NeurIPS 2024 Papers — Page 10
Conference on Neural Information Processing Systems · 4035 papers
DiffGS: Functional Gaussian Splatting Diffusion
Junsheng Zhou (Tsinghua University), Yu-Shen Liu (Tsinghua University)
GenerationData SynthesisDiffusion modelAuto EncoderGaussian SplattingPoint CloudMesh
🎯 What it does: This paper proposes DiffGS, a 3D Gaussian spray generation framework based on latent diffusion models, capable of generating high-quality 3DGS with any number of Gaussians.
DiffHammer: Rethinking the Robustness of Diffusion-Based Adversarial Purification
Kaibo Wang (Hong Kong University of Science and Technology), Yang Xiang (Hong Kong University of Science and Technology)
Computational EfficiencyAdversarial AttackDiffusion modelImage
🎯 What it does: This paper proposes DiffHammer, a selective attack method for diffusion-based adversarial purification that effectively circumvents the gradient dilemma and enhances evaluation efficiency.
DiffLight: A Partial Rewards Conditioned Diffusion Model for Traffic Signal Control with Missing Data
Hanyang Chen (Beijing Jiaotong University), Huaiyu Wan (Beijing Jiaotong University)
Autonomous DrivingTransformerReinforcement LearningDiffusion modelTime Series
🎯 What it does: In the context of missing data, DiffLight is proposed, which jointly utilizes conditional diffusion models to achieve traffic signal control and data missing compensation.
DiffNorm: Self-Supervised Normalization for Non-autoregressive Speech-to-speech Translation
Weiting Tan (Johns Hopkins University), Philipp Koehn (Johns Hopkins University)
TransformerDiffusion modelAudio
🎯 What it does: Proposed DIFFNORM (self-supervised normalization based on diffusion models) and Classifier-Free Guidance to improve non-autoregressive speech-to-speech translation systems.
DiffPano: Scalable and Consistent Text to Panorama Generation with Spherical Epipolar-Aware Diffusion
Weicai Ye, Guofeng Zhang
GenerationData SynthesisDiffusion modelImageVideoText
🎯 What it does: The DiffPano framework is proposed to achieve scalable and cross-view consistent panoramic image generation based on text and camera poses.
DiffPhyCon: A Generative Approach to Control Complex Physical Systems
Long Wei (Westlake University), Tailin Wu (Westlake University)
OptimizationRobotic IntelligenceReinforcement LearningDiffusion modelTime SeriesPhysics Related
🎯 What it does: This paper proposes DiffPhyCon, a generative control framework based on diffusion models, which simultaneously generates control sequences and physical system trajectories while optimizing the overall control objectives.
DiffPO: A causal diffusion model for learning distributions of potential outcomes
Yuchen Ma (LMU Munich), Stefan Feuerriegel (LMU Munich)
Diffusion model
🎯 What it does: A causal diffusion model named DiffPO is proposed to learn the complete distribution of potential outcomes (PO) from observational data, providing both point estimates and uncertainty quantification.
DiffSF: Diffusion Models for Scene Flow Estimation
Yushan Zhang (Linköping University), Michael Felsberg (Linköping University)
Autonomous DrivingTransformerDiffusion modelPoint Cloud
🎯 What it does: Proposes the DiffSF model, which combines diffusion models with Transformers for point cloud scene flow estimation.
DiffTORI: Differentiable Trajectory Optimization for Deep Reinforcement and Imitation Learning
Weikang Wan (University of California San Diego), David Held (Carnegie Mellon University)
OptimizationRobotic IntelligenceReinforcement LearningAgentic AIImagePoint Cloud
🎯 What it does: This paper proposes DiffTORI, which uses differentiable trajectory optimization as a policy representation for deep reinforcement learning and imitation learning, and directly updates the dynamics and cost function of trajectory optimization through gradients.
DiffuBox: Refining 3D Object Detection with Point Diffusion
Xiangyu Chen (Cornell University), Kilian Q Weinberger
Object DetectionDomain AdaptationAutonomous DrivingDiffusion modelPoint Cloud
🎯 What it does: This paper proposes a point diffusion-based box correction method called DiffuBox, aimed at improving the localization and size of 3D object detection boxes across different domains.
