödinger Bridge for Density Ratio Estimation
Wei Chen (South China University of Technology), Delu Zeng (South China University of Technology)
Diffusion modelScore-based ModelTabularOrdinary Differential Equation
🎯 What it does: The DRE3 framework is proposed, utilizing the dequantized diffusion bridge (DDBI) and the dequantized Schrödinger bridge (DSBI) for robust density ratio estimation, addressing the density-chasm and support-chasm issues.
Design Considerations in Offline Preference-based RL
Alekh Agarwal (Google Research), Teodor Vanislavov Marinov
Reinforcement Learning from Human FeedbackTransformerReinforcement LearningText
🎯 What it does: This paper conducts a theoretical analysis of design choices (loss function, baseline policy, data sampling strategy) in offline reinforcement learning with human preferences (Offline RLHF) and provides a realizable suboptimal baseline.
Designing Cyclic Peptides via Harmonic SDE with Atom-Bond Modeling
Xiangxin Zhou (ByteDance Seed, School of Artificial Intelligence, University of Chinese Academy of Sciences, New Laboratory of Pattern Recognition, State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of Sciences), Quanquan Gu (Institute for AI Industry Research, Tsinghua University)
GenerationDrug DiscoveryProtein Structure PredictionDiffusion modelGraphBiomedical DataStochastic Differential Equation
🎯 What it does: Proposes CPSDE, a generative model based on harmonic SDE, which utilizes atom-bond representation to simultaneously generate the three-dimensional structure and sequence of cyclic peptides, supporting all four types of cyclization;