NeurIPS 2025 Papers — Page 12
Conference on Neural Information Processing Systems · 5275 papers
Diff-ICMH: Harmonizing Machine and Human Vision in Image Compression with Generative Prior
Ruoyu Feng (University of Science and Technology of China), Zhibo Chen (University of Science and Technology of China)
CompressionDiffusion modelImage
🎯 What it does: Proposes Diff-ICMH, an image compression framework that balances machine analysis and human perception.
DiffBreak: Is Diffusion-Based Purification Robust?
Andre Kassis (University of Waterloo), Yaoliang Yu (University of Waterloo)
Adversarial AttackConvolutional Neural NetworkTransformerDiffusion modelScore-based ModelAuto EncoderImage
🎯 What it does: Evaluated and demonstrated that Diffusion-based Purification (DBP) lacks robustness in the presence of precise gradients, and proposed methods such as the DiffBreak tool, DiffGrad gradient implementation, and low-frequency attacks;
DiffE2E: Rethinking End-to-End Driving with a Hybrid Diffusion-Regression-Classification Policy
Rui Zhao (Jilin University), Zhenhai Gao (Jilin University)
Autonomous DrivingTransformerDiffusion modelMultimodality
🎯 What it does: This paper proposes DiffE2E, an end-to-end autonomous driving framework that integrates diffusion, regression, and classification to achieve multimodal perception to trajectory generation.
Differentiable Constraint-Based Causal Discovery
Jincheng Zhou (Purdue University), Bruno Ribeiro (Purdue University)
OptimizationGraph Neural NetworkGraph
🎯 What it does: A differentiable d-separation framework is proposed, and based on this framework, DAGPA is implemented for causal structure learning, which can directly utilize the conditional independence information in the observed data for gradient optimization.
Differentiable Cyclic Causal Discovery Under Unmeasured Confounders
Muralikrishnna Guruswamy Sethuraman, Faramarz Fekri (Georgia Institute of Technology)
Graph Neural NetworkFlow-based ModelGraphTabular
🎯 What it does: A differentiable causal structure learning framework DCCD-CONF is proposed, which can learn nonlinear causal graphs from experimental data in the presence of hidden confounding variables and directed cycles.
Differentiable Decision Tree via "ReLU+Argmin" Reformulation
Qiangqiang Mao (University of British Columbia), Yankai Cao (University of British Columbia)
ClassificationOptimizationExplainability and InterpretabilityTabular
🎯 What it does: A differentiable slope decision tree based on ReLU+Argmin is proposed, utilizing gradient methods for one-time optimization of the entire tree, achieving a high-precision and interpretable model.
Differentiable extensions with rounding guarantees for combinatorial optimization over permutations
Robert R Nerem (Georgia Institute of Technology), Yusu Wang (Georgia Institute of Technology)
Optimization
🎯 What it does: This study investigates how to continuously extend the permutation function to double stochastic matrices and provides a differentiable Birkhoff extension.
Differentiable Generalized Sliced Wasserstein Plans
Laetitia Chapel (IRISA L'Institut Agro Rennes Angers), Samuel Vaiter (CNRS Université Côte d'Azur)
GenerationOptimizationImage
🎯 What it does: A differentiable generalized sliced Wasserstein plan (DGSWP) is proposed, achieving OT approximation in high dimensions and manifolds through neural network projection and Stein smoothing;
Differentiable Hierarchical Visual Tokenization
Marius Aasan (University of Oslo), Adín Ramírez Rivera (University of Oslo)
ClassificationSegmentationConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper proposes a differentiable hierarchical visual segmenter (∂HT) that can adaptively generate visual tokens with pixel-level accuracy in end-to-end training.
Differentiable Sparsity via $D$-Gating: Simple and Versatile Structured Penalization
Chris Kolb (LMU Munich), David Rügamer (LMU Munich)
OptimizationConvolutional Neural NetworkImage
🎯 What it does: The D-Gating method is proposed, which achieves differentiable structural sparsity by splitting each structural group into main weights and multi-scale gating factors, compatible with standard SGD and applicable to any neural network architecture.
Differentiable Structure Learning and Causal Discovery for General Binary Data
Chang Deng (University of Chicago), Bryon Aragam (University of Chicago)
GraphTabular
🎯 What it does: A differentiable structure learning framework based on the multinomial Bernoulli distribution is proposed, which can capture any dependency relationships in discrete data and identify the sparsest Markov equivalence class with only observational data.
Differential Privacy for Euclidean Jordan Algebra with Applications to Private Symmetric Cone Programming
Zhao Song (University of California), Lichen Zhang (Massachusetts Institute of Technology)
OptimizationSafty and Privacy
🎯 What it does: This paper proposes a general Gaussian mechanism for achieving differential privacy in Euclidean Jordan algebra (EJA) spaces and applies it to symmetric cone programming (SCP) problems, particularly including semidefinite programming (SDP) and second-order cone programming (SOCP).
Differential Privacy on Fully Dynamic Streams
Yuan Qiu (Southeast University), Ke Yi (Hong Kong University of Science and Technology)
Safty and PrivacyTime Series
🎯 What it does: This paper studies mechanisms for achieving differential privacy in fully dynamic streams, proposing an efficient black-box construction method that can continuously publish private answers to linear queries at each time point.
Differentially Private Bilevel Optimization: Efficient Algorithms with Near-Optimal Rates
Andrew Lowy (CISPA Helmholtz Center for Information Security), Daogao Liu (Google Research)
OptimizationSafty and Privacy
🎯 What it does: A theoretical framework for differential privacy (DP) double-layer optimization is proposed, providing efficient algorithms and error upper bounds in both convex and non-convex scenarios, achieving a near-optimal convergence rate comparable to single-layer DP ERM in the convex case.
Differentially Private Federated Low Rank Adaptation Beyond Fixed-Matrix
Ming Wen (Fudan University), DINGDING HAN
Federated LearningSafty and Privacy
🎯 What it does: In the federated learning environment, the FedASK framework is proposed, using a two-stage sketching technique to achieve differential privacy secure updates for LoRA low-rank adapters;
Differentially Private Gomory-Hu Trees
Anders Aamand (University of Copenhagen), Yinzhan Xu (University of California San Diego)
OptimizationSafty and PrivacyGraph Neural NetworkGraph
🎯 What it does: An ε-differential privacy algorithm is designed to output an approximate Gomory-Hu tree, thereby achieving an approximation of the minimum cut for all pairs of vertices, with an error of ˜(O(n/ε)), and a polynomial time implementation is provided.
Differentially Private High-dimensional Variable Selection via Integer Programming
Petros Prastakos (Massachusetts Institute of Technology), Rahul Mazumder (Massachusetts Institute of Technology)
OptimizationSafty and PrivacyTabular
🎯 What it does: This paper proposes two scalable pure differential privacy variable selection algorithms (topR and mistakes) for support recovery in sparse high-dimensional learning;
Differentially Private Quantiles with Smaller Error
Jacob Imola (University of Copenhagen), Rasmus Pagh (University of Copenhagen)
Safty and PrivacyComputational EfficiencyTabular
🎯 What it does: A new mechanism for estimating multiple quantiles under pure differential privacy and approximate differential privacy is proposed, utilizing continuous counting techniques to achieve lower rank error.
Differentially Private Relational Learning with Entity-level Privacy Guarantees
Yinan Huang (Georgia Institute of Technology), Pan Li (Georgia Institute of Technology)
Safty and PrivacyGraph Neural NetworkSupervised Fine-TuningTextGraph
🎯 What it does: This paper proposes a differential privacy training framework suitable for relational learning, providing entity-level DP guarantees.
Differentiation Through Black-Box Quadratic Programming Solvers
Connor W. Magoon (University of North Carolina), Shahar Z. Kovalsky (Université de Montréal)
Optimization
🎯 What it does: A general, plug-and-play differentiable layer (dQP) is proposed, which can be used with any black-box quadratic programming (QP) solver, allowing for the simultaneous computation of solutions and gradients with just one KKT matrix decomposition.
DiffEye: Diffusion-Based Continuous Eye-Tracking Data Generation Conditioned on Natural Images
Ozgur Kara (University of Illinois), James Matthew Rehg
GenerationData SynthesisDiffusion modelImage
🎯 What it does: A framework for generating eye movement trajectories based on diffusion models, called DiffEye, has been designed and implemented. It can generate eye movement trajectories, scanning paths, and saliency maps for natural images by training on raw continuous eye movement trajectories.
DiffLiG: Diffusion-enhanced Liquid Graph with Attention Propagation for Grid-to-Station Precipitation Correction
Yuxiang Li (Sun Yat-sen University), Juepeng Zheng (Sun Yat-sen University)
Graph Neural NetworkGraphTime SeriesOrdinary Differential Equation
🎯 What it does: A graph neural network framework named DiffLiG is proposed, specifically designed to refine grid forecasts to the level of meteorological stations and address site heterogeneity and over-smoothing issues.
DIFFSSR: Stereo Image Super-resolution Using Differential Transformer
Dafeng Zhang (Samsung Research and Development Institute China)
RestorationSuper ResolutionTransformerImage
🎯 What it does: The DIFFSSR model is proposed, which improves the stereo image super-resolution task;
Diffusing DeBias: Synthetic Bias Amplification for Model Debiasing
Massimiliano Ciranni (University of Genoa), Vittorio Murino (Istituto Italiano di Tecnologia)
ClassificationData SynthesisDiffusion modelImage
🎯 What it does: In an unsupervised scenario, synthetic images with bias alignment are generated using a conditional diffusion model, training a Bias Amplifier, which is then used as a plugin for debiasing learning in image classification models.
Diffusion Adaptive Text Embedding for Text-to-Image Diffusion Models
Byeonghu Na (Korea Advanced Institute of Science and Technology), Il-chul Moon
GenerationData SynthesisDiffusion modelImageText
🎯 What it does: A dynamic text embedding update method called DATE is proposed, which can adjust text embeddings in real-time during the sampling process of diffusion models to better match the generated images;
Diffusion Beats Autoregressive in Data-Constrained Settings
Mihir Prabhudesai (Carnegie Mellon University), Deepak Pathak (Carnegie Mellon University)
TransformerDiffusion modelText
🎯 What it does: This study investigates the performance of mask diffusion models compared to autoregressive models under data-constrained conditions, finding that diffusion models can better utilize repeated data through multiple rounds of training and ultimately outperform autoregressive models.
Diffusion Feature Field for Text-based 3D Editing with Gaussian Splatting
Eunseo Koh (Sungkyunkwan University), Jae-Pil Heo (Sungkyunkwan University)
GenerationData SynthesisDiffusion modelAuto EncoderGaussian SplattingPoint Cloud
🎯 What it does: This paper proposes a DFFSplat framework that utilizes a 3D consistent diffusion feature field and a dual-encoder for text-driven 3D Gaussian Splatting editing, significantly enhancing multi-view consistency and addressing the Janus problem.
Diffusion Federated Dataset
Seok-Ju Hahn (Argonne National Laboratory), Junghye Lee (Seoul National University)
Data SynthesisFederated LearningDiffusion modelImage
🎯 What it does: A synthetic data generation framework DfD for parameter-free model synchronization in federated learning is designed, utilizing inference from local diffusion models instead of parameter exchange, and sampling from local distribution mixtures through ULA.
Diffusion Generative Modeling on Lie Group Representations
Marco Bertolini (Pfizer Worldwide Research and Development), Djork-Arné Clevert (Pfizer Worldwide Research and Development)
GenerationData SynthesisDrug DiscoveryDiffusion modelScore-based ModelTabularSequentialStochastic Differential Equation
🎯 What it does: In this paper, the authors propose a new diffusion process based on score matching, modeling directly in the representation space of arbitrary Lie groups. They derive a discretized Langevin dynamics using generalized score matching and prove that these dynamics are solutions to a dual stochastic differential equation system that can be solved analytically.
Diffusion Guided Adversarial State Perturbations in Reinforcement Learning
Xiaolin Sun (Tulane University), Zizhan Zheng (Tulane University)
Autonomous DrivingAdversarial AttackReinforcement LearningDiffusion modelAuto EncoderImageVideo
🎯 What it does: This paper proposes a diffusion model-based adversarial state perturbation attack (SHIFT) that can perform semantic-level attacks on visual inputs in reinforcement learning while maintaining historical consistency and trajectory authenticity.
Diffusion Model as a Noise-Aware Latent Reward Model for Step-Level Preference Optimization
Tao Zhang (Chinese Academy of Sciences), Chunhong Pan (Chinese Academy of Sciences)
GenerationOptimizationReinforcement LearningDiffusion modelImage
🎯 What it does: This study proposes using the pre-trained diffusion model itself as a noise-aware latent reward model (LRM) and directly performing stepwise preference optimization (LPO) in the noise latent space, significantly improving image quality and training efficiency.
Diffusion Models and the Manifold Hypothesis: Log-Domain Smoothing is Geometry Adaptive
Tyler Farghly (University of Oxford), Jakiw Pidstrigach (University of Oxford)
GenerationData SynthesisDiffusion modelScore-based ModelAuto EncoderImage
🎯 What it does: This study investigates how the smoothing of the score function in diffusion models during training implicitly adapts to the low-dimensional manifold structure of the data, and verifies its impact on generalization through theoretical and experimental methods.
Diffusion Models Meet Contextual Bandits
Imad Aouali (Criteo AI Lab)
Recommendation SystemOptimizationReinforcement LearningDiffusion modelTabular
🎯 What it does: This paper proposes a context-based Thompson sampling algorithm (dTS) based on a pre-trained diffusion model, utilizing the diffusion model as an expressive prior to capture the complex correlations between actions, and achieving efficient online posterior updates and sampling through hierarchical sampling and Gaussian approximation.
Diffusion on Demand: Selective Caching and Modulation for Efficient Generation
Hee Min Choi (Samsung Electronics), Nam Ik Cho (Seoul National University)
GenerationData SynthesisComputational EfficiencyTransformerDiffusion modelImageTextOrdinary Differential Equation
🎯 What it does: A cache-based framework is proposed, which significantly reduces inference computation by selectively adjusting cached features in the diffusion Transformer.
Diffusion Transformers as Open-World Spatiotemporal Foundation Models
Yuan Yuan (Tsinghua University), Yong Li (Tsinghua University)
TransformerPrompt EngineeringDiffusion modelGraphTime Series
🎯 What it does: This paper presents UrbanDiT, an open-world spatiotemporal foundation model based on diffusion Transformers, capable of uniformly handling various spatiotemporal data such as grids and graphs, and supporting multiple tasks including forward prediction, reverse prediction, temporal interpolation, spatial extrapolation, and spatiotemporal filling in a single pass.
Diffusion Transformers for Imputation: Statistical Efficiency and Uncertainty Quantification
Zeqi Ye (Northwestern University), Minshuo Chen (Northwestern University)
TransformerDiffusion modelTime Series
🎯 What it does: This paper studies the statistical efficiency of Diffusion Transformers in time series missing value imputation and uncertainty quantification, providing upper bounds on sample complexity and confidence interval construction, and proposing a mixed masking training strategy.
Diffusion Tree Sampling: Scalable inference‑time alignment of diffusion models
Vineet Jain (Mila - Quebec Artificial Intelligence Institute), Siamak Ravanbakhsh (Mila - Quebec Artificial Intelligence Institute)
GenerationData SynthesisComputational EfficiencyReinforcement LearningDiffusion modelImageText
🎯 What it does: This paper proposes a tree search-based alignment method for diffusion model inference called Diffusion Tree Sampling (DTS) and its search variant DTS⋆. By treating the reverse diffusion process as a finite-depth tree and using soft value backup for global credit assignment, it can align the prior distribution to the posterior distribution based on a reward function during inference, significantly reducing computational costs while maintaining high-quality samples.
Diffusion-Based Hierarchical Graph Neural Networks for Simulating Nonlinear Solid Mechanics
Tobias Würth (Karlsruhe Institute of Technology), Luise Kärger (Karlsruhe Institute of Technology)
Graph Neural NetworkDiffusion modelGraphPhysics Related
🎯 What it does: A learned simulator called ROBIN, based on diffusion models and hierarchical graph neural networks, is proposed for efficiently and accurately simulating nonlinear solid mechanics problems.
Diffusion-Classifier Synergy: Reward-Aligned Learning via Mutual Boosting Loop for FSCIL
Ruitao Wu (Beihang University), Jia Li (Beihang University)
ClassificationGenerationData SynthesisReinforcement LearningDiffusion modelImage
🎯 What it does: This paper proposes a Diffusion-Classifier Synergy (DCS) based on a mutual enhancement cycle between diffusion models and classifiers, generating high-quality and targeted training samples in few-shot incremental learning through a reward-aligned learning strategy, addressing issues of semantic mismatch and insufficient diversity.
Diffusion-Driven Progressive Target Manipulation for Source-Free Domain Adaptation
Yuyang Huang (Shanghai Jiao Tong University), Qi Tian (Huawei Inc.)
Domain AdaptationDiffusion modelImage
🎯 What it does: This paper proposes a source-end unsupervised domain adaptation framework based on a latent diffusion model—Diffusion-Driven Progressive Target Manipulation (DPTM). By dividing target domain samples into reliable and unreliable sets, it directly trains the target model using pseudo-labels from the reliable set, and after label assignment for the unreliable set samples, it employs a three-step diffusion manipulation (Target-guided Initialization, Semantic Feature Injection, Domain-specific Feature Preservation) to transform their semantics into new labels while preserving target domain features. Subsequently, a progressive refinement mechanism is used to iteratively update the gap between the pseudo target domain and the real target domain, gradually improving the performance of the target model.
Diffusion-Driven Two-Stage Active Learning for Low-Budget Semantic Segmentation
Jeongin Kim (Ewha Womans University), Junhyug Noh (Ewha Womans University)
SegmentationDiffusion modelImage
🎯 What it does: A two-stage active learning framework for low-budget semantic segmentation is proposed, which first selects diverse candidate pixels from multi-scale features extracted from a pre-trained diffusion model using MaxHerding, and then selects the final labeled pixels using eDALD (mutual information + entropy);
Diffusion-Guided Graph Data Augmentation
Maria Marrium (Information Technology University), Wenxiong Kang (South China University of Technology)
Representation LearningData-Centric LearningGraph Neural NetworkDiffusion modelAuto EncoderGraph
🎯 What it does: A label-agnostic diffusion-based graph data augmentation framework D-GDA is proposed, which generates diverse and consistent incremental samples in the latent space using a graph variational autoencoder and a latent diffusion model.
Dimension-adapted Momentum Outscales SGD
Damien Ferbach (Mila and Université de Montréal), Courtney Paquette (Google DeepMind and McGill University)
OptimizationRecurrent Neural NetworkTextStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper studies the scaling laws of the stochastic momentum algorithm under power-law random feature models and proves that the DANA (Dimension-Adaptive Nesterov Acceleration), adjusted according to model size and data complexity, can significantly enhance the scaling exponent of the loss, thereby 'surpassing' traditional SGD;
Dimension-free Score Matching and Time Bootstrapping for Diffusion Models
Syamantak Kumar (University of Texas at Austin), Purnamrita Sarkar (University of Texas at Austin)
Diffusion modelScore-based ModelTime Series
🎯 What it does: This paper proves that when using a single function approximator to jointly estimate score functions at different noise levels, it is possible to achieve almost dimension-independent sample complexity, and proposes the Bootstrapped Score Matching (BSM) algorithm that utilizes previously learned scores to achieve variance reduction.
Dimension-Reduction Attack! Video Generative Models are Experts on Controllable Image Synthesis
Hengyuan Cao, Bin Wang
GenerationData SynthesisTransformerVision Language ModelImageVideoText
🎯 What it does: This paper proposes a Dimension-Reduction Attack (DRA-Ctrl) framework that transfers a pre-trained high-dimensional video generation model to a low-dimensional controllable image generation task, achieving functions such as subject-driven image generation, spatial alignment, and style transfer at the image level.
Dimensional Collapse in VQVAEs: Evidence and Remedies
Jiayou Zhang (Mohammed Bin Zayed University of Artificial Intelligence), Eric P. Xing (Carnegie Mellon University)
RestorationGenerationData SynthesisConvolutional Neural NetworkTransformerAuto EncoderImage
🎯 What it does: This study addresses the phenomenon of dimensional collapse in VQ-VAE and proposes Divide-and-Conquer VQ (DCVQ), which enhances effective dimensionality and reconstruction quality by splitting the latent space into multiple low-dimensional subspaces and quantizing them independently.
Dimensionality Mismatch Between Brains and Artificial Neural Networks
Santiago Galella (Frankfurt Institute for Advanced Studies and Goethe University Frankfurt), Matthias Kaschube (Frankfurt Institute for Advanced Studies and Goethe University Frankfurt)
ImageMagnetic Resonance Imaging
🎯 What it does: This paper systematically quantifies and compares the linear and nonlinear dimensions of human brain fMRI and artificial neural networks during the process of viewing natural images.
DINGO: Constrained Inference for Diffusion LLMs
Tarun Suresh (University of Illinois Urbana-Champaign), Gagandeep Singh (University of Illinois Urbana-Champaign)
GenerationOptimizationTransformerLarge Language ModelDiffusion modelText
🎯 What it does: A constraint reasoning method for diffusion language models called DINGO is proposed, which can enforce compliance with user-defined regular expressions or structural constraints during the generation process, ensuring valid outputs.
DINO-Foresight: Looking into the Future with DINO
Efstathios Karypidis (Archimedes Athena Research Center), Nikos Komodakis (IACM-Forth)
SegmentationAutonomous DrivingComputational EfficiencyTransformerContrastive LearningVideo
🎯 What it does: Proposes the DINO-Foresight framework, which utilizes the high-dimensional semantic feature space of a pre-trained Vision Foundation Model (VFM) for self-supervised future frame semantic prediction.
DIPO: Dual-State Images Controlled Articulated Object Generation Powered by Diverse Data
Ruiqi Wu (Nankai University), Ming-Ming Cheng (Nankai University)
GenerationData SynthesisTransformerLarge Language ModelDiffusion modelImageMultimodalityChain-of-Thought
🎯 What it does: Using paired still and motion state images, controllable three-dimensional joint objects are generated through a diffusion model.
Direct Alignment with Heterogeneous Preferences
Ali Shirali, Ariel D. Procaccia
Recommendation SystemReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningTabular
🎯 What it does: This paper studies how to obtain a unified strategy that accommodates multiple user types by directly aligning preference data in scenarios where user preferences are highly heterogeneous.
Direct Fisher Score Estimation for Likelihood Maximization
Sherman Khoo (University of Bristol), Mark Beaumont (University of Bristol)
OptimizationGenerative Adversarial NetworkTabular
🎯 What it does: This study investigates a technique for achieving maximum likelihood estimation through local Fisher score matching (FSM) in situations where the likelihood cannot be analytically resolved in likelihood-free inference.
Direct Numerical Layout Generation for 3D Indoor Scene Synthesis via Spatial Reasoning
Xingjian Ran, Bo Dai
GenerationData SynthesisLarge Language ModelReinforcement LearningVision Language ModelTextPoint CloudChain-of-Thought
🎯 What it does: Developed the DirectLayout framework, which directly generates 3D indoor scene layouts from text descriptions, including BEV generation, 3D enhancement, and iterative alignment.
Direct3D-S2: Gigascale 3D Generation Made Easy with Spatial Sparse Attention
Shuang Wu (Nanjing University), Yao Yao (Nanjing University)
GenerationData SynthesisTransformerDiffusion modelPoint CloudMesh
🎯 What it does: A 3D generation framework based on sparse voxels, Direct3D-S2, is proposed, achieving high-quality 3D shape generation at a resolution of 1024³.
Directed-Tokens: A Robust Multi-Modality Alignment Approach to Large Language-Vision Models
Thanh-Dat Truong (University of Arkansas), Khoa Luu (University of Arkansas)
RecognitionGenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: Proposes a sequential reconstruction task of images and texts and introduces directed-token and Image-to-Response Guided Loss to enhance the visual-text alignment and robustness of large multimodal models (LMM).
DISC: Dynamic Decomposition Improves LLM Inference Scaling
Jonathan Light (Rensselaer Polytechnic Institute), Haifeng Chen (NEC Laboratories America)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: A dynamic decomposition method called DISC is proposed, which can adaptively partition steps and allocate computational resources during the inference process.
DISCO: DISCrete nOise for Conditional Control in Text-to-Image Diffusion Models
Longquan Dai (Nanjing University of Science and Technology), Jinhui Tang (Nanjing University of Science and Technology)
GenerationData SynthesisTransformerDiffusion modelImage
🎯 What it does: This paper proposes the DISCO method, which utilizes a pre-constructed discrete noise codebook to replace continuous conditional adjustments, achieving conditional generation in text-to-image diffusion models.
DISCO: Disentangled Communication Steering for Large Language Models
Max Torop (Northeastern University), Jennifer Dy (Northeastern University)
TransformerLarge Language ModelText
🎯 What it does: This paper proposes a new vector injection method—DISCO Steering, which directly injects the average difference vector into the query and value representation space of the Transformer attention heads to regulate the output of large language models. It also provides theoretical analysis, linear separability assessment, and various benchmark comparison experiments.
DisCO: Reinforcing Large Reasoning Models with Discriminative Constrained Optimization
Gang Li (Texas A&M University), Tianbao Yang (Texas A&M University)
OptimizationReinforcement LearningText
🎯 What it does: A reinforcement learning framework called DisCO based on discriminative constraint optimization is proposed to enhance the mathematical reasoning performance of large-scale inference models.
DISCOVER: Automated Curricula for Sparse-Reward Reinforcement Learning
Leander Diaz-Bone (ETH Zurich), Andreas Krause (ETH Zurich)
Reinforcement Learning
🎯 What it does: This paper proposes an automatic course learning method based on goal selection called DISCOVER, aimed at addressing the sparse reward long-term reinforcement learning problem.
Discovering Compositional Hallucinations in LVLMs
Sibei Yang (Sun Yat-sen University), Cheng Shi (University of Hong Kong)
Knowledge DistillationRepresentation LearningTransformerVision Language ModelMultimodalityBenchmark
🎯 What it does: This paper proposes the concept of Simple Combination Hallucination (SCHall), constructs the SCBench benchmark, and designs the VLR-distillation method to mitigate this phenomenon.
Discovering Data Structures: Nearest Neighbor Search and Beyond
Omar Salemohamed (University of Montreal), Gregory Valiant (Stanford University)
TransformerImageTabular
🎯 What it does: This paper proposes an end-to-end learning framework that allows neural networks to learn to construct and query data structures from scratch, focusing on nearest neighbor search and extending it to problems like frequency estimation.
Discovering Important Experts for Mixture-of-Experts Models Pruning Through a Theoretical Perspective
Weizhong Huang (Xiamen University), Liujuan Cao (Xiamen University)
CompressionOptimizationComputational EfficiencyLarge Language ModelMixture of ExpertsText
🎯 What it does: This paper proposes a Shapley value-based Mixture-of-Experts (MoE) expert pruning method that automatically identifies and removes experts with low contributions to model performance, significantly reducing memory and computational costs.
Discovering Latent Graphs with GFlowNets for Diverse Conditional Image Generation
Bailey Trang (Stanford University), Ehsan Adeli (Stanford University)
GenerationData SynthesisRecurrent Neural NetworkDiffusion modelFlow-based ModelImageBiomedical DataMagnetic Resonance ImagingAlzheimer's Disease
🎯 What it does: On any pre-trained conditional generative model, by constructing a latent graph and using GFlowNets to sample multiple trajectories, an uncertain condition is decomposed into various latent representations, generating corresponding diversified and interpretable images.
Discovering Opinion Intervals from Conflicts in Signed Graphs
Peter Blohm (Aalto University), Stefan Neumann (TU Wien)
Graph Neural NetworkGraph
🎯 What it does: The BEST INTERVAL APPROXIMATION problem is proposed, aiming to assign overlapping opinion intervals to nodes in a signed graph to explain the conflicts and cooperation of positive and negative edges.
Discovering Symbolic Partial Differential Equation by Abductive Learning
En-Hao Gao (Nanjing University), Zhi-Hua Zhou (Nanjing University)
TabularPhysics Related
🎯 What it does: Based on the abductive learning framework and FOL knowledge base, the ABL-PDE method is proposed to discover symbolic partial differential equations from data and estimate their coefficients.
Discrete Diffusion Models: Novel Analysis and New Sampler Guarantees
Yuchen Liang (Ohio State University), Ness Shroff (Ohio State University)
GenerationDrug DiscoveryDiffusion modelText
🎯 What it does: A new analytical framework based on differential inequalities is proposed, providing convergence guarantees for sampling methods of discrete diffusion models (especially τ-leaping, Euler method, and Tweedie τ-leaping).
Discrete Neural Flow Samplers with Locally Equivariant Transformer
Zijing Ou (Imperial College London), Yingzhen Li (Imperial College London)
OptimizationTransformerFlow-based ModelGraph
🎯 What it does: This paper proposes Discrete Neural Flow Samplers (DNFS), which achieve efficient sampling of discrete unnormalized distributions by learning the rate matrix of continuous-time Markov chains.
Discrete Spatial Diffusion: Intensity-Preserving Diffusion Modeling
Javier E. Santos (Los Alamos National Laboratory), Yen Ting Lin (Los Alamos National Laboratory)
RestorationGenerationData SynthesisDiffusion modelImage
🎯 What it does: A Discrete Spatial Diffusion (DSD) model is proposed, achieving a complete diffusion generation process in discrete pixel and particle counting space while strictly maintaining the total particle count, supporting conditional generation and image inpainting.
Discretization-free Multicalibration through Loss Minimization over Tree Ensembles
Hongyi Henry Jin (University of California Los Angeles), Steven Wu
ClassificationOptimizationSupervised Fine-TuningTabular
🎯 What it does: A multi-calibration method that does not require discretization is proposed, achieving balanced calibration for multiple groups by minimizing the empirical risk of squared loss on an ensemble of depth-2 trees based on a benchmark predictor.
Disentangled Concepts Speak Louder Than Words: Explainable Video Action Recognition
Jongseo Lee (Kyung Hee University), Jinwoo Choi (Kyung Hee University)
RecognitionExplainability and InterpretabilityTransformerLarge Language ModelVideo
🎯 What it does: This paper proposes a concept-based explainable video action recognition framework called DANCE, which generates structured explanations by breaking down action predictions into three types of interpretable concepts: motion dynamics, objects, and scenes.
Disentangled Cross-Modal Representation Learning with Enhanced Mutual Supervision
Lu Gao (Central South University), Cheng Liang (Shandong Normal University)
GenerationRepresentation LearningAuto EncoderContrastive LearningImageMultimodality
🎯 What it does: Proposes the DCMEM framework, which utilizes a multimodal variational autoencoder to achieve cross-modal representation learning, decoupling shared and modality-specific information, and realizes separable shared and private latent variables through enhanced mutual supervision and information bottleneck.
Disentangled Representation Learning via Modular Compositional Bias
Whie Jung (Korea Advanced Institute of Science and Technology), Seunghoon Hong (Korea Advanced Institute of Science and Technology)
GenerationRepresentation LearningDiffusion modelAuto EncoderImage
🎯 What it does: A general framework is proposed that can achieve attribute, object, and their joint disentanglement by adjusting the mixing strategy;
Disentangling Hyperedges through the Lens of Category Theory
Yoonho Lee (Korea Advanced Institute of Science and Technology), Chanyoung Park (Korea Advanced Institute of Science and Technology)
ClassificationGraph Neural NetworkGraphBiomedical Data
🎯 What it does: This paper decouples hyperedges in hypergraph structures, proposes a naturality condition criterion based on category theory, constructs the Natural-HNN model, and verifies its effectiveness in classifying cancer subtypes based on gene pathways.
Disentangling Latent Shifts of In-Context Learning with Weak Supervision
Josip Jukić (University of Zagreb), Jan Šnajder (University of Zagreb)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes a method that views in-context learning (ICL) as weak supervision, encoding implicit shifts generated by demonstrations into lightweight adapters through a teacher-student framework, enabling promptless query reasoning and supporting large-scale demonstration combinations.
Disentangling misreporting from genuine adaptation in strategic settings: a causal approach
Dylan Zapzalka (University of Michigan), Maggie Makar (University of Michigan)
TabularFinance Related
🎯 What it does: This paper proposes a method based on causal inference to distinguish between agents' misreporting (upcoding) and true adaptation in strategic decision-making scenarios, and to identify and estimate the misreporting rate when only misreported features are observed.
Disentangling Superpositions: Interpretable Brain Encoding Model with Sparse Concept Atoms
Alicia Zeng (University of California), Jack L. Gallant
Explainability and InterpretabilityTextBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A Sparse Concept Encoding Model is proposed, which maps dense word embeddings to a higher-dimensional, sparse, and non-negative concept atom space through sparse dictionary learning, eliminating the unrecognizability caused by feature superposition and enhancing the interpretability of brain encoding models.
DisMo: Disentangled Motion Representations for Open-World Motion Transfer
Thomas Ressler-Antal (CompVis), Björn Ommer (CompVis)
GenerationRepresentation LearningTransformerDiffusion modelOptical FlowVideo
🎯 What it does: This paper proposes an unsupervised learning framework that utilizes only image reconstruction objectives to learn appearance-independent abstract motion representations, applying them as conditional inputs for cross-category, cross-view video motion transfer and zero-shot action classification.
DIsoN: Decentralized Isolation Networks for Out-of-Distribution Detection in Medical Imaging
Felix Wagner (University of Oxford), Konstantinos Kamnitsas (University of Oxford)
Anomaly DetectionFederated LearningConvolutional Neural NetworkImageBiomedical DataUltrasound
🎯 What it does: A decentralized isolation network (DIsoN) and its class-conditional variant CC-DIsoN are proposed for detecting outlier samples in medical images through model parameter exchange without sharing training data.
Distance Adaptive Beam Search for Provably Accurate Graph-Based Nearest Neighbor Search
Yousef Al-Jazzazi (New York University), Torsten Suel (New York University)
RetrievalOptimizationGraph Neural NetworkGraphBenchmark
🎯 What it does: This paper proposes a distance-threshold-based adaptive Beam Search termination strategy to replace the traditional fixed-width termination rules, thereby achieving more efficient nearest neighbor retrieval in graph search.
Distance-informed Neural Processes
Aishwarya Venkataramanan (Friedrich Schiller University Jena), Joachim Denzler (Friedrich Schiller University Jena)
ClassificationAnomaly DetectionTabularTime Series
🎯 What it does: This paper proposes a Distance-Aware Neural Process (DNP) that improves uncertainty estimation and generalization by incorporating a double Lipschitz constraint into both global and local latent variables.
Distances for Markov chains from sample streams
Sergio Calo (Universitat Pompeu Fabra), Javier Segovia-Aguas (Universitat Pompeu Fabra)
OptimizationSequentialStochastic Differential Equation
🎯 What it does: A stochastic optimization method based on sample flow to estimate the similarity between Markov chains is proposed, particularly for estimating bisimulation metrics.
Distil-E2D: Distilling Image-to-Depth Priors for Event-Based Monocular Depth Estimation
Jie Long Lee (National University of Singapore), Gim Hee Lee (National University of Singapore)
Depth EstimationAutonomous DrivingKnowledge DistillationTransformerImageMultimodality
🎯 What it does: A framework named Distil-E2D has been developed, which distills dense depth prior knowledge from image-based foundational depth models into the event domain. It achieves monocular depth estimation on event cameras by utilizing synthetic pseudo-labels, confidence-guided calibration loss, context transformers, and dual-decoder training.
Distillation Robustifies Unlearning
Bruce W. Lee (University of Pennsylvania), Alexander Matt Turner (University of Pennsylvania)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes a method to enhance the robustness of unlearning in large language models (LLMs) through distillation, demonstrating that distillation can suppress the relearning of forgotten knowledge, and introduces the UNDO method based on noise perturbation to achieve a computable and robust unlearning.
Distilled Decoding 2: One-step Sampling of Image Auto-regressive Models with Conditional Score Distillation
Enshu Liu (Tsinghua University), Yu Wang (Tsinghua University)
GenerationComputational EfficiencyKnowledge DistillationTransformerScore-based ModelImage
🎯 What it does: Proposes Distilled Decoding 2 (DD2), achieving one-step sampling for image autoregressive models;
Distilling LLM Agent into Small Models with Retrieval and Code Tools
Minki Kang (KAIST), Sung Ju Hwang (KAIST)
RetrievalKnowledge DistillationTransformerLarge Language ModelTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: A framework called Agent Distillation is proposed by transferring the interactive reasoning and tool usage behavior of large language models to smaller models;
Distilling LLM Prior to Flow Model for Generalizable Agent’s Imagination in Object Goal Navigation
Badi Li (The University of Hong Kong), Wei-Shi Zheng (Sun Yat-sen University)
Knowledge DistillationRobotic IntelligenceTransformerLarge Language ModelPrompt EngineeringFlow-based ModelPoint Cloud
🎯 What it does: A flow matching-based generative model called GOAL is proposed to perform semantic reasoning in unobserved areas for object goal navigation (ObjectNav), assisting agents in planning paths.
Distortion of AI Alignment: Does Preference Optimization Optimize for Preferences?
Paul Gölz (Cornell University), Kunhe Yang (University of California Berkeley)
OptimizationReinforcement Learning from Human Feedback
🎯 What it does: This study investigates the distortion problem of AI alignment methods (RLHF, DPO, NLHF) under diverse user preferences and quantifies their performance in achieving average utility.
Distributed mediation analysis with communication efficiency
Shaomin Li (Beijing Jiaotong University)
Tabular
🎯 What it does: A distributed version of the Sobel and MaxP tests is proposed within a distributed framework for detecting mediation effects with high communication efficiency.
Distributed Multi-Agent Bandits Over Erdős-Rényi Random Networks
Jingyuan Liu (Nanjing University), Mengfan Xu (University of Massachusetts Amherst)
Recommendation SystemOptimizationFederated LearningGraph Neural NetworkReinforcement LearningTabular
🎯 What it does: This paper studies the multi-agent multi-armed bandit (MA-MAB) problem under a random Erdős–Rényi communication network generated by a fixed base graph, and proposes a successful elimination (GSE) algorithm combined with gossip communication.
Distribution Learning Meets Graph Structure Sampling
Arnab Bhattacharyya (University of Warwick), N. V. Vinodchandran (University of Nebraska Lincoln)
Graph Neural NetworkGraph
🎯 What it does: Study high-dimensional Bayesian networks and propose a framework that combines graph structure sampling with online learning to obtain a PAC learning algorithm for multi-class Bayesian networks.
Distribution-Aligned Decoding for Efficient LLM Task Adaptation
Senkang Hu (City University of Hong Kong), Yuguang Fang (City University of Hong Kong)
Domain AdaptationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This paper redefines the task adaptation of large language models as the alignment of output distributions and proposes the SVDecode method, which adjusts logits using task-specific steering vectors during the decoding phase, enabling task transfer without retraining weights.
Distribution-Aware Tensor Decomposition for Compression of Convolutional Neural Networks
Alper KALLE, mohamed Tamaazousti
CompressionConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a tensor decomposition compression method based on data distribution, which directly optimizes the low-rank representation of convolutional layers using distribution-aware norms, avoiding or reducing subsequent fine-tuning.
Distributional Adversarial Attacks and Training in Deep Hedging
Guangyi He (Imperial College London), Lukas Gonon (University of St. Gallen)
OptimizationAdversarial AttackTime SeriesFinance Related
🎯 What it does: This paper studies the robustness of traditional deep hedging strategies under distributional drift and proposes a distributed adversarial attack and training framework based on the Wasserstein ball, utilizing adversarial training to enhance the robustness of hedging strategies.
Distributional Autoencoders Know the Score
Andrej Leban (University of Michigan)
OptimizationRepresentation LearningAuto EncoderTabular
🎯 What it does: This paper proposes and theoretically proves that the Distributed Principal Component Autoencoder (DPA) can simultaneously achieve learning of data distribution (by aligning with data scores) and determination of latent space dimensions (through the conditional independence of additional latent variables) within a single model, providing precise geometric and information-theoretic properties.
Distributional LLM-as-a-Judge
Luyu Chen (Renmin University of China), Xu Chen (Renmin University of China)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: A distributed LLM-as-a-Judge framework is proposed, utilizing KL divergence, cross-entropy assistance, and adversarial training to ensure that the judgment distribution output by the model aligns with human evaluation distribution.
Distributional Training Data Attribution: What do Influence Functions Sample?
Bruno Kacper Mlodozeniec (University of Cambridge), Roger Baker Grosse (University of Toronto)
OptimizationExplainability and InterpretabilityData-Centric LearningTransformerDiffusion modelImageTabular
🎯 What it does: This paper proposes a Distributed Training Data Attribution (d-TDA) framework, which redefines data attribution using the distribution of randomly trained data, and proves that Influence Functions (IFs) are essentially distributed, allowing for efficient estimation of changes in the model output distribution after the removal of training samples without retraining.
Distributionally Robust Feature Selection
Maitreyi Swaroop (Carnegie Mellon University), Bryan Wilder (Carnegie Mellon University)
Tabular
🎯 What it does: This paper proposes a noise-smoothing based distributionally robust feature selection method aimed at maintaining model prediction performance across multiple subgroups.
Distributionally Robust Performative Optimization
Zhuangzhuang Jia (University of Illinois), Grani A. Hanasusanto (University of Illinois)
OptimizationTabularFinance Related
🎯 What it does: A distributionally robust performative optimization framework (DRPO) is proposed, along with an iterative algorithm—Repeated Robust Risk Minimization (RRRM)—that can solve optimal decisions in situations where decisions lead to distribution changes and the distribution is unknown.
Distributive Fairness in Large Language Models: Evaluating Alignment with Human Values
Hadi Hosseini (Penn State University), Samarth Khanna (Penn State University)
TransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: This paper evaluates the fairness of resource allocation by large language models (LLMs) and their consistency with human preferences. It systematically compares various fairness axioms (fairness, no envy, maximizing the minimum benefit) and efficiency metrics, and analyzes the impact of factors such as prompting methods and role settings on model performance.