DiffuLT: Diffusion for Long-tail Recognition Without External Knowledge
Jie Shao (Nanjing University), Jianxin Wu (Nanjing University)
ClassificationRecognitionDiffusion modelImage
🎯 What it does: This paper proposes a long-tail recognition pipeline based on diffusion models, which directly trains the diffusion model on the original long-tail data to generate approximate distribution samples to balance the dataset, thereby improving classification performance.
DiffuPac: Contextual Mimicry in Adversarial Packets Generation via Diffusion Model
Abdullah Bin Jasni (Nagaoka University of Technology), Kohei Watabe (Saitama University)
GenerationAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningDiffusion modelTabular
🎯 What it does: A generative adversarial model named DiffuPac is proposed and implemented, capable of generating covert malicious traffic that can bypass NIDS while maintaining malicious functionality.
DiffuserLite: Towards Real-time Diffusion Planning
Zibin Dong (Tianjin University), YAN ZHENG
Computational EfficiencyRobotic IntelligenceReinforcement LearningDiffusion modelScore-based ModelRectified FlowSequentialBenchmark
🎯 What it does: This paper proposes DiffuserLite, a lightweight diffusion planning framework that utilizes the Planning Refinement Process (PRP) to achieve coarse-to-fine trajectory generation, significantly enhancing decision frequency.
Diffusing Differentiable Representations
Yash Savani (Carnegie Mellon University), J Zico Kolter
GenerationData SynthesisDiffusion modelNeural Radiance FieldImage
🎯 What it does: A training-free sampling method is proposed, which directly samples in the differentiable representation (diffrep) space using a pre-trained diffusion model to generate high-quality images, panoramas, and 3D NeRF.
Diffusion Actor-Critic with Entropy Regulator
Yinuo Wang (Tsinghua University), Shengbo Eben Li (Tsinghua University)
Reinforcement LearningDiffusion modelMultimodality
🎯 What it does: Designed and implemented an online reinforcement learning algorithm DACER based on the reverse process of diffusion models, aimed at learning multimodal policies and controlling exploration through an entropy regulator.
Diffusion for World Modeling: Visual Details Matter in Atari
Eloi Alonso (University of Geneva), François Fleuret (University of Geneva)
Reinforcement LearningDiffusion modelVideoBenchmark
🎯 What it does: Developed and trained DIAMOND, a world model based on diffusion models, to learn reinforcement learning agents through imagination.
Diffusion Forcing: Next-token Prediction Meets Full-Sequence Diffusion
Boyuan Chen (Massachusetts Institute of Technology), Vincent Sitzmann (Massachusetts Institute of Technology)
GenerationRobotic IntelligenceRecurrent Neural NetworkTransformerReinforcement LearningDiffusion modelVideoTime SeriesSequential
🎯 What it does: Proposes the Diffusion Forcing training paradigm, which trains a causal sequence generator capable of denoising at independent noise levels for each token;
Diffusion Imitation from Observation
Bo-Ruei Huang (National Taiwan University), Shao-Hua Sun (National Taiwan University)
Robotic IntelligenceReinforcement LearningDiffusion modelSequential
🎯 What it does: A new learning from observation (LfO) framework is proposed—Diffusion Imitation from Observation (DIFO), which uses diffusion models as discriminators to provide rewards for policies, achieving action-free label imitation from state sequences.
Diffusion Model with Cross Attention as an Inductive Bias for Disentanglement
Tao Yang (Xi'an Jiaotong University), Nanning Zheng (Xi'an Jiaotong University)
GenerationRepresentation LearningConvolutional Neural NetworkDiffusion modelImage
🎯 What it does: Proposes the EncDiff framework, utilizing diffusion models and cross-attention as prior biases to achieve unsupervised separable representation learning.
Diffusion Models are Certifiably Robust Classifiers
Huanran Chen (Tsinghua University), Jun Zhu (RealAI)
ClassificationComputational EfficiencyDiffusion modelImage
🎯 What it does: This paper theoretically proves that diffusion classifiers have an O(1) Lipschitz constant and derives their verifiable robustness; it further extends diffusion classifiers to data with arbitrary noise levels, constructing Noised Diffusion Classifiers (NDC) and proposing two implementations: Exact Posterior and Approximate Posterior, utilizing random smoothing to enhance the robustness radius; at the same time, it designs two techniques, variance reduction and Sift-and-Refine, to significantly reduce computational complexity.
Diffusion Models With Learned Adaptive Noise
Subham Sekhar Sahoo (Cornell University), Volodymyr Kuleshov (Cornell University)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: A learnable multivariate adaptive noise (MuLAN) diffusion process is proposed to improve the log-likelihood estimation of image generation.
Diffusion of Thought: Chain-of-Thought Reasoning in Diffusion Language Models
Jiacheng Ye (University of Hong Kong), Lingpeng Kong
Diffusion modelTextChain-of-Thought
🎯 What it does: This paper proposes Diffusion-of-Thought (DoT), which combines diffusion models with chain-of-thought (CoT) to achieve parallel reasoning steps, and conducts experiments in areas such as multi-step reasoning, error self-correction, and reasoning efficiency adjustment.
Diffusion PID: Interpreting Diffusion via Partial Information Decomposition
Shaurya Rajat Dewan (Carnegie Mellon University), Yonatan Bisk (Carnegie Mellon University)
GenerationExplainability and InterpretabilityTransformerPrompt EngineeringDiffusion modelImageText
🎯 What it does: This study investigates a method based on Partial Information Decomposition (PID) and Conditional PID (CPID) to explain the generative process of text-to-image diffusion models, quantify the impact of words and their interactions on image pixels, and apply it to bias detection, synonym, and related word analysis.
Diffusion Policies Creating a Trust Region for Offline Reinforcement Learning
Tianyu Chen (University of Texas at Austin), Mingyuan Zhou (University of Texas at Austin)
Computational EfficiencyReinforcement LearningDiffusion modelTabularBenchmark
🎯 What it does: This paper proposes Diffusion Trusted Q-Learning (DTQL), which achieves efficient training and inference in offline reinforcement learning through a dual policy (diffusion policy and first-order policy) and a diffusion trust region loss.
Diffusion Policy Attacker: Crafting Adversarial Attacks for Diffusion-based Policies
Yipu Chen (Georgia Institute of Technology), Yongxin Chen (Georgia Institute of Technology)
Adversarial AttackRobotic IntelligenceReinforcement LearningDiffusion modelImage
🎯 What it does: This study investigates adversarial attacks on behavior cloning strategies based on diffusion models, proposing the DP-Attacker algorithm and evaluating its effectiveness in various attack scenarios.
Diffusion Priors for Variational Likelihood Estimation and Image Denoising
Jun Cheng (Huazhong University of Science and Technology), Shan Tan (Huazhong University of Science and Technology)
RestorationDiffusion modelImage
🎯 What it does: A self-supervised image denoising method based on diffusion priors and variational Bayes is proposed: during the reverse diffusion process, the noise precision posterior is adaptively estimated, the independent and identically distributed (i.i.d.) Gaussian likelihood is dynamically updated, and local Gaussian convolution is used to correct the noise variance, ultimately obtaining the denoised image through MAP inference; at the same time, the local prior of the low-resolution diffusion model is used to directly process high-resolution noisy images, and the method is extended to non-Gaussian noise and demosaicing tasks.
Diffusion Spectral Representation for Reinforcement Learning
Dmitry Shribak (Georgia Institute of Technology), Bo Dai (Georgia Institute of Technology)
Reinforcement LearningDiffusion modelTabularBenchmark
🎯 What it does: This paper proposes the Diffusion Spectral Representation (Diff-SR), which views diffusion models as energy-based models to learn a spectral representation that can fully express any policy value function, thereby achieving efficient planning and exploration in reinforcement learning.
Diffusion Tuning: Transferring Diffusion Models via Chain of Forgetting
Jincheng Zhong (Tsinghua University), Mingsheng Long (Tsinghua University)
GenerationData SynthesisTransformerDiffusion modelImage
🎯 What it does: Transfer learning is applied to pre-trained diffusion models, proposing the Diff-Tuning method, which utilizes the chain forgetting trend of the reverse process for knowledge retention and reconstruction.
Diffusion Twigs with Loop Guidance for Conditional Graph Generation
Giangiacomo Mercatali (HES-SO Genève University of Manchester), Vikas Garg (YaiYai Ltd & Aalto University)
GenerationOptimizationDrug DiscoveryGraph Neural NetworkDiffusion modelScore-based ModelGraphStochastic Differential Equation
🎯 What it does: This paper studies a fraction-based diffusion model called Twigs for conditional graph generation, particularly for tasks such as molecular design and network graph optimization.
Diffusion-based Layer-wise Semantic Reconstruction for Unsupervised Out-of-Distribution Detection
Ying Yang (Xidian University), Xinbo Gao (Chongqing University of Posts and Telecommunications)
Anomaly DetectionConvolutional Neural NetworkDiffusion modelImage
🎯 What it does: A hierarchical semantic reconstruction framework based on diffusion models is proposed, using multi-layer feature extraction and diffusion reconstruction to identify unsupervised anomaly distributions.
Diffusion-based Curriculum Reinforcement Learning
Erdi Sayar (Technical University of Munich), Alois Knoll (Technical University of Munich)
Robotic IntelligenceReinforcement LearningDiffusion modelSequential
🎯 What it does: A course reinforcement learning method based on conditional diffusion models, DiCuRL, is proposed, which utilizes diffusion models to generate intermediate goals with progressively increasing difficulty, helping agents learn efficiently in environments without domain knowledge.
Diffusion-based Reinforcement Learning via Q-weighted Variational Policy Optimization
Shutong Ding (ShanghaiTech University), Ye Shi (ShanghaiTech University)
Reinforcement LearningDiffusion modelMultimodality
🎯 What it does: An online reinforcement learning algorithm QVPO is proposed, utilizing diffusion models to achieve multimodal policies.
Diffusion-DICE: In-Sample Diffusion Guidance for Offline Reinforcement Learning
Liyuan Mao (Shanghai Jiao Tong University), Amy Zhang (University of Texas at Austin)
Reinforcement LearningDiffusion modelMultimodality
🎯 What it does: This paper proposes Diffusion-DICE, an offline reinforcement learning algorithm that utilizes diffusion models to convert behavior distributions into optimal policies, employing a guide-then-select mechanism to generate high-quality actions.
Diffusion-Inspired Truncated Sampler for Text-Video Retrieval
Jiamian Wang (Rochester Institute of Technology), ZHIQIANG TAO
RetrievalDiffusion modelContrastive LearningVideoText
🎯 What it does: This paper proposes a Diffusion-based Truncated Sampler (DITS) to improve the modality alignment issue in text-video retrieval systems.
Diffusion-Reward Adversarial Imitation Learning
Chun-Mao Lai (National Taiwan University), Shao-Hua Sun (National Taiwan University)
Robotic IntelligenceReinforcement LearningDiffusion modelSequential
🎯 What it does: A new adversarial imitation learning framework called DRAIL is proposed, which embeds diffusion models into GAIL to construct a more robust and smoother discriminator and generate rewards.
Diffusion4D: Fast Spatial-temporal Consistent 4D generation via Video Diffusion Models
HANWEN LIANG, Yunchao Wei (Beijing Jiaotong University)
GenerationData SynthesisDiffusion modelGaussian SplattingVideoMultimodality
🎯 What it does: A 4D content generation framework called Diffusion4D based on video diffusion models is proposed, which can quickly and consistently generate complete trajectory videos of dynamic 3D assets and explicitly construct them using high-quality 4D Gaussian Splatting.
DiffusionBlend: Learning 3D Image Prior through Position-aware Diffusion Score Blending for 3D Computed Tomography Reconstruction
Bowen Song (University of Michigan), Liyue Shen (University of Michigan)
RestorationDiffusion modelImageBiomedical DataComputed Tomography
🎯 What it does: Proposes two methods, DiffusionBlend and DiffusionBlend++, which utilize 3D-patch diffusion score fusion to learn 3D CT image priors and achieve 3D reconstruction under sparse view/limited angle conditions.
DiffusionFake: Enhancing Generalization in Deepfake Detection via Guided Stable Diffusion
Ke Sun (Xiamen University), Rongrong Ji (Xiamen University)
ClassificationDomain AdaptationTransformerDiffusion modelImageVideo
🎯 What it does: Proposes the DiffusionFake framework, which utilizes the pre-trained Stable Diffusion to perform inverse reconstruction of source and target images, guiding the detector to learn source/target features, thereby enhancing the cross-domain generalization ability of deepfake detection.
DiffusionPDE: Generative PDE-Solving under Partial Observation
Jiahe Huang (University of Michigan), Jeong Joon Park (University of Michigan)
GenerationDiffusion modelTime SeriesPhysics Related
🎯 What it does: A general PDE solving framework based on diffusion models, DiffusionPDE, is proposed, which can simultaneously recover PDE coefficients (or initial states) and solutions (or final states) with only a few observation points.
DigiRL: Training In-The-Wild Device-Control Agents with Autonomous Reinforcement Learning
Hao Bai (University of Illinois at Urbana-Champaign), Aviral Kumar (Carnegie Mellon University)
Robotic IntelligenceReinforcement LearningVision Language ModelSequentialBenchmark
🎯 What it does: A self-reinforcement learning framework called DigiRL was constructed and trained to perform device control tasks on Android devices using a pre-trained VLM.
DiGRAF: Diffeomorphic Graph-Adaptive Activation Function
Krishna Sri Ipsit Mantri (Purdue University), Moshe Eliasof (University of Cambridge)
ClassificationGraph Neural NetworkGraph
🎯 What it does: A graph adaptive activation function DIGRAF based on Continuous Piecewise Affine Transformations (CPAB) is proposed, which can learn differentiable, invertible, and bounded activation functions in graph neural networks.
Dimension-free deterministic equivalents and scaling laws for random feature regression
Leonardo Defilippis (École Normale Supérieure PSL CNRS), Theodor Misiakiewicz (Yale University)
ImageTabular
🎯 What it does: This paper studies the generalization performance of Random Feature Ridge Regression (RFRR) and proposes a non-asymptotic, dimension-independent deterministic equivalence to approximate the test error.
Dimension-free Private Mean Estimation for Anisotropic Distributions
Yuval Dagan (Tel Aviv University), Nikita Zhivotovskiy (University of California Berkeley)
OptimizationSafty and PrivacyTabular
🎯 What it does: A differentially private algorithm is proposed for high-dimensional mean estimation, particularly targeting non-isotropic distributions, overcoming the sample complexity issues of previous algorithms in high-dimensional cases.
DiMSUM: Diffusion Mamba - A Scalable and Unified Spatial-Frequency Method for Image Generation
Hao Phung (VinAI Research), Anh Tuan Tran
GenerationData SynthesisDiffusion modelAuto EncoderImage
🎯 What it does: This paper proposes a diffusion model architecture that integrates spatial and frequency information—DiMSUM, which combines Mamba state space networks with discrete wavelet transforms, and further fuses spatial features and frequency sub-bands through cross-attention to generate high-quality images.
DINTR: Tracking via Diffusion-based Interpolation
Pha Nguyen (University of Arkansas), Khoa Luu (University of Arkansas)
Object TrackingDiffusion modelVideoBenchmark
🎯 What it does: Proposes DINTR, a visual interpolation tracking framework based on latent diffusion models, achieving efficient and stable tracking of targets.
DiP-GO: A Diffusion Pruner via Few-step Gradient Optimization
haoweiz, Emad Barsoum (Advanced Micro Devices)
GenerationOptimizationComputational EfficiencyDiffusion modelImageText
🎯 What it does: A differential model pruning method through few-step gradient optimization (DiP-GO) is proposed, significantly improving inference speed while keeping the pre-trained model unchanged;
DiPEx: Dispersing Prompt Expansion for Class-Agnostic Object Detection
Jia Syuen Lim (University of Queensland), Yadan Luo (University of Queensland)
Object DetectionPrompt EngineeringVision Language ModelImage
🎯 What it does: A self-supervised Dispersing Prompt Expansion (DiPEx) method is proposed for class-agnostic object detection, enhancing recall through the diffusion of non-overlapping prompts.
Direct Consistency Optimization for Robust Customization of Text-to-Image Diffusion models
Kyungmin Lee (Korea Advanced Institute of Science and Technology), Jinwoo Shin (Korea Advanced Institute of Science and Technology)
GenerationOptimizationDiffusion modelImageText
🎯 What it does: This paper proposes Direct Consistency Optimization (DCO), a fine-tuning method for text-to-image diffusion models tailored for low sample customization, which achieves high consistency and text alignment while preserving the knowledge of the pre-trained model.
Direct Preference-Based Evolutionary Multi-Objective Optimization with Dueling Bandits
Tian Huang (University of Electronic Science and Technology of China), Ke Li (University of Exeter)
OptimizationProtein Structure Prediction
🎯 What it does: The D-PBEMO framework is proposed, utilizing direct feedback from dual decision-makers to guide evolutionary multi-objective optimization in search of solutions of interest (SOI).
Direct Unlearning Optimization for Robust and Safe Text-to-Image Models
Yong-Hyun Park (Korea University), Gayoung Lee (NAVER AI Lab)
GenerationOptimizationSafty and PrivacyDiffusion modelImageTextStochastic Differential Equation
🎯 What it does: This paper proposes a framework called DUO for directly performing concept forgetting on images, aiming to remove NSFW content from text-to-image models while maintaining generative capabilities for other subjects.
Direct3D: Scalable Image-to-3D Generation via 3D Latent Diffusion Transformer
Shuang Wu (DreamTech), Yao Yao (Nanjing University)
GenerationTransformerDiffusion modelMesh
🎯 What it does: Proposes Direct3D, a local 3D generation model that can directly generate high-quality 3D assets from a single image.
Directional Smoothness and Gradient Methods: Convergence and Adaptivity
Aaron Mishkin (Stanford University), Robert M. Gower (Flatiron Institute)
OptimizationTabular
🎯 What it does: A gradient descent suboptimality bound based on Directional Smoothness is proposed, along with a more compact, path-dependent convergence analysis compared to traditional global L-smoothness.
Director3D: Real-world Camera Trajectory and 3D Scene Generation from Text
Xinyang Li (Xiamen University), Rongrong Ji (Xiamen University)
GenerationData SynthesisTransformerDiffusion modelImageText
🎯 What it does: The Director3D framework is proposed to generate real-world 3D scenes from text and automatically generate suitable camera trajectories.
DisC-GS: Discontinuity-aware Gaussian Splatting
Haoxuan Qu (Lancaster University), Jun Liu (Lancaster University)
Gaussian SplattingImage
🎯 What it does: Proposes the DisC-GS framework, enabling Gaussian Splatting to consider and accurately present discontinuities and boundaries in images during rendering.
DisCEdit: Model Editing by Identifying Discriminative Components
Chaitanya Murti (Indian Institute of Science), Chiranjib Bhattacharyya (Indian Institute of Science)
ClassificationConvolutional Neural NetworkTransformerImage
🎯 What it does: A distribution-based model editing method called DISCEDIT is proposed, which utilizes discriminative filters to identify and prune key network components for structured pruning and category forgetting.
Discovering Creative Behaviors through DUPLEX: Diverse Universal Features for Policy Exploration
Borja G. León, Peter Stone (Sony AI)
Autonomous DrivingReinforcement LearningVideo
🎯 What it does: The DUPLEX method is proposed, which achieves robust adaptation by learning multiple sets of near-optimal diversified strategies in multi-context environments.
Discovering plasticity rules that organize and maintain neural circuits
David G Bell, Adrienne Fairhall (University of Washington)
Meta LearningSpiking Neural NetworkSequential
🎯 What it does: This study investigates how to discover self-supervised plasticity rules through meta-learning, enabling neural networks to self-organize and maintain sparse sequence generation dynamics in the face of biological noise such as synaptic turnover, drawing inspiration from the temporal coding mechanism of the avian HVC.
Discovering Preference Optimization Algorithms with and for Large Language Models
Chris Lu (Sakana AI), Robert Tjarko Lange (Sakana AI)
Recommendation SystemOptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Utilizing large language models to automatically generate and evaluate preference optimization loss functions, thereby discovering new offline preference optimization algorithms; ultimately proposing DiscoPOP (Logarithmic Ratio Modulation Loss) and validating its effectiveness across various dialogue, summarization, and sentiment generation tasks.
Discovering Sparsity Allocation for Layer-wise Pruning of Large Language Models
Lujun Li (Hong Kong University of Science and Technology), Yike Guo (Hong Kong University of Science and Technology)
OptimizationTransformerLarge Language ModelTextMultimodality
🎯 What it does: This paper proposes the DSA framework, which automatically searches for hierarchical sparse allocation functions for hierarchical pruning of large language models, significantly reducing the model parameter count while maintaining or improving performance.
Discovery of the Hidden World with Large Language Models
Chenxi Liu (Hong Kong Baptist University), Kun Zhang (Carnegie Mellon University)
TransformerLarge Language ModelPrompt EngineeringTextTime SeriesBiomedical DataMagnetic Resonance Imaging
🎯 What it does: The COAT framework is proposed, which utilizes large language models to automatically generate and annotate high-level variables from unstructured text, and identifies the Markov Blanket of the target variable and its causal structure through causal discovery algorithms.
Discrete Dictionary-based Decomposition Layer for Structured Representation Learning
Taewon Park (Kyungpook National University), Minho Lee (ALI Co, Ltd)
Representation LearningRecurrent Neural NetworkTransformerText
🎯 What it does: A discrete dictionary decomposition layer (D3) is proposed to enhance the decomposition capability of tensor product representation (TPR) models in system generalization tasks.
Discrete Flow Matching
Itai Gat (Meta FAIR), Yaron Lipman (Weizmann Institute of Science)
GenerationData SynthesisFlow-based ModelImageText
🎯 What it does: A Discrete Flow Matching (DFM) framework is proposed for non-autoregressive generation of high-dimensional discrete data.
Discrete Modeling via Boundary Conditional Diffusion Processes
Yuxuan Gu (Harbin Institute of Technology), Bing Qin (Harbin Institute of Technology)
GenerationData SynthesisTransformerDiffusion modelImageTextOrdinary Differential Equation
🎯 What it does: A diffusion model based on discrete boundary conditions is proposed, which first estimates the boundaries of discrete samples and scales the trajectories in both the forward and backward processes, making the continuous diffusion model more suitable for generating discrete data.
Discrete-state Continuous-time Diffusion for Graph Generation
Zhe Xu (University of Illinois Urbana-Champaign), Hanghang Tong (University of Illinois Urbana-Champaign)
GenerationData SynthesisGraph Neural NetworkDiffusion modelGraph
🎯 What it does: The first discrete state continuous time graph diffusion generative model DISCO is proposed;
Discretely beyond $1/e$: Guided Combinatorial Algortihms for Submodular Maximization
Yixin Chen (Texas A&M University), Alan Kuhnle (Texas A&M University)
OptimizationVideo
🎯 What it does: A combinatorial algorithm based on fast local search guidance is designed, breaking the 1/e barrier, achieving approximation ratios of 0.385-ε and 0.305-ε under size and matrix constraints, with a query complexity of O(kn/ε); a near-linear query deterministic algorithm with an approximation ratio of 0.377-ε is also introduced.
DisenGCD: A Meta Multigraph-assisted Disentangled Graph Learning Framework for Cognitive Diagnosis
Shangshang Yang (Anhui University), Xingyi Zhang (Anhui University)
Representation LearningGraph Neural NetworkGraph
🎯 What it does: The DisenGCD framework is proposed, which uses three decoupled graphs to learn representations of students, exercises, and knowledge concepts, and achieves robustness in student representations through a meta multigraph learning module.
Disentangled Representation Learning in Non-Markovian Causal Systems
Adam Li (Columbia University), Elias Bareinboim (Columbia University)
Representation LearningImage
🎯 What it does: A theoretical framework and algorithm based on causal structure diagrams (CRID) is proposed to determine which latent causal variables can be disentangled in non-Markovian distributions with mixed observational variables from multiple domains.
Disentangled Style Domain for Implicit $z$-Watermark Towards Copyright Protection
Junqiang Huang (Fudan University), Xinpeng Zhang (Fudan University)
Data SynthesisSafty and PrivacyAuto EncoderContrastive LearningImageText
🎯 What it does: This paper proposes an implicit Zero-Watermarking scheme based on decoupled style domains for detecting unauthorized data usage in text-to-image models.
Disentangled Unsupervised Skill Discovery for Efficient Hierarchical Reinforcement Learning
Jiaheng Hu (University of Texas at Austin), Roberto Martín-Martín (University of Texas at Austin)
Robotic IntelligenceReinforcement LearningImage
🎯 What it does: The DUSDi method is proposed, utilizing the factorized structure of the state space to achieve parallel and reconfigurable decoupled skill learning through an information-theoretic objective, which is directly applied to downstream tasks in hierarchical reinforcement learning.
Disentangling and mitigating the impact of task similarity for continual learning
Naoki Hiratani (Washington University in St Louis)
Knowledge DistillationImage
🎯 What it does: A linear teacher-student model with a low-dimensional latent structure was constructed, and the impact of task similarity (similarity of input features and similarity of output readouts) on knowledge transfer and retention in continual learning was analyzed; further, the alleviating effects of weight regularization using activity gating, plasticity gating, soft thresholding, and Euclidean and Fisher information metrics were evaluated.
Disentangling Interpretable Factors with Supervised Independent Subspace Principal Component Analysis
Jiayu Su (Columbia University), Raul Rabadan (Columbia University)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningBiomedical Data
🎯 What it does: A linear multi-subspace PCA extension named sisPCA is proposed to decouple and explain multiple independent subspaces in high-dimensional biological data.
Disentangling Linear Quadratic Control with Untrusted ML Predictions
Tongxin Li (Chinese University of Hong Kong), Yisong Yue (California Institute of Technology)
OptimizationReinforcement Learning from Human FeedbackReinforcement LearningTime Series
🎯 What it does: An online control strategy named DISC is proposed, which can adaptively learn the confidence parameters of machine learning predictions in the presence of unobservable disturbances in linear quadratic control problems, and achieve control by combining decoupled predictive information.
Disentangling the Roles of Distinct Cell Classes with Cell-Type Dynamical Systems
Aditi Jha (Princeton University), Jonathan W. Pillow (Princeton University)
Recurrent Neural NetworkReinforcement LearningTime SeriesBiomedical Data
🎯 What it does: A cell type-specific linear dynamical system (CTDS) was developed to explain the dynamics of neuronal populations and predict the behavioral effects of cell-specific optogenetic perturbations by assigning independent latent variables to excitatory and inhibitory neurons and constraining their interactions.
DISP-LLM: Dimension-Independent Structural Pruning for Large Language Models
Shangqian Gao (Florida State University), Yen-Chang Hsu (Samsung Research America)
OptimizationTransformerLarge Language ModelText
🎯 What it does: A dimension-independent structural pruning method called DISP-LLM is proposed, which can find high-quality sub-networks in LLMs without updating weights.
Dissect Black Box: Interpreting for Rule-Based Explanations in Unsupervised Anomaly Detection
Yu Zhang (Tsinghua University), Yong Jiang (Tsinghua University)
Anomaly DetectionExplainability and InterpretabilityTabular
🎯 What it does: A global interpretability framework based on SCD-Tree and Gaussian Boundary Delineation is proposed to automatically extract rules from black-box unsupervised anomaly detection models.
Dissecting Query-Key Interaction in Vision Transformers
Xu Pan (Harvard University), Odelia Schwartz (University of Miami)
TransformerImage
🎯 What it does: This paper analyzes the attention preferences of self-attention at different levels for similar and dissimilar tokens by performing singular value decomposition on the query-key interaction matrix in visual Transformers, and visualizes semantically interpretable interaction patterns.
Dissecting the Failure of Invariant Learning on Graphs
Qixun Wang (Peking University), Xianghua Ying (Peking University)
Domain AdaptationGraph Neural NetworkGraphBenchmark
🎯 What it does: Analyzed the reasons for the failure of traditional invariant learning methods IRM and VREx in node-level graph OOD tasks, and proposed Cross-environment Intra-class Alignment (CIA) and its environment-free label variant CIA-LRA to enhance the generalization performance of node-level graph OOD.
Dissecting the Interplay of Attention Paths in a Statistical Mechanics Theory of Transformers
Lorenzo Tiberi (Harvard University), Haim Sompolinsky (Hebrew University of Jerusalem)
ClassificationRecognitionCