ICML 2025 Papers — Page 8
International Conference on Machine Learning · 3257 papers
Detecting Strategic Deception with Linear Probes
Nicholas Goldowsky-Dill (Apollo Research), Marius Hobbhahn (Apollo Research)
Anomaly DetectionTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper identifies the strategic deception behavior of large language models by training linear detectors and evaluates their effectiveness in various adversarial and real-world scenarios.
Determinant Estimation under Memory Constraints and Neural Scaling Laws
Siavash Ameli (University of California), Michael W. Mahoney (Lawrence Berkeley National Laboratory)
OptimizationComputational EfficiencyImage
🎯 What it does: A new algorithm MEMDET for computing the log-determinant of large positive definite matrices in memory-constrained environments is proposed, and a FLODANCE estimation method is developed based on the neural network scaling law, which can achieve accurate or efficient estimation on extremely large NTK Gram matrices.
Determining Layer-wise Sparsity for Large Language Models Through a Theoretical Perspective
Weizhong Huang (Xiamen University), Rongrong Ji (Xiamen University)
TransformerLarge Language ModelText
🎯 What it does: This paper proposes a theoretical-based hierarchical sparsity allocation method (ATP), which provides an arithmetic series distribution of sparsity rates increasing with hierarchy by analyzing the explosion of reconstruction error during the sparsification process.
Deterministic Sparse Fourier Transform for Continuous Signals with Frequency Gap
Xiaoyu Li (University of New South Wales), Shenghao Xie (Texas A&M University)
🎯 What it does: A deterministic sparse Fourier transform (SFT) algorithm for continuous signal scenarios is proposed, capable of recovering sparse frequencies in sublinear time and sample complexity.
Devil is in the Details: Density Guidance for Detail-Aware Generation with Flow Models
Rafal Karczewski, Vikas K Garg
GenerationData SynthesisDiffusion modelFlow-based ModelImageStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes two methods, Score Alignment and Density Guidance, to achieve precise control over the sample density generated by Continuous Normalizing Flows (CNF) and diffusion models, thereby finely adjusting image details.
DexScale: Automating Data Scaling for Sim2Real Generalizable Robot Control
Guiliang Liu (Chinese University of Hong Kong), Kui Jia (Chinese University of Hong Kong)
Domain AdaptationRobotic IntelligenceTransformerLarge Language ModelReinforcement LearningDiffusion modelVideoMultimodality
🎯 What it does: DexScale is proposed, an automated data engine that can map real task descriptions to simulation environments and automate domain randomization and adaptation, thereby generating robot control datasets for zero-shot Sim2Real transfer.
Diagonal Symmetrization of Neural Network Solvers for the Many-Electron Schrödinger Equation
Kevin Han Huang (University College London), Ryan P Adams
OptimizationComputational EfficiencyGraph Neural NetworkTabularPhysics Related
🎯 What it does: This study investigates the symmetrization strategy under diagonal symmetry (G_diag) in a neural network solver for the multi-electron Schrödinger equation, comparing the effects of data augmentation, group averaging, and normalization on training and inference.
Dialogue Without Limits: Constant-Sized KV Caches for Extended Response in LLMs
Ravi Ghadia (University of Texas at Austin), Poulami Das
GenerationCompressionComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper proposes MorphKV, a technique that maintains a constant size KV cache during inference by dynamically selecting the oldest tokens that are most strongly associated with the recent tokens to retain key information, significantly reducing memory consumption.
Diff-MoE: Diffusion Transformer with Time-Aware and Space-Adaptive Experts
Kun Cheng (State Key Laboratory of Integrated Services Networks Xidian University), Jie Hu (Huawei Noah's Ark Lab)
GenerationData SynthesisTransformerMixture of ExpertsDiffusion modelImage
🎯 What it does: Proposes Diff-MoE, a generative model that combines Diffusion Transformer with Mixture-of-Experts;
DiffAdvMAP: Flexible Diffusion-Based Framework for Generating Natural Unrestricted Adversarial Examples
Zhengzhao Pan (Hunan University), Xiaogang Zhang (Hunan University)
GenerationAdversarial AttackDiffusion modelImage
🎯 What it does: A flexible diffusion-based unconstrained adversarial attack framework called DiffAdvMAP is proposed for generating naturally realistic unconstrained adversarial samples.
Differentiable Quadratic Optimization For the Maximum Independent Set Problem
Ismail Alkhouri, Alvaro Velasquez (University of Colorado)
OptimizationGraph
🎯 What it does: A new continuous quadratic optimization framework pCQO-MIS based on maximum clique information is proposed to solve the maximum independent set problem.
Differentiable Solver Search for Fast Diffusion Sampling
Shuai Wang (Nanjing University), Limin Wang (Nanjing University)
GenerationOptimizationDiffusion modelImageStochastic Differential Equation
🎯 What it does: A differentiable solver search algorithm is proposed to find better numerical integration methods for diffusion models under a limited number of sampling steps.
Differentiable Structure Learning with Ancestral Constraints
Taiyu Ban (University of Science and Technology of China), Huanhuan Chen (University of Science and Technology of China)
Graph Neural NetworkGraphTabular
🎯 What it does: This paper proposes a differentiable structure learning framework that integrates ancestral constraints (path existence constraints) into DAG learning using binary mask continuous relaxation and order-guided optimization.
Differential Coding for Training-Free ANN-to-SNN Conversion
Zihan Huang (Peking University), Tiejun Huang (Peking University)
Spiking Neural NetworkTransformerImage
🎯 What it does: This paper proposes a training-independent ANN-to-SNN conversion method based on differential coding.
Differential Privacy Guarantees of Markov Chain Monte Carlo Algorithms
Andrea Bertazzi (Ecole Polytechnique), Alain Oliviero Durmus
Safty and Privacy
🎯 What it does: This paper aims to provide differential privacy (DP) guarantees for Markov Chain Monte Carlo (MCMC) algorithms. First, it establishes DP guarantees for the output samples of MCMC algorithms and their associated Monte Carlo estimators, with a particular emphasis on the condition that the target distribution itself needs to be differentially private. Second, it specifically analyzes the unadjusted Langevin algorithm and stochastic gradient Langevin dynamics, establishing their (Rényi) DP guarantees.
Differential Privacy Under Class Imbalance: Methods and Empirical Insights
Lucas Rosenblatt (New York University), Rachel Cummings (Columbia University)
ClassificationSafty and PrivacyTransformerTabular
🎯 What it does: The study investigates how to achieve high-quality binary classification learning under the constraint of differential privacy (DP) in scenarios with severe data imbalance; it evaluates the privacy feasibility and performance of various preprocessing and internal processing methods through theory and experiments.
Differentially Private Analysis for Binary Response Models: Optimality, Estimation, and Inference
Ce Zhang (University of Alberta), Bei Jiang (University of Alberta)
OptimizationSafty and PrivacyTabular
🎯 What it does: This paper proposes a method for optimal estimation and inference of binary response models using the random response (RR) mechanism within the framework of label differential privacy (LabelDP).
Differentially Private Boxplots
Kelly Ramsay (York University), Jairo Diaz-Rodriguez (York University)
Safty and PrivacyTabular
🎯 What it does: A differentially private boxplot (DPBoxplot) is proposed, which estimates quartiles and extremes by combining JointExp and unbounded quantization algorithms.
Differentially Private Federated $k$-Means Clustering with Server-Side Data
Jonathan Scott (Institute of Science and Technology Austria), David Saulpic (CNRS and Université Paris Cité)
Federated LearningSafty and PrivacyText
🎯 What it does: This paper proposes a differential privacy k-means clustering algorithm FedDP-KMeans implemented within a federated learning framework. It utilizes a small amount of potentially differently distributed data on the server side for initialization, followed by multiple rounds of DP noise addition during the Lloyd iteration between the client and server.
Differentially Private Space-Efficient Algorithms for Counting Distinct Elements in the Turnstile Model
Rachel Cummings (Columbia University), Peilin Zhong (Google Research)
Safty and PrivacyComputational Efficiency
🎯 What it does: Implementing differentially private counting of distinct elements in a sustainable streaming data model with space less than linear.
DiffMS: Diffusion Generation of Molecules Conditioned on Mass Spectra
Montgomery Bohde (Massachusetts Institute of Technology), Connor W. Coley (Massachusetts Institute of Technology)
GenerationData SynthesisDrug DiscoveryGraph Neural NetworkTransformerDiffusion modelGraph
🎯 What it does: DiffMS is constructed, a condition-based molecular generation model based on mass spectra, utilizing a discrete graph diffusion decoder with formula constraints to achieve structural inference.
Diffuse Everything: Multimodal Diffusion Models on Arbitrary State Spaces
Kevin Rojas (Georgia Institute of Technology), Molei Tao (Georgia Institute of Technology)
GenerationData SynthesisTransformerDiffusion modelImageTextMultimodalityTabular
🎯 What it does: A framework has been constructed that can perform multimodal diffusion models in arbitrary state spaces, achieving joint generation of native multimodal data.
Diffusion Adversarial Post-Training for One-Step Video Generation
Shanchuan Lin (ByteDance), Lu Jiang (ByteDance)
GenerationData SynthesisTransformerDiffusion modelGenerative Adversarial NetworkImageVideo
🎯 What it does: Based on the pre-trained diffusion model (DiT), adversarial post-training (APT) is applied for one-step generation directly on real data, successfully achieving one-step generation of high-resolution videos (1280×720, 24fps, 2 seconds) and images (1024px).
Diffusion Counterfactual Generation with Semantic Abduction
Rajat R Rasal, Ben Glocker (Imperial College London)
GenerationData SynthesisDiffusion modelImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: This paper studies how to combine diffusion models with structural causal models (SCM) for high-fidelity, identity-preserving counterfactual image generation, and proposes three mechanisms: spatial, semantic, and dynamic semantic induction.
Diffusion Instruction Tuning
Chen Jin (AstraZeneca), Philip Alexander Teare
GenerationData SynthesisOptimizationTransformerSupervised Fine-TuningVision Language ModelDiffusion modelMultimodality
🎯 What it does: This paper presents Lavender, a supervised fine-tuning framework that aligns the text-visual attention maps of diffusion models with visual-language models.
Diffusion Models are Secretly Exchangeable: Parallelizing DDPMs via Auto Speculation
Hengyuan Hu (Stanford University), Nima Anari (Stanford University)
GenerationComputational EfficiencyRobotic IntelligenceDiffusion modelImage
🎯 What it does: A parallel sampling algorithm called Autospeculative Decoding (ASD) is proposed during the inference process of diffusion models (DDPM), which significantly accelerates inference while maintaining sample quality by utilizing hidden exchangeability.
Diffusion models for Gaussian distributions: Exact solutions and Wasserstein errors
Emile Pierret (Universite d'Orleans), Bruno Galerne (Institut Universitaire de France)
Diffusion modelImageStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: The study analyzes the analytical solution of diffusion models under Gaussian distribution, the sources of error, and their numerical sampling errors, providing a closed-form 2-Wasserstein error analysis.
Diffusion on Language Model Encodings for Protein Sequence Generation
Viacheslav Meshchaninov (Constructor University), Dmitry Vetrov (AIRI)
GenerationData SynthesisProtein Structure PredictionTransformerLarge Language ModelDiffusion modelBiomedical Data
🎯 What it does: DiMA has been developed, a continuous latent diffusion framework operating on protein language model encoding, for generating high-quality and diverse protein sequences.
Diffusion Sampling Correction via Approximately 10 Parameters
Guangyi Wang (Xiamen University), Song-Zhi Su
GenerationData SynthesisOptimizationDiffusion modelImageOrdinary Differential Equation
🎯 What it does: A PCA-based sampling correction method called PAS is proposed, which can improve the sampling quality of DPMs without significantly increasing training costs.
Diffusion-based Adversarial Purification from the Perspective of the Frequency Domain
Gaozheng Pei (University of Chinese Academy of Sciences), Qingming Huang (University of Chinese Academy of Sciences)
RestorationAdversarial AttackDiffusion modelImage
🎯 What it does: A frequency-domain based diffusion adversarial purification method called FreqPure is proposed, which retains only the low-frequency amplitude and phase spectra during the reverse diffusion process to remove adversarial perturbations while preserving semantic information.
DiffusionVLA: Scaling Robot Foundation Models via Unified Diffusion and Autoregression
Junjie Wen (Midea Group), Feifei Feng (Midea Group)
Robotic IntelligenceTransformerVision-Language-Action ModelDiffusion modelMultimodality
🎯 What it does: This paper proposes DiffusionVLA, a unified framework for autoregressive and diffusion models for robot vision-language action tasks.
DiLQR: Differentiable Iterative Linear Quadratic Regulator via Implicit Differentiation
Shuyuan Wang (University of British Columbia), Wei Pan (University of Manchester)
OptimizationReinforcement Learning from Human FeedbackReinforcement LearningSequential
🎯 What it does: This paper proposes a differentiable iterative linear quadratic regulator (DiLQR) that achieves efficient gradient propagation for iLQR through implicit differentiation, supporting end-to-end training and model identification.
DiMa: Understanding the Hardness of Online Matching Problems via Diffusion Models
Boyu Zhang (Huazhong University of Science and Technology), Xianjun Deng (Huazhong University of Science and Technology)
OptimizationTransformerReinforcement LearningDiffusion model
🎯 What it does: A unified framework called DiMa is proposed, utilizing the denoising diffusion probabilistic model (DDPM) to generate more challenging online matching instances, and fine-tuned through reinforcement learning (shortcut policy gradient, SPG) to improve the upper bound of competitive rates.
DIME: Diffusion-Based Maximum Entropy Reinforcement Learning
Onur Celik (Karlsruhe Institute of Technology), Gerhard Neumann (Karlsruhe Institute of Technology)
Reinforcement LearningDiffusion modelTabular
🎯 What it does: This paper proposes a maximum entropy reinforcement learning framework based on diffusion models, called DIME, which utilizes approximate inference to obtain an entropy lower bound and achieves provable policy iteration.
Dimension-Free Adaptive Subgradient Methods with Frequent Directions
Sifan Yang (Nanjing University), Lijun Zhang (Nanjing University)
OptimizationTabular
🎯 What it does: This paper proposes a new online adaptive subgradient algorithm FTSL (Follow the Sketchy Leader), which utilizes Frequent Directions (FD) for low-rank approximation of the gradient outer product matrix, and incorporates cumulative discarded information within the original ADA-FULL's origin-bilateral framework, resulting in a dimension-independent regret analysis.
Dimension-Independent Rates for Structured Neural Density Estimation
Robert A. Vandermeulen, Bryon Aragam
GenerationData SynthesisGraph Neural NetworkAuto EncoderImageAudio
🎯 What it does: A structured neural density estimation framework based on Markov Random Fields (MRF) is proposed, and it is proven that a dimension-independent convergence rate can be achieved under the assumption of local conditional independence.
Dimensionality Reduction on Complex Vector Spaces for Euclidean Distance with Dynamic Weights
Simone Moretti (University of Padova), Francesco Silvestri (University of Padova)
🎯 What it does: A linear mapping for dimensionality reduction in complex vector spaces is proposed, which can estimate the weighted Euclidean distance when weights change dynamically;
DINO-WM: World Models on Pre-trained Visual Features enable Zero-shot Planning
Gaoyue Zhou (New York University), Lerrel Pinto (New York University)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerWorld ModelImage
🎯 What it does: Construct an offline world model based on pre-trained visual features, capable of zero-shot planning using MPC directly during testing.
DipLLM: Fine-Tuning LLM for Strategic Decision-making in Diplomacy
Kaixuan Xu (University of Chinese Academy of Sciences), Dongbin Zhao (Chinese Academy of Sciences)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIText
🎯 What it does: A self-regressive factor decomposition agent DipLLM based on a large language model is constructed to learn equilibrium strategies for Diplomacy and directly generate multi-unit actions.
Direct Density Ratio Optimization: A Statistically Consistent Approach to Aligning Large Language Models
Rei Higuchi (University of Tokyo), Taiji Suzuki (University of Tokyo)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A direct estimation method for the density ratio of preferred and non-preferred output distributions, called DDRO, is proposed, and its statistical consistency is proven.
Direct Discriminative Optimization: Your Likelihood-Based Visual Generative Model is Secretly a GAN Discriminator
Kaiwen Zheng (NVIDIA), Qinsheng Zhang (NVIDIA)
GenerationOptimizationDiffusion modelGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes the Direct Discriminative Optimization (DDO) method, which implicitly couples the discriminator with likelihood-based models to achieve efficient fine-tuning of pre-trained diffusion and autoregressive models, thereby overcoming the mode coverage limitations of MLE.
Direct Motion Models for Assessing Generated Videos
Kelsey R Allen, Sjoerd van Steenkiste (Google Research)
GenerationData SynthesisTransformerAuto EncoderOptical FlowVideo
🎯 What it does: This paper proposes a video generation quality assessment method based on the trajectory autoencoder TRAJAN, which can measure the motion consistency of a single video and locate spatiotemporal errors.
Direct Prediction Set Minimization via Bilevel Conformal Classifier Training
Yuanjie Shi (Washington State University), Yan Yan (Washington State University)
ClassificationOptimizationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: This paper proposes a new two-layer optimization framework that directly minimizes the size of the classifier's prediction set during the training process, achieving a more compact prediction set that meets coverage requirements.
Directed Graph Grammars for Sequence-based Learning
Michael Sun (Massachusetts Institute of Technology), Jie Chen (IBM Research)
GenerationOptimizationGraph Neural NetworkTransformerAuto EncoderGraph
🎯 What it does: A DAG encoding method based on graph grammar, DIGGED, is proposed, which losslessly compresses the DAG into a unique sequence of rules and constructs an end-to-end autoencoder for generation, prediction, and Bayesian optimization.
Directly Forecasting Belief for Reinforcement Learning with Delays
Qingyuan Wu (University of Southampton), Chao Huang (University of Southampton)
TransformerReinforcement LearningSequential
🎯 What it does: To address the problem of reinforcement learning with observation delays, this paper proposes the Direct Forecasting Belief Transformer (DFBT) and its derivative method DFBT-SAC, which utilizes one-shot sequence prediction to directly reconstruct unobservable states, thereby avoiding the cumulative errors associated with recursive predictions.
DIS-CO: Discovering Copyrighted Content in VLMs Training Data
André V. Duarte (Instituto Superior Técnico), Lei Li (Carnegie Mellon University)
RecognitionGenerationAnomaly DetectionTransformerPrompt EngineeringVision Language ModelImageVideo
🎯 What it does: The DIS-CO method is proposed, which infers whether the content has been included in the training set by sending movie frames to the VLM and allowing it to complete with free text.
DISCO: learning to DISCover an evolution Operator for multi-physics-agnostic prediction
Rudy Morel (Flatiron Institute), Edouard Oyallon (Sorbonne Université)
TransformerTime SeriesPhysics RelatedOrdinary Differential Equation
🎯 What it does: This paper proposes DISCO, which learns the evolution operator of PDEs from short trajectories using a transformer hypernetwork and performs continuous time prediction with this operator.
Discovering a Zero (Zero-Vector Class of Machine Learning)
Harikrishna Metta (Indian Institute of Science), Venkatesh Babu Radhakrishnan
ClassificationData SynthesisConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: The paper proposes treating categories in machine learning as vectors in a vector space, and constructs a vector space by defining addition (corresponding to the union of classes) and scalar multiplication (corresponding to the complement of classes). It further discovers that the zero vector of this space corresponds to a category (Metta-Class) that can be viewed as uniformly distributed noise in the feature space.
Discovering Global False Negatives On the Fly for Self-supervised Contrastive Learning
Vicente Balmaseda (Texas A&M University), Tianbao Yang (Texas A&M University)
OptimizationRepresentation LearningContrastive LearningImageMultimodality
🎯 What it does: This paper proposes an online learning global threshold method GLOFND, which automatically discovers and filters 'false negatives' in self-supervised contrastive learning to improve representation quality.
Discovering Latent Causal Graphs from Spatiotemporal Data
Kun Wang (University of California), Rose Yu (University of California)
Explainability and InterpretabilityComputational EfficiencyTime Series
🎯 What it does: An end-to-end framework SPACY based on variational inference is proposed for jointly learning latent time series and corresponding causal graphs from high-dimensional spatiotemporal grid data.
Discovering Physics Laws of Dynamical Systems via Invariant Function Learning
Shurui Gui (Texas A&M University), Shuiwang Ji (Texas A&M University)
OptimizationExplainability and InterpretabilityMeta LearningTime SeriesPhysics RelatedOrdinary Differential Equation
🎯 What it does: A distinguishable function learning framework based on causal analysis (DIF) is proposed to extract the essential invariant functions of dynamical systems from observational trajectories in different environments, thereby achieving the discovery of physical laws.
Discovering Spoofing Attempts on Language Model Watermarks
Thibaud Gloaguen (ETH Zurich), Martin Vechev (ETH Zurich)
Large Language ModelPrompt EngineeringText
🎯 What it does: This paper studies the observable traces left by learning-based watermark forgery attacks in text and designs statistical testing methods based on this to distinguish between genuine watermark texts and forged texts.
Discovering Symbolic Cognitive Models from Human and Animal Behavior
Pablo Samuel Castro (Google DeepMind), Kim Stachenfeld
OptimizationReinforcement Learning from Human FeedbackRecurrent Neural NetworkLarge Language ModelTabularTime SeriesSequential
🎯 What it does: Using a large language model-driven evolutionary search to automatically generate symbolic cognitive models that fit the behavior of humans, mice, and fruit flies in reward learning tasks.
Discrepancies are Virtue: Weak-to-Strong Generalization through Lens of Intrinsic Dimension
Yijun Dong (New York University), Qi Lei (New York University)
Knowledge DistillationTransformerGaussian SplattingImage
🎯 What it does: This paper studies the transfer process from weak teacher models to strong student models (Weak-to-Strong Generalization, W2S) and proposes an explanation for why strong students can achieve better generalization after receiving pseudo-labels from weak teachers, based on the intrinsic dimensions of the models and the differences in teacher-student characteristics.
Discrepancy Minimization in Input-Sparsity Time
Yichuan Deng (University of Washington), OMRI WEINSTEIN
OptimizationComputational EfficiencyGaussian Splatting
🎯 What it does: A novel combinatorial algorithm is proposed to achieve real-valued matrix contradiction minimization within sparse time input.
Discrete and Continuous Difference of Submodular Minimization
George Orfanides (McGill University), Marwa El Halabi (Samsung AI Lab)
Optimization
🎯 What it does: This paper studies the problem of minimizing the difference of submodular functions (Difference of Submodular, DS) in both discrete and continuous domains, and proposes an improved DC algorithm (DCA‑LS) to find approximate local optimal solutions.
Discrete Markov Probabilistic Models: An Improved Discrete Score-Based Framework with sharp convergence bounds under minimal assumptions
Le-Tuyet-Nhi PHAM, Alain Oliviero Durmus
GenerationData SynthesisDiffusion modelScore-based ModelImage
🎯 What it does: A new discrete generative model DMPM is proposed, using continuous-time Markov chains as the noise process and performing reverse sampling through a discrete score function in the time-reversal process.
Discrete Neural Algorithmic Reasoning
Gleb Rodionov (Yandex Research), Liudmila Prokhorenkova (Yandex Research)
TransformerGraphBenchmark
🎯 What it does: Construct an interpretable discrete neural algorithm interpreter that forces the network to execute algorithm steps on a finite set of predefined state combinations.
Discriminative Finetuning of Generative Large Language Models without Reward Models and Human Preference Data
Siqi Guo (Texas A&M University), Tianbao Yang (Texas A&M University)
GenerationOptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A discriminative fine-tuning framework (DFT/DFT2) is proposed that does not rely on reward models or human preference data, shifting LLM from traditional generative fine-tuning to likelihood-based learning;
Discriminative Policy Optimization for Token-Level Reward Models
Hongzhan Chen (Sun Yat-sen University), Ting Yao (Tencent Inc)
OptimizationReinforcement Learning from Human FeedbackTransformerReinforcement LearningText
🎯 What it does: A Q-RM (Q-Function Reward Model) based on a discriminative strategy is proposed, which learns token-level rewards from preference data, avoiding conflicts between the generative model and reward modeling.
Disentangled Graph Spectral Domain Adaptation
Liang Yang (Hebei University of Technology), Yuanfang Guo (Beihang University)
Domain AdaptationGraph Neural NetworkGraph
🎯 What it does: The Disentangled Graph Spectral Domain Adaptation (DGSDA) framework is proposed, decoupling attribute alignment from topological alignment, and directly aligning spectral filters to achieve unsupervised graph domain adaptation.
Disentangling and Integrating Relational and Sensory Information in Transformer Architectures
Awni Altabaa (Yale University), John Lafferty (Yale University)
TransformerImage
🎯 What it does: This paper proposes a Dual Attention Transformer (DAT) that decouples and integrates perceptual information and relational information into the Transformer through two mechanisms: perceptual attention and relational attention.
Disentangling Invariant Subgraph via Variance Contrastive Estimation under Distribution Shifts
Haoyang Li (Tsinghua University), Wenwu Zhu (Tsinghua University)
ClassificationDomain AdaptationGraph Neural NetworkContrastive LearningGraph
🎯 What it does: To address the issue of distribution drift in graph data, the VIVACE method is proposed, which estimates variant subgraphs through self-supervised contrastive learning and identifies invariant subgraphs that are independent of labels using inverse propensity weighting, achieving OOD generalization for graph-level classification.
Disparate Conditional Prediction in Multiclass Classifiers
Sivan Sabato (McMaster University), Elad Yom-Tov (Bar-Ilan University)
ClassificationOptimizationTabular
🎯 What it does: A fairness auditing method for multi-classifiers is proposed, defining and extending the Disparate Conditional Prediction (DCP) metric to quantify the bias of multi-class models relative to multi-class equal opportunity.
Diss-l-ECT: Dissecting Graph Data with Local Euler Characteristic Transforms
Julius von Rohrscheidt (Institute of AI for Health Helmholtz Munich), Bastian Rieck (University of Fribourg)
ClassificationGraph Neural NetworkGraph
🎯 What it does: This paper proposes the Local Euler Characteristic Transformation (ℓ-ECT), which extends the global ECT to the local neighborhood of graphs, preserving the geometric-topological information of node neighborhoods for node classification and graph space alignment.
Dissecting Submission Limit in Desk-Rejections: A Mathematical Analysis of Fairness in AI Conference Policies
Yuefan Cao (University of Washington), Jiahao Zhang (University of Washington)
OptimizationTabular
🎯 What it does: This study investigates the fairness issue of desk-rejection caused by submission limits in artificial intelligence conferences and proposes an optimization-based desk-rejection framework based on fairness metrics.
Distillation of Discrete Diffusion through Dimensional Correlations
Satoshi Hayakawa (Sony Group Corporation), Yuki Mitsufuji (Sony AI)
GenerationKnowledge DistillationMixture of ExpertsDiffusion modelImageText
🎯 What it does: This paper proposes a discrete diffusion model distillation framework based on dimension correlation, Di4C, which can compress a multi-step diffusion process into fewer sampling steps while capturing the correlation between dimensions through a mixture model.
Distillation Scaling Laws
Dan Busbridge (Apple), Russell Webb
Knowledge DistillationTransformerLarge Language ModelText
🎯 What it does: This paper proposes a distillation extension law that can predict the cross-entropy performance of the student model after distillation based on computational budget and the resource allocation between teacher and student.
Distilling the Knowledge in Data Pruning
Emanuel Ben Baruch (Amazon), Gerard Medioni (Amazon)
ClassificationKnowledge DistillationImage
🎯 What it does: The study uses knowledge distillation (KD) after data pruning to improve model accuracy.
DistiLLM-2: A Contrastive Approach Boosts the Distillation of LLMs
Jongwoo Ko (Korea Advanced Institute of Science and Technology), Se-Young Yun (Korea Advanced Institute of Science and Technology)
Knowledge DistillationTransformerLarge Language ModelContrastive LearningTextMultimodality
🎯 What it does: Proposes DISTILLM-2, a LLM distillation framework that utilizes the idea of contrastive learning;
Distinguishing Cause from Effect with Causal Velocity Models
Johnny Xi (University of British Columbia), Benjamin Bloem-Reddy (University College London)
Score-based ModelTabularOrdinary Differential Equation
🎯 What it does: This paper proposes a method to view bivariate causal structure models as dynamical systems, utilizing causal speed and score functions to achieve causal direction determination without simulation.
Distributed Conformal Prediction via Message Passing
Haifeng Wen (Hong Kong University of Science and Technology), Osvaldo Simeone (King's College London)
OptimizationFederated LearningReinforcement LearningImage
🎯 What it does: Two distributed conformal prediction methods based on message passing in a fully decentralized network (where devices can only communicate with neighbors) are proposed: Q-DCP (quantile regression + ADMM) and H-DCP (consistent histogram estimation), achieving reliable inference and coverage guarantees.
Distributed Differentially Private Data Analytics via Secure Sketching
Jakob Burkhardt (Aarhus University), Chris Schwiegelshohn (Aarhus University)
OptimizationSafty and PrivacyTabular
🎯 What it does: This paper proposes a Linear Transformation Model (LTM), under which differentially private low-rank approximations and ridge regression are achieved through secure multi-party computation after linear compression is performed.
Distributed Event-Based Learning via ADMM
Guner Dilsad ER, Michael Muehlebach (Max Planck Institute for Intelligent Systems)
OptimizationFederated LearningReinforcement LearningImage
🎯 What it does: A distributed learning framework driven by events is proposed, which triggers communication only when local model changes exceed a threshold, significantly reducing communication volume.
Distributed Nonparametric Estimation: from Sparse to Dense Samples per Terminal
Deheng Yuan (Tsinghua University), Zhongyi Huang (Tsinghua University)
🎯 What it does: This paper studies the nonparametric function estimation problem in a distributed environment with communication constraints, where each terminal has a different number of samples. It provides the globally optimal asymptotic error rate and describes the phase transition from sparse to dense samples.
Distributed Parallel Gradient Stacking(DPGS): Solving Whole Slide Image Stacking Challenge in Multi-Instance Learning
Boyuan Wu (Huzhou University), Lianxin Hu (Huzhou University)
ClassificationOptimizationComputational EfficiencyImageBiomedical Data
🎯 What it does: Proposes the Distributed Parallel Gradient Stacking (DPGS) framework, which achieves batch processing of Multi-Instance Learning (MIL) through gradient stacking, thereby solving the problem of stacking bags of different lengths, and on this basis, designs the Deep Model-Gradient Compression (DMGC) to further compress the communication volume of gradients and weights.
Distributed Retraction-Free and Communication-Efficient Optimization on the Stiefel Manifold
Yilong Song (Peking University), Kun Yuan (Peking University)
OptimizationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a distributed, non-degenerate, and communication-efficient EF-Landing algorithm for solving distributed stochastic optimization problems with orthogonal constraints on the Stiefel manifold.
Distribution-aware Fairness Learning in Medical Image Segmentation From A Control-Theoretic Perspective
Yujin Oh (Massachusetts General Hospital), Quanzheng Li (Massachusetts General Hospital)
SegmentationConvolutional Neural NetworkTransformerMixture of ExpertsImageBiomedical DataComputed TomographyOrdinary Differential Equation
🎯 What it does: A distribution-aware mixture of experts (dMoE) framework is proposed to enhance the fairness and robustness of medical image segmentation.
Distributional Diffusion Models with Scoring Rules
Valentin De Bortoli (Google DeepMind), Arnaud Doucet (Gatsby UCL)
GenerationData SynthesisDiffusion modelScore-based ModelImage
🎯 What it does: By replacing the regression loss of traditional diffusion models with score-based methods (energy score/kernel score), we learn the complete conditional posterior distribution p0|t, allowing for fast sampling at coarse time steps.
Distributionally Robust Active Learning for Gaussian Process Regression
Shion Takeno (Nagoya University), Ichiro Takeuchi (Nagoya University)
Tabular
🎯 What it does: A distributionally robust active learning framework for Gaussian process regression is proposed, aiming to minimize the worst-case expected error over a candidate set of target distributions and providing a theoretical convergence upper bound.
Distributionally Robust Multi-Agent Reinforcement Learning for Dynamic Chute Mapping
Guangyi Liu (Amazon Robotics), Michael M. Zavlanos (Duke University)
OptimizationRobotic IntelligenceReinforcement LearningTabular
🎯 What it does: This paper proposes and validates a Distributionally Robust Multi-Agent Reinforcement Learning (DRMARL) framework for dynamic conveyor mapping in Amazon robotic sorting warehouses, significantly reducing the package delivery cycle rate.
Distributionally Robust Policy Learning under Concept Drifts
Jingyuan Wang (New York University), Zhengyuan Zhou (Arena Technologies)
OptimizationReinforcement LearningTabularTime Series
🎯 What it does: This paper studies distributionally robust policy learning under concept drift and proposes a doubly robust policy value estimation and optimal learning algorithm.
DiTAR: Diffusion Transformer Autoregressive Modeling for Speech Generation
Dongya Jia (ByteDance), Yuxuan Wang (ByteDance)
GenerationData SynthesisComputational EfficiencyTransformerLarge Language ModelDiffusion modelOrdinary Differential EquationAudio
🎯 What it does: DiTAR is proposed, a patch-based autoregressive framework that combines language models with diffusion Transformers for generating continuous speech representations.
Diverging Preferences: When do Annotators Disagree and do Models Know?
Michael JQ Zhang, Valentina Pyatkin (Allen Institute for Artificial Intelligence)
Reinforcement Learning
🎯 What it does: This paper studies the sources of disagreement among humans when labeling preference data and proposes a classification of ten types of disagreement reasons.
Diverse Prototypical Ensembles Improve Robustness to Subpopulation Shift
Nguyen Nhat Minh To, Rahul Krishnan
ClassificationDomain AdaptationSupervised Fine-TuningImage
🎯 What it does: This paper proposes a Diversified Prototype Ensemble (DPE) method that does not require subgroup labels to improve the robustness of models under subgroup distribution shifts.
Diversified Flow Matching with Translation Identifiability
Sagar Shrestha (Oregon State University), Xiao Fu (Oregon State University)
Image TranslationData SynthesisDomain AdaptationOptimizationFlow-based ModelGenerative Adversarial NetworkImageOrdinary Differential Equation
🎯 What it does: A distribution matching framework based on flow matching is proposed—Diverse Flow Matching (DFM), which achieves unaligned domain translation recognition by learning a unified flow field.
Diversifying Policy Behaviors with Extrinsic Behavioral Curiosity
Zhenglin Wan (Chinese University of Hong Kong), Ivor Tsang
Robotic IntelligenceReinforcement LearningGenerative Adversarial Network
🎯 What it does: This paper combines inverse reinforcement learning with quality diversity (QD) methods, proposing an external behavior curiosity-based reward mechanism (EBC) that enhances the behavior diversity and performance of the IL model under limited demonstration data by rewarding unoccupied behavior space units.
Diversity By Design: Leveraging Distribution Matching for Offline Model-Based Optimization
Michael S Yao, Osbert Bastani (University of Pennsylvania)
OptimizationAuto EncoderGenerative Adversarial Network
🎯 What it does: A method named DynAMO is proposed, which explicitly introduces diversity into the objective function in offline model-based optimization (MBO) using distribution matching and an adversarial source evaluator, helping to generate a set of design candidates that are both high-quality and diverse.
Divide and Conquer: Exploring Language-centric Tree Reasoning for Video Question-Answering
Zhaohe Liao (Shanghai Jiao Tong University), Liqing Zhang (Shanghai Jiao Tong University)
Explainability and InterpretabilityLarge Language ModelVideoTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: By constructing a language-centric logical tree and using recursive reasoning, the reasoning ability and interpretability of multimodal large language models in video question answering are enhanced.
Divide and Conquer: Grounding LLMs as Efficient Decision-Making Agents via Offline Hierarchical Reinforcement Learning
Zican Hu (Nanjing University), Yu Cheng (Chinese University of Hong Kong)
TransformerLarge Language ModelReinforcement LearningAgentic AIText
🎯 What it does: This paper proposes a GLIDER framework for LLM agents based on offline hierarchical reinforcement learning, which can decompose large language models into high-level planning and low-level execution strategies, enabling efficient learning and deployment of long-term decision-making tasks.
Divide and Conquer: Learning Label Distribution with Subtasks
Haitao Wu (Nanjing University of Science and Technology), Xiuyi Jia (Nanjing University of Science and Technology)
ClassificationRecognitionSegmentationRepresentation LearningImageTextMagnetic Resonance Imaging
🎯 What it does: A S-LDL framework is proposed, which automatically constructs sub-tasks (sub-label spaces) without prior knowledge to generate additional supervisory information, thereby enhancing the performance of label distribution learning (LDL).
Diving into Self-Evolving Training for Multimodal Reasoning
Wei Liu (Hong Kong University of Science and Technology), Junxian He (Hong Kong University of Science and Technology)
OptimizationReinforcement LearningPrompt EngineeringMultimodality
🎯 What it does: This paper proposes a self-evolving training framework for multimodal reasoning, M-STAR, which systematically studies and optimizes training methods, reward models, and prompt variants, and achieves continuous improvement through dynamic balancing of exploration and exploitation.
DLP: Dynamic Layerwise Pruning in Large Language Models
Yuli Chen (Beijing University of Posts and Telecommunications), Shuhao Zhang (Beijing University of Posts and Telecommunications)
TransformerLarge Language ModelText
🎯 What it does: A dynamic layer pruning method (DLP) is proposed, which calculates the importance of each layer by combining weight and activation information, thereby allocating non-uniform sparsity rates;
DMM: Distributed Matrix Mechanism for Differentially-Private Federated Learning Based on Constant-Overhead Linear Secret Resharing
Alexander Bienstock (J.P. Morgan), Antigoni Polychroniadou (J.P. Morgan)
Federated LearningSafty and PrivacyText
🎯 What it does: A Distributed Matrix Mechanism (DMM) is proposed in federated learning, which achieves distributed differential privacy through linear secret redistribution in each round and generates relevant noise in multiple rounds using the matrix mechanism.
DMOSpeech: Direct Metric Optimization via Distilled Diffusion Model in Zero-Shot Speech Synthesis
Yinghao Aaron Li (Columbia University), Zeyu Jin (Adobe Research)
Data SynthesisOptimizationKnowledge DistillationTransformerDiffusion modelAuto EncoderAudio
🎯 What it does: A zero-shot TTS system based on diffusion models was trained, compressing the inference steps from 128 to 4 using distribution matching distillation, and achieving end-to-end differentiable metric optimization in one-click inference.
Do Bayesian Neural Networks Actually Behave Like Bayesian Models?
Gábor Pituk (University of Oxford), Tom Rainforth (University of Oxford)
ClassificationOptimizationImageText
🎯 What it does: Evaluate whether mainstream approximate Bayesian neural networks (VI, Laplace, SWAG, SGLD, HMC) adhere to Bayesian theory in terms of functional space, sequential consistency, and predictive consistency, and conduct experiments on tasks such as synthetic regression, CIFAR-10/100, and IMDB text classification.
Do Multiple Instance Learning Models Transfer?
Daniel Shao (Massachusetts Institute of Technology), Faisal Mahmood (Harvard University)
ClassificationDomain AdaptationRepresentation LearningTransformerBiomedical Data
🎯 What it does: Under the multi-instance learning (MIL) framework, transfer learning experiments were conducted on 19 computational pathology tasks, evaluating the transfer effects of 11 MIL architectures and 21 pre-training tasks. A 'foundation model' at the slide level based on pancancer supervised pre-training was proposed and validated, along with the provision of unified implementation and weight resources.
Do Not Mimic My Voice : Speaker Identity Unlearning for Zero-Shot Text-to-Speech
Taesoo Kim, Gyeong-Moon Park (Korea University)
GenerationData SynthesisAudio
🎯 What it does: A 'Guided Forgetting' framework is proposed for zero-shot text-to-speech (ZS-TTS) models, which allows the model to completely forget the voice information of a specified speaker while maintaining high-quality synthesis for speakers that have not been forgotten.
Do NOT Think That Much for 2+3=? On the Overthinking of Long Reasoning Models
Xingyu Chen (Shanghai Jiao Tong University), Dong Yu (Tencent)
OptimizationComputational EfficiencyTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: This paper studies the phenomenon of overthinking in long-chain reasoning models (such as Qwen-QwQ and DeepSeek-R1) during the reasoning process. It proposes efficiency metrics for results and processes, and reduces the number of generated tokens without sacrificing accuracy through a self-training length preference optimization method.
Do Vision-Language Models Really Understand Visual Language?
Yifan Hou (ETH Zurich), Mrinmaya Sachan (ETH Zurich)
RecognitionObject DetectionTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageMultimodalityPhysics RelatedChain-of-Thought
🎯 What it does: This study evaluates the understanding of graphic symbol language by large-scale visual language models (LVLMs) through the construction of a comprehensive test set (including synthetic and real images), focusing on entity recognition, relationship recognition, and reasoning abilities.