NeurIPS 2025 Papers — Page 11
Conference on Neural Information Processing Systems · 5275 papers
Dataset Distillation for Pre-Trained Self-Supervised Vision Models
George Cazenavette (Massachusetts Institute of Technology), Vincent Sitzmann (Massachusetts Institute of Technology)
ClassificationRepresentation LearningData-Centric LearningContrastive LearningImage
🎯 What it does: This study investigates dataset distillation for training linear probes on pre-trained self-supervised visual models (such as CLIP, DINO-v2, etc.) and proposes the Linear Gradient Matching method, which synthesizes a single image per class to train high-performance linear classifiers.
Dataset Distillation of 3D Point Clouds via Distribution Matching
Jae-Young Yim (Ulsan National Institute of Science and Technology), Jae-Young Sim (Ulsan National Institute of Science and Technology)
Data SynthesisPose EstimationKnowledge DistillationPoint Cloud
🎯 What it does: This paper proposes a dataset distillation framework for 3D point clouds that jointly optimizes the geometric structure and pose of synthetic samples.
Datasets, Documents, and Repetitions: The Practicalities of Unequal Data Quality
Alex Fang (Stanford), Tom Gunter (Apple)
TransformerLarge Language ModelText
🎯 What it does: This study investigates the impact of data filtering, deduplication, and reuse (multi-round training) on model performance and computational costs in large-scale language model training. It explores how to enhance the efficiency of limited data utilization by adjusting training techniques (such as weight decay) and document-level sampling strategies.
DAWP: A framework for global observation forecasting via Data Assimilation and Weather Prediction in satellite observation space
Junchao Gong (Shanghai Jiao Tong University), LEI BAI (Shanghai AI Laboratory)
TransformerAuto EncoderMultimodalityTime Series
🎯 What it does: The DAWP framework is proposed, which initializes through the Artificial Intelligence Data Assimilation (AIDA) module and uses a Cross-Boundary Condition (CBC) transformer to achieve direct predictions from global satellite observations.
DBLoss: Decomposition-based Loss Function for Time Series Forecasting
Xiangfei Qiu (East China Normal University), Bin Yang (East China Normal University)
TransformerTime Series
🎯 What it does: A loss function called DBLoss based on time series decomposition is proposed, which uses EMA to decompose the predicted values and true values into seasonal and trend components within the prediction interval, and calculates the loss separately before summing them with weights;
DC4GS: Directional Consistency-Driven Adaptive Density Control for 3D Gaussian Splatting
Moonsoo Jeong (Sungkyunkwan University), Sungkil Lee (Sungkyunkwan University)
OptimizationComputational EfficiencyGaussian SplattingPoint Cloud
🎯 What it does: This paper proposes an adaptive density control method based on Direction Consistency (DC) called DC4GS, aimed at improving the original splitting and sub-primitive position selection in 3D Gaussian splatting, thereby reducing the number of Gaussians and enhancing rendering quality without increasing computational load.
DCA: Graph-Guided Deep Embedding Clustering for Brain Atlases
Mo wang, Quanying Liu (Southern University of Science and Technology)
SegmentationRepresentation LearningConvolutional Neural NetworkSupervised Fine-TuningBiomedical DataMagnetic Resonance ImagingAlzheimer's Disease
🎯 What it does: A graph-guided deep embedding clustering framework DCA is proposed for generating personalized and voxel-level brain region partitions.
DCI: Dual-Conditional Inversion for Boosting Diffusion-Based Image Editing
Zixiang Li (Beijing Jiaotong University), Yao Zhao (Beijing Jiaotong University)
GenerationOptimizationDiffusion modelImageBenchmark
🎯 What it does: A Dual-Conditional Inversion method is designed, which combines text prompts and reference images to guide the reverse process of the diffusion model, aiming to obtain more accurate and editable latent noise.
DEAL: Diffusion Evolution Adversarial Learning for Sim-to-Real Transfer
Wentao Xu (Nanjing University), Chunlin Chen (Nanjing University)
OptimizationRobotic IntelligenceReinforcement LearningDiffusion modelGenerative Adversarial NetworkMultimodality
🎯 What it does: Proposes the DEAL framework, which combines diffusion evolution with adversarial learning to iteratively optimize simulation parameters using limited real data, thereby achieving high-dimensional Sys-Id and sim-to-real transfer;
Debate or Vote: Which Yields Better Decisions in Multi-Agent Large Language Models?
Hyeong Kyu Choi (University of Wisconsin Madison), Sharon Li (University of Wisconsin Madison)
TransformerLarge Language ModelAgentic AIText
🎯 What it does: This paper studies Multi-Agent Debate (MAD) systems, breaking it down into two parts: voting and debate. It conducts large-scale experiments and theoretical analysis, proving that majority voting is the main source of improvement, while the debate itself does not increase accuracy in expectation, and proposes improvement strategies.
DeblurDiff: Real-Word Image Deblurring with Generative Diffusion Models
Lingshun Kong (Nanjing University of Science and Technology), Jinshan Pan (Harbin Institute of Technology)
RestorationDiffusion modelImage
🎯 What it does: This paper proposes DeblurDiff, which combines the latent space spatial variant convolution kernel prediction network LKPN with the pre-trained Stable Diffusion diffusion model, utilizing pixel-level adaptive convolution (EAC) and iterative kernel prediction to achieve real image deblurring.
DeCaFlow: A deconfounding causal generative model
Alejandro Almodóvar (Universidad Politécnica de Madrid), Isabel Valera (Saarland University)
GenerationData SynthesisFlow-based ModelAuto EncoderTabular
🎯 What it does: A causal generative model named DeCaFlow is proposed, which can accurately estimate causal interventions and counterfactual queries for continuous variables in the presence of hidden confounding, using only observational data and a known causal graph, after a single training session.
Decentralized Dynamic Cooperation of Personalized Models for Federated Continual Learning
Danni Yang (Tsinghua University), Mingming Gong (University of Melbourne)
Federated LearningKnowledge DistillationImage
🎯 What it does: This paper proposes a decentralized dynamic collaboration framework for personalized federated continual learning, allowing clients to dynamically form non-overlapping cooperative alliances at each task stage based on the trade-off between knowledge acquisition and retention of previous learning, thereby mitigating catastrophic forgetting.
Deciphering the Extremes: A Novel Approach for Pathological Long-tailed Recognition in Scientific Discovery
Zhe Zhao, Yang Wang
ClassificationRecognitionOptimizationConvolutional Neural NetworkGraph Neural NetworkContrastive LearningImageGraph
🎯 What it does: This paper proposes an end-to-end framework for pathological long-tail distribution data in scientific discovery, aiming to enhance the recognition performance of rare tail classes.
Decoder-Hybrid-Decoder Architecture for Efficient Reasoning with Long Generation
Liliang Ren (Microsoft), yelong shen
GenerationComputational EfficiencyTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes the Gated Memory Unit (GMU) and designs the SambaY decoder-hybrid-decoder architecture based on it to significantly improve the decoding efficiency of long sequence inference.
Decoding Causal Structure: End-to-End Mediation Pathways Inference
Yulong Li (Mohamed bin Zayed University of Artificial Intelligence), Eran Segal (Weizmann Institute of Science)
Flow-based ModelAuto EncoderMultimodalityBiomedical Data
🎯 What it does: An end-to-end causal mediation analysis framework SIGMA has been designed and implemented, integrating structure learning, path identification, and effect estimation.
DecompNet: Enhancing Time Series Forecasting Models with Implicit Decomposition
Donghao Luo (Tsinghua University), Xue Wang (Tsinghua University)
TransformerMixture of ExpertsTime Series
🎯 What it does: The concept of Implicit Decomposition is proposed, and based on this idea, the DecompNet framework is constructed, which can inject seasonal and trend knowledge into the model without explicitly decomposing the input sequence, achieving performance improvement with no additional inference cost.
Decomposing Interventional Causality into Synergistic, Redundant, and Unique Components
Abel Jansma (Dutch Institute for Emergent Phenomena, University of Amsterdam)
TransformerText
🎯 What it does: This paper proposes a framework that decomposes intervention causal effects into synergistic, redundant, and unique components, extending partial information decomposition (PID) to the causal intervention domain.
Decomposing motor units through elimination for real-time intention driven assistive neurotechnology
Nicholas Tacca (Battelle Memorial Institute), David A. Friedenberg (Battelle Memorial Institute)
Spiking Neural NetworkBiomedical Data
🎯 What it does: The MUelim algorithm is proposed and implemented for real-time high-density EMG motor unit decomposition, and its effectiveness is validated in the control of neural prosthetics for clinical SCI patients.
Decomposing stimulus-specific sensory neural information via diffusion models
Steeve Laquitaine (Institut de la Vision), Matthew Chalk (Institut de la Vision)
Diffusion modelImage
🎯 What it does: A new information decomposition method is proposed to understand how neurons encode stimuli, particularly how mutual information can be decomposed into contributions from individual stimuli and their features.
Decoupled Entropy Minimization
Jing Ma (Huazhong University of Science and Technology), Xiang Xiang (Huazhong University of Science and Technology)
Domain AdaptationReinforcement LearningImage
🎯 What it does: A new self-supervised entropy minimization method called AdaDEM is proposed, which decouples and improves traditional entropy minimization, addressing the issues of reward collapse and class bias.
Decoupling Contrastive Decoding: Robust Hallucination Mitigation in Multimodal Large Language Models
Wei Chen (Hong Kong University of Science and Technology), Long Chen (Hong Kong University of Science and Technology)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningVision Language ModelContrastive LearningMultimodality
🎯 What it does: A decoupled contrastive decoding framework (DCD) is proposed, which suppresses the hallucinations of multimodal large models by separately learning the projections of positive and negative samples and comparing positive and negative visual features during inference.
Decreasing Entropic Regularization Averaged Gradient for Semi-Discrete Optimal Transport
Ferdinand Genans (Sorbonne Université), Olivier Wintenberger (Sorbonne Université)
OptimizationPoint Cloud
🎯 What it does: This paper proposes and analyzes the DRAG algorithm, which uses a stepwise decreasing entropy regularization method for the average gradient descent in the semi-discrete optimal transport problem.
Deep Compositional Phase Diffusion for Long Motion Sequence Generation
Ho Yin Au (Hong Kong Baptist University), Jingyu Xiang (Hong Kong Baptist University)
GenerationData SynthesisTransformerDiffusion modelAuto EncoderVideoTextSequential
🎯 What it does: A scalable 'Compositional Phase Diffusion' framework is proposed to generate long sequences, composite, and intermediate human actions, achieving smooth transitions between segments.
Deep Continuous-Time State-Space Models for Marked Event Sequences
Yuxin Chang (University of California), Andrew Warrington (GE HealthCare)
Time SeriesSequentialBiomedical DataElectronic Health Records
🎯 What it does: A continuous-time marked event sequence model S2P2 based on a deep state space model is proposed to capture the strong correlations and long-range dependencies of events at irregular time points;
Deep Edge Filter: Return of the Human-Crafted Layer in Deep Learning
Dongkwan Lee (Seoul National University), Nojun Kwak (Seoul National University)
Domain AdaptationConvolutional Neural NetworkTransformerImageTextMultimodalityAudio
🎯 What it does: Proposes the Deep Edge Filter, which utilizes the high-frequency components extracted from deep features after low-pass filtering to enhance the model's generalization ability.
Deep Gaussian from Motion: Exploring 3D Geometric Foundation Models for Gaussian Splatting
Yu Chen (National University of Singapore), Gim Hee Lee (National University of Singapore)
GenerationPose EstimationTransformerGaussian SplattingSimultaneous Localization and MappingImage
🎯 What it does: DeepGfM reconstructs neural fields from unordered image collections without relying on SfM, and utilizes a pre-trained 3D geometric foundation model to predict 3D Gaussian primitives and camera poses.
Deep learning for continuous-time stochastic control with jumps
Patrick Cheridito (ETH Zurich), Donatien Hainaut (UCLouvain)
OptimizationReinforcement LearningTabularFinance RelatedPhysics RelatedStochastic Differential Equation
🎯 What it does: This paper proposes two iterative algorithms based on neural networks (GPI-PINN and GPI-CBU) for solving finite time slot, jump-inclusive continuous-time stochastic control problems, and provides approximations for the global value function and optimal control.
Deep Learning with Plausible Deniability
Wenxuan Bao (University of Florida), Yiwei Cai (Visa Research)
OptimizationSafty and PrivacyTransformerSupervised Fine-TuningImageText
🎯 What it does: A suspicious denial-style stochastic gradient descent algorithm PD-SGD is proposed, which utilizes noise injection and privacy testing to perform rejection sampling on gradients, thereby suppressing membership inference attacks during training.
Deep Legendre Transform
Aleksey Minabutdinov (ETH Zurich), Patrick Cheridito (ETH Zurich)
OptimizationConvolutional Neural Network
🎯 What it does: Proposes the Deep Legendre Transform (DLT) method, which uses deep learning to compute the Legendre transform (convex conjugate) of differentiable convex functions.
Deep RL Needs Deep Behavior Analysis: Exploring Implicit Planning by Model-Free Agents in Open-Ended Environments
Riley Simmons-Edler (Harvard University), Kanaka Rajan (Harvard University)
Recurrent Neural NetworkReinforcement LearningSequential
🎯 What it does: This paper presents ForageWorld, a naturalistic open-ended reinforcement learning environment, and combines tools from neuroscience and animal behavior to jointly analyze the behavior and internal representations of deep RL agents, revealing their implicit planning and memory capabilities.
Deep Taxonomic Networks for Unsupervised Hierarchical Prototype Discovery
Zekun Wang (Georgia Institute of Technology), Christopher J. MacLellan
Representation LearningAuto EncoderContrastive LearningImage
🎯 What it does: This paper proposes a Deep Taxonomic Network, which achieves unsupervised hierarchical clustering by using a complete binary tree mixture of Gaussian priors within a VAE framework, and automatically learns prototypes at each level.
Deep Tree Tensor Networks
Chang Nie (Nanjing University of Science and Technology)
ClassificationRecommendation SystemOptimizationConvolutional Neural NetworkTransformerImagePhysics Related
🎯 What it does: A deep tree tensor network (DTTN) is designed and implemented, which achieves exponential-order feature interaction through stacking anti-symmetric interaction modules (AIM) without the need for activation functions, and is applied to image classification, recommendation, and solving physical problems.
Deep Value Benchmark: Measuring Whether Models Generalize Deep values or Shallow Preferences
Joshua Ashkinaze (University of Michigan), Ceren Budak (University of Michigan)
Large Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: This study investigates whether large language models learn deep human values or merely memorize surface preferences, and proposes the Deep Value Benchmark (DVB) to directly measure the models' generalization ability regarding deep values.
Deep Video Discovery: Agentic Search with Tool Use for Long-form Video Understanding
Xiaoyi Zhang (Microsoft Research), Yan Lu (Microsoft Research)
RecognitionRetrievalTransformerLarge Language ModelAgentic AIVision Language ModelVideoText
🎯 What it does: Proposed and implemented the Deep Video Discovery (DVD) agent, which utilizes a multi-granularity video database and toolchain to achieve autonomous search and reasoning for long videos.
DeepASA: An Object-Oriented Multi-Purpose Network for Auditory Scene Analysis
Dongheon Lee (Korea Advanced Institute of Science and Technology), Jung-Woo Choi (Korea Advanced Institute of Science and Technology)
ClassificationRecognitionTransformerAudio
🎯 What it does: A unified framework DeepASA is proposed, compatible with multi-tasks such as source separation, echo cancellation, sound event detection, audio classification, and direction estimation;
DeepDiver: Adaptive Web-Search Intensity Scaling via Reinforcement Learning
Wenxuan Shi (Huawei), Lifeng Shang (Huawei)
RetrievalTransformerLarge Language ModelReinforcement LearningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This work studies the information-seeking ability of large language models in real-time web search environments, proposing the WebPuzzle dataset and the DeepDiver framework based on reinforcement learning;
Deeper with Riemannian Geometry: Overcoming Oversmoothing and Oversquashing for Graph Foundation Models
Li Sun (Beijing University of Posts and Telecommunications), Philip S. Yu (University of Illinois)
ClassificationGraph Neural NetworkGraph
🎯 What it does: Proposes a non-homogeneous boundary condition based on local Riemannian geometry to simultaneously address the issues of over-smoothing and over-compression in graph message passing networks;
DeepHalo: A Neural Choice Model with Controllable Context Effects
Shuhan Zhang (Chinese University of Hong Kong), Shuang Li (Chinese University of Hong Kong)
Recommendation SystemExplainability and InterpretabilityTabular
🎯 What it does: The DeepHalo framework is proposed for interpretable context-aware choice modeling, supporting controllable interaction orders;
DeepKD: A Deeply Decoupled and Denoised Knowledge Distillation Trainer
Haiduo Huang (Xi'an Jiaotong University), Pengju Ren (Xi'an Jiaotong University)
Object DetectionKnowledge DistillationImage
🎯 What it does: A deep decoupling and denoising knowledge distillation training framework, DeepKD, is designed, which achieves independent optimization of task gradients, target class gradients, and non-target class gradients through GSNR-driven momentum allocation and dynamic top-k masking, while filtering low-confidence dark knowledge.
DeepVideo-R1: Video Reinforcement Fine-Tuning via Difficulty-aware Regressive GRPO
Jinyoung Park (Korea Advanced Institute of Science and Technology), Hyunwoo J. Kim (Korea Advanced Institute of Science and Technology)
Large Language ModelReinforcement LearningVideoBenchmark
🎯 What it does: A video large language model DeepVideo-R1 is proposed, which uses regression GRPO (Reg-GRPO) and difficulty-aware data augmentation for video reinforcement fine-tuning to enhance video reasoning capabilities.
Defending Multimodal Backdoored Models by Repulsive Visual Prompt Tuning
Zhifang Zhang (Southeast University), Lei Feng (Southeast University)
ClassificationRecognitionAdversarial AttackTransformerPrompt EngineeringContrastive LearningImage
🎯 What it does: A rejection-based training method is proposed, utilizing a small number of clean samples to defend against backdoor-infected CLIP models.
Deferring Concept Bottleneck Models: Learning to Defer Interventions to Inaccurate Experts
Andrea Pugnana (University of Trento), Davide Bacciu (University of Pisa)
Explainability and InterpretabilityConvolutional Neural NetworkImage
🎯 What it does: Designed and trained Deferring Concept Bottleneck Models (DCBMs), enabling traditional concept bottleneck models to automatically determine when to request intervention from human experts and when to make predictions independently, thus achieving interpretable human-machine collaboration.
Defining and Discovering Hyper-meta-paths for Heterogeneous Hypergraphs
Yaming Yang (Xidian University), Ziyu Guan (Xidian University)
Recommendation SystemExplainability and InterpretabilityRepresentation LearningGraph Neural NetworkGraph
🎯 What it does: Proposes and uses hyper-meta-paths for representation learning on heterogeneous hypergraphs.
DEFT: Decompositional Efficient Fine-Tuning for Text-to-Image Models
Komal Kumar (Mohamed bin Zayed University of Artificial Intelligence), Hisham Cholakkal (Mohamed bin Zayed University of Artificial Intelligence)
GenerationData SynthesisTransformerSupervised Fine-TuningImageText
🎯 What it does: Designed and implemented DEFT, an efficient fine-tuning framework that splits the weight updates of pre-trained text-to-image models into two parts: projection to a low-rank subspace and low-rank updates, supporting tasks such as personalization, general generation, and multi-concept combinations.
DEGauss: Defending Against Malicious 3D Editing for Gaussian Splatting
Lingzhuang Meng (China University of Petroleum), Xiang Lv (China University of Petroleum)
RestorationAdversarial AttackDiffusion modelGaussian SplattingPoint Cloud
🎯 What it does: A 3D defense framework named DEGauss is proposed, which defends against malicious edits based on Gaussian Splatting by adding small perturbations to 3D Gaussian point clouds.
Degradation-Aware Dynamic Schrödinger Bridge for Unpaired Image Restoration
Jingjun Yi (University of Alberta), Yefeng Zheng (Westlake University)
RestorationConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes a framework for unpaired image restoration based on the Schrödinger Bridge, called DDSB. It utilizes degradation-aware dynamic optimal transport (OT) and consistency constraints, introducing a degradation model at different stages to achieve a gradual transition from the degraded domain to the clear domain.
Degrees of Freedom for Linear Attention: Distilling Softmax Attention with Optimal Feature Efficiency
Naoki Nishikawa (University of Tokyo), Taiji Suzuki (University of Tokyo)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelText
🎯 What it does: Proposes an adaptive linear attention feature dimension selection method based on statistical degrees of freedom, combined with hierarchical training to learn nonlinear features, aimed at distilling the softmax attention of a pre-trained Transformer into an efficient linear attention model.
Deliberation on Priors: Trustworthy Reasoning of Large Language Models on Knowledge Graphs
Jie Ma (Xi'an Jiaotong University), su zhou
RetrievalOptimizationKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningGraph
🎯 What it does: A trustworthy reasoning framework DP is proposed, which enhances the answer credibility of large language models in KG semantic retrieval and multi-hop reasoning by utilizing the structural prior and constraint prior of knowledge graphs.
Delta Attention: Fast and Accurate Sparse Attention Inference by Delta Correction
Jeffrey Willette (KAIST), Sung Ju Hwang (KAIST)
TransformerLarge Language ModelText
🎯 What it does: A post-processing method called Δ Attention is proposed to correct the distribution shift caused by sparse attention inference, thereby improving the accuracy of long sequence reasoning.
DeltaFlow: An Efficient Multi-frame Scene Flow Estimation Method
Qingwen Zhang (KTH Royal Institute of Technology), Patric Jensfelt (Linköping University)
Autonomous DrivingOptimizationComputational EfficiencyOptical FlowPoint Cloud
🎯 What it does: A lightweight multi-frame scene flow estimation framework called DeltaFlow is proposed;
DeltaFormer: Unlock the state space of Transformer
Mingyu Xu, Shu Zhong
TransformerText
🎯 What it does: The DeltaFormer model is proposed, which combines the Delta rule with kernel functions to construct a model that can overcome the representation limitations of Transformer TC0.
DeltaPhi: Physical States Residual Learning for Neural Operators in Data-Limited PDE Solving
Xihang Yue (Zhejiang University), Linchao Zhu (Zhejiang University)
OptimizationData-Centric LearningTabularPhysics Related
🎯 What it does: This paper proposes a framework called DeltaPhi that transforms the task of solving PDEs from direct mapping to learning the residuals of similar physical states. It utilizes residual learning to achieve implicit data augmentation, significantly enhancing the performance of neural operators in data-scarce scenarios.
DeltaProduct: Improving State-Tracking in Linear RNNs via Householder Products
Julien Siems (University of Freiburg), Riccardo Grazzi (Istituto Italiano di Tecnologia)
Recurrent Neural NetworkTextSequential
🎯 What it does: DeltaProduct is proposed, which extends DeltaNet by performing multi-step gradient descent on each token, resulting in a state transition matrix composed of several Householder transformations, enhancing the representational capacity of linear RNNs.
Delving into Cascaded Instability: A Lipschitz Continuity View on Image Restoration and Object Detection Synergy
Qing Zhao (Sun Yat-sen University), Liang Lin
RestorationObject DetectionImage
🎯 What it does: The study focuses on the synergy between image restoration and object detection in harsh environments, and proposes a Lipschitz regularization framework to alleviate the instability caused by the functional mismatch between the two.
Delving into Large Language Models for Effective Time-Series Anomaly Detection
Junwoo Park (KAIST), Jaewoong Cho (KAIST)
Anomaly DetectionTransformerLarge Language ModelPrompt EngineeringTime Series
🎯 What it does: This paper discusses the shortcomings of large language models (LLMs) in time series anomaly detection (TSAD) and proposes a zero-shot method that combines statistical decomposition with index-aware prompts.
Delving into RL for Image Generation with CoT: A Study on DPO vs. GRPO
Chengzhuo Tong (Shanghai AI Lab), Pheng-Ann Heng (Chinese University of Hong Kong)
GenerationReinforcement LearningImageChain-of-Thought
🎯 What it does: The paper systematically compares and analyzes the in-domain and out-of-domain performance of GRPO and DPO in autoregressive image generation tasks, and studies the impact of reward models and scaling strategies on the two RL algorithms.
Democratizing Clinical Risk Prediction with Cross-Cohort Cross-Modal Knowledge Transfer
Qiannan Zhang (Weill Cornell Medicine), Fei Wang (Weill Cornell Medicine)
Domain AdaptationKnowledge DistillationGraph Neural NetworkTransformerMultimodalityBiomedical DataAlzheimer's DiseaseElectronic Health Records
🎯 What it does: A cross-queue and cross-modal knowledge transfer framework, C M3, is proposed, enabling models trained on national multimodal data to be transferred to local clinical environments with only EHR.
Demystifying Language Model Forgetting with Low-rank Example Associations
Xisen Jin (University of Southern California), Xiang Ren (University of Southern California)
Large Language ModelSupervised Fine-TuningText
🎯 What it does: This paper analyzes the relationship between upstream samples lost during fine-tuning of LLMs and new tasks, and discovers its simplicity through low-rank matrix approximation.
Demystifying Reasoning Dynamics with Mutual Information: Thinking Tokens are Information Peaks in LLM Reasoning
Chen Qian (Renmin University of China), Jing Shao (Shanghai Artificial Intelligence Laboratory)
TransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: This study investigates the internal reasoning processes of large reasoning models, proposing the phenomenon of mutual information peak and proving its correlation with reasoning accuracy.
Demystifying Spectral Feature Learning for Instrumental Variable Regression
Dimitri Meunier (University College London), Arthur Gretton (University College London)
Contrastive LearningImageBenchmark
🎯 What it does: This paper studies nonparametric instrumental variable regression (NPIV) using spectral features in the presence of hidden confounding variables and provides a generalization error upper bound. It reveals the impact of two major factors, spectral alignment and singular value decay, on estimation performance, and proposes a classification of three scenarios: 'good/bad/ugly'. It also presents a spectral feature learning method based on contrastive learning and provides practical steps for estimating spectral features and alignment from data. The theory and methods are validated through synthetic experiments and the dSprites dataset.
Dendritic Resonate-and-Fire Neuron for Effective and Efficient Long Sequence Modeling
Dehao Zhang (University of Electronic Science and Technology of China), Haizhou Li (Chinese University of Hong Kong)
Spiking Neural NetworkSequential
🎯 What it does: This paper proposes multi-branch oscillatory discharge (D-RF) spiking neurons for efficient and sparse long sequence modeling.
Deno-IF: Unsupervised Noisy Visible and Infrared Image Fusion Method
Han Xu (Southeast University), Guangcan Liu (Southeast University)
RestorationTransformerImage
🎯 What it does: Proposes an unsupervised visible and infrared image fusion method Deno-IF, which unifies denoising and fusion;
DenoiseRotator: Enhance Pruning Robustness for LLMs via Importance Concentration
Tianteng Gu (Shanghai Jiao Tong University), Yanmin Qian (Shanghai Jiao Tong University)
TransformerLarge Language ModelText
🎯 What it does: A pruning preconditioning method called DenoiseRotator is proposed, which centralizes parameter importance through learning orthogonal transformations, enhancing the robustness of LLM pruning.
Denoising Trajectory Biases for Zero-Shot AI-Generated Image Detection
Yachao Liang (Institute of Information Engineering Chinese Academy of Sciences), Weiqing Huang (Institute of Information Engineering Chinese Academy of Sciences)
ClassificationRecognitionGenerationDiffusion modelGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes an unsupervised (zero-shot) AI-generated image detection method that utilizes the denoising trajectory during the reverse sampling process of diffusion models to distinguish between real images and synthetic images.
Dense Associative Memory with Epanechnikov Energy
Benjamin Hoover (IBM Research), Parikshit Ram (IBM Research)
GenerationData SynthesisImage
🎯 What it does: A new energy function LSR (log-sum-ReLU) is proposed for dense associative memory networks, achieving exponential memory capacity while generating new 'emergent memories' while maintaining perfect recall.
Dense Backpropagation Improves Training for Sparse Mixture-of-Experts
Ashwinee Panda (University of Maryland), Supriyo Chakraborty (Capital One)
OptimizationComputational EfficiencyTransformerMixture of ExpertsText
🎯 What it does: A new sparse mixture of experts (MoE) training method called Default MoE is proposed, which generates default outputs for inactive experts, allowing the router to receive gradients from all experts during training, thereby improving the model's convergence speed and final performance.
Dense Metric Depth Estimation via Event-based Differential Focus Volume Prompting
Boyu Li (Peking University), Boxin Shi (Peking University)
Depth EstimationConvolutional Neural NetworkImage
🎯 What it does: By combining events triggered by focal sweeping and images, an Event Differential Focal Volume (EDFV) is constructed to predict sparse depth, which is then integrated with a single image foundation model (IFM) through a prompting network to obtain high-quality full dense metric depth.
Dense SAE Latents Are Features, Not Bugs
Xiaoqing Sun (Massachusetts Institute of Technology), Max Tegmark (Massachusetts Institute of Technology)
Auto EncoderText
🎯 What it does: This paper systematically studies the dense latent variables generated by sparse autoencoders (SAE), proving that they are not noise but capture meaningful information in the model residual flow.
DenseDPO: Fine-Grained Temporal Preference Optimization for Video Diffusion Models
Ziyi Wu (Snap Research), Aliaksandr Siarohin (Snap Research)
GenerationRecommendation SystemOptimizationVision Language ModelDiffusion modelRectified FlowVideoBenchmark
🎯 What it does: Proposes DenseDPO, a fine-grained temporal preference optimization method for video diffusion models.
Density Ratio-Free Doubly Robust Proxy Causal Learning
Bariscan Bozkurt (University College London), Arthur Gretton (DeepMind)
ImageTabular
🎯 What it does: Two types of doubly robust kernel estimators (DRKPV and DRPMMR) are proposed within the framework of Proxy Causal Learning, capable of estimating dose-response curves without the need for explicit density ratio estimation, suitable for continuous or high-dimensional treatment variables.
DePass: Unified Feature Attributing by Simple Decomposed Forward Pass
Xiangyu Hong (Tsinghua University), Bowen Zhou (Tsinghua University)
GenerationExplainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: DePass is proposed, a method for achieving interpretability of the Transformer mechanism through single decomposition forward propagation.
Dependency Matters: Enhancing LLM Reasoning with Explicit Knowledge Grounding
Xiangyu Wen (Chinese University of Hong Kong), Qiang Xu (Chinese University of Hong Kong)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A reasoning framework called GRiD based on a knowledge dependency graph is proposed, which explicitly extracts knowledge from the internal knowledge of LLMs and constructs a graph structure of knowledge and reasoning nodes, combined with a lightweight verifier to ensure logical consistency at each step.
Dependency Parsing is More Parameter-Efficient with Normalization
Paolo Gajo (University of Bologna), Alberto Barrón-Cedeño (University of Bologna)
Recurrent Neural NetworkGraph
🎯 What it does: The study and verification of normalizing (scaling) the biaffine scores in dependency parsing can improve model performance and reduce parameters.
Deployment Efficient Reward-Free Exploration with Linear Function Approximation
Zihan Zhang (Hong Kong University of Science and Technology), Ruosong Wang (Peking University)
OptimizationReinforcement Learning
🎯 What it does: A linear MDP exploration algorithm is designed that can complete learning and return an approximately optimal policy during the planning phase with only H deployments, without relying on reward information.
Depth-Bounds for Neural Networks via the Braid Arrangement
Moritz Leo Grillo (Max Planck Institute for Mathematics in the Sciences), Georg Loho (Freie Universität Berlin)
🎯 What it does: This paper studies the minimum hidden layer depth required to implement continuous piecewise linear functions, particularly the maximum function, in B0_d-conforming ReLU and maxout networks (i.e., where breakpoints are only located on the braid fan), proving that at least Ω(log log d) layers are needed.
Depth-Supervised Fusion Network for Seamless-Free Image Stitching
Zhiying Jiang (Dalian Maritime University), Jinyuan Liu (Dalian University of Technology)
Image TranslationDepth EstimationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a deep supervision-based image stitching method that achieves precise alignment and seamless stitching in large parallax scenes through two-stage depth-aware transformation estimation and flexible seam fusion.
Depth-Width Tradeoffs for Transformers on Graph Tasks
Gilad Yehudai (New York University), Amir Globerson (Google Research)
TransformerGraph
🎯 What it does: This study investigates the trade-off between depth and width of Transformers in graph tasks, proving that linear width can achieve various graph algorithms, with the option to increase width to quadratic when necessary.
DepthVanish: Optimizing Adversarial Interval Structures for Stereo-Depth-Invisible Patches
Yun Xing (University of Alberta), Qing Guo (Agency for Science, Technology and Research)
Depth EstimationAutonomous DrivingOptimizationAdversarial AttackImage
🎯 What it does: This paper proposes an attack method for stereo depth estimation systems that can be deployed in physical environments—DepthVanish, which creates a 'vanishing' attack by jointly optimizing texture elements and spacing structures;
DERD-Net: Learning Depth from Event-based Ray Densities
Diego de Oliveira Hitzges, Guillermo Gallego (Technische Universität Berlin)
Depth EstimationAutonomous DrivingConvolutional Neural NetworkRecurrent Neural NetworkPoint Cloud
🎯 What it does: Using an event camera and camera pose to generate a 3D ray density map (DSI), the lightweight deep learning network DERD-Net predicts pixel-level depth, supporting both monocular and binocular scenes.
Derivative-Free Guidance in Continuous and Discrete Diffusion Models with Soft Value-based Decoding
Xiner Li (Texas A&M University), Masatoshi Uehara (EvolutionaryScale)
GenerationOptimizationReinforcement LearningDiffusion modelImageSequential
🎯 What it does: A no fine-tuning, non-differentiable inference algorithm SVDD is proposed, which guides the diffusion model with a soft value function to optimize downstream rewards.
Design-Based Bandits Under Network Interference: Trade-Off Between Regret and Statistical Inference
Zichen Wang (University of Illinois Urbana-Champaign), Huazheng Wang (Oregon State University)
OptimizationReinforcement Learning from Human FeedbackGraph
🎯 What it does: This paper studies the multi-armed bandit problem with network interference (MABNI), constructs the theoretical Pareto front in adversarial environments, and proposes the EXP3-N-CS algorithm that can achieve continuous inference while reducing cumulative regret.
DesignX: Human-Competitive Algorithm Designer for Black-Box Optimization
Hongshu Guo (South China University of Technology), Yue-Jiao Gong (South China University of Technology)
OptimizationTransformerLarge Language ModelReinforcement LearningTabular
🎯 What it does: DesignX is proposed, an automatic black-box optimizer design framework based on dual-agent reinforcement learning, capable of generating human-level optimizers for any black-box problem within seconds.
Detecting Data Deviations in Electronic Health Records
Kaiping Zheng (National University of Singapore), Beng Chin Ooi (Zhejiang University)
Anomaly DetectionKnowledge DistillationBiomedical DataElectronic Health Records
🎯 What it does: A task-agnostic EHR data reliability predictor is constructed using dual-layer knowledge distillation to detect record deviations in real-time.
Detecting Generated Images by Fitting Natural Image Distributions
Yonggang Zhang (Hong Kong University of Science and Technology), Bo Han (Hong Kong Baptist University)
Anomaly DetectionTransformerFlow-based ModelContrastive LearningImageVideo
🎯 What it does: A generative image detection framework called ConV is proposed, which verifies the consistency of natural image distributions. It constructs two functions to achieve consistency detection using the gradient orthogonality principle of self-supervised models, and actively amplifies the distribution differences between generated images and natural images through regularization flow.
Detecting High-Stakes Interactions with Activation Probes
Alex McKenzie (LASR Labs), Dmitrii Krasheninnikov (University of Cambridge)
Anomaly DetectionComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This paper proposes and evaluates activation probes for detecting high-stakes situations in interactions with large language models, and validates their feasibility in real-world scenarios.
Detoxifying Large Language Models via Autoregressive Reward Guided Representation Editing
Yisong Xiao (Beihang University), Dacheng Tao (Nanyang Technological University)
Representation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: A testing-time detoxification method called ARGRE is proposed, which utilizes toxic transfer paths in the representation space to achieve dense supervision of sparse toxicity annotations, and controls the toxicity of LLM-generated content through an autoregressive reward model and a two-step representation editing.
DETree: DEtecting Human-AI Collaborative Texts via Tree-Structured Hierarchical Representation Learning
Yongxin He (Institute of Computing Technology, Chinese Academy of Sciences), Ping Luo (Institute of Computing Technology, Chinese Academy of Sciences)
ClassificationRecognitionRepresentation LearningContrastive LearningText
🎯 What it does: This paper proposes a framework called DETree, aimed at identifying text generated through human-machine collaboration.
DevFD : Developmental Face Forgery Detection by Learning Shared and Orthogonal LoRA Subspaces
Tianshuo Zhang (University of Chinese Academy of Sciences), Zhen Lei (Chinese Academy of Sciences)
RecognitionAnomaly DetectionTransformerMixture of ExpertsImage
🎯 What it does: A development-oriented Mixture of Experts framework, DevFD, has been constructed, utilizing the LoRA subspace for continual learning to achieve facial forgery detection.
DexFlyWheel: A Scalable and Self-improving Data Generation Framework for Dexterous Manipulation
Kefei Zhu (Harbin Institute of Technology), Yuanpei Chen (Peking University)
Robotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningMultimodality
🎯 What it does: A scalable and self-improving data generation framework, DexFlyWheel, has been constructed to generate diverse and high-quality multi-finger robot grasping and manipulation data with minimal human demonstrations.
DexGarmentLab: Dexterous Garment Manipulation Environment with Generalizable Policy
Yuran Wang (Peking University), Hao Dong (Peking University)
Robotic IntelligenceReinforcement LearningPoint Cloud
🎯 What it does: Proposed the DexGarmentLab environment, an automated data collection pipeline, and a general policy HALO to achieve high-dimensional and generalizable control over two-handed grasping and manipulation of garments.
DEXTER: Diffusion-Guided EXplanations with TExtual Reasoning for Vision Models
Simone Carnemolla (University of Catania), Concetto Spampinato (University of Catania)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringDiffusion modelImageText
🎯 What it does: The DEXTER framework is proposed, utilizing diffusion models and large language models to generate global text explanations in the absence of data, aimed at revealing the decision mechanisms and biases of visual classifiers.
DGH: Dynamic Gaussian Hair
Junying Wang (University of Southern California), Tony Tung (Meta Reality Labs Research)
GenerationData SynthesisConvolutional Neural NetworkGaussian SplattingImageVideo
🎯 What it does: A dynamic Gaussian model-based hairstyle simulation framework (DGH) has been developed, capable of learning the dynamic deformation and lighting rendering of any hairstyle with head movement, supporting real-time reanimation and high-fidelity rendering.
DGS-LRM: Real-Time Deformable 3D Gaussian Reconstruction From Monocular Videos
Chieh Hubert Lin (University of California Merced), Zhengqin Li (University of California Santa Barbara)
GenerationData SynthesisDepth EstimationTransformerGaussian SplattingOptical FlowVideo
🎯 What it does: An end-to-end Transformer model DGS-LRM has been constructed, capable of predicting deformable 3D Gaussian splats from monocular videos with poses in a single pass, achieving real-time dynamic scene reconstruction, view synthesis, and 3D scene flow estimation.
DGSolver: Diffusion Generalist Solver with Universal Posterior Sampling for Image Restoration
Hebaixu Wang (Wuhan University), Bo Du (Wuhan University)
RestorationSuper ResolutionDiffusion modelImageOrdinary Differential Equation
🎯 What it does: DGSolver is designed—a high-order ODE solver implemented on a general diffusion model, specifically for various image restoration tasks combined with general posterior sampling.
DICEPTION: A Generalist Diffusion Model for Visual Perceptual Tasks
Canyu Zhao (Zhejiang University), Chunhua Shen (Zhejiang University)
SegmentationPose EstimationDepth EstimationDiffusion modelFlow-based ModelImage
🎯 What it does: A visual general perception model DICEPTION based on a pre-trained text-image diffusion model has been developed, capable of performing up to six types of visual perception tasks (depth estimation, normal estimation, interactive segmentation, entity segmentation, instance segmentation, human keypoint estimation) within a single network while maintaining performance comparable to specialized models.
DiCo: Revitalizing ConvNets for Scalable and Efficient Diffusion Modeling
Yuang Ai (Chinese Academy of Sciences), Huaibo Huang (Chinese Academy of Sciences)
GenerationComputational EfficiencyConvolutional Neural NetworkDiffusion modelImageText
🎯 What it does: Designed and trained a fully convolutional diffusion model DiCo, replacing self-attention with convolution and enhancing performance through compact channel attention, significantly improving computational efficiency.
DiCoFlex: Model-Agnostic Diverse Counterfactuals with Flexible Control
Oleksii Furman (Wrocław University of Science and Technology), Marek Śmieja (Jagiellonian University)
Explainability and InterpretabilityComputational EfficiencyFlow-based ModelTabular
🎯 What it does: A model-free framework DiCoFlex has been designed and implemented to generate diverse and interpretable adversarial explanations with a single forward inference.
DictPFL: Efficient and Private Federated Learning on Encrypted Gradients
Jiaqi Xue (University of Central Florida), Qian Lou (University of Central Florida)
Federated LearningSafty and PrivacyComputational EfficiencyTransformerImageText
🎯 What it does: The DictPFL framework is proposed, which splits model weights into a static dictionary and a learnable lookup table, encrypting only the lookup table to achieve efficient HE federated learning.
DiEP: Adaptive Mixture-of-Experts Compression through Differentiable Expert Pruning
Sikai Bai (Hong Kong University of Science and Technology), Song Guo (Hong Kong University of Science and Technology)
CompressionComputational EfficiencyMixture of ExpertsText
🎯 What it does: This paper studies a differentiable expert pruning method for the Mixture-of-Experts model called DiEP, which enables hierarchical non-uniform sparsity and provides an online expert skipping mechanism, significantly reducing model storage and inference costs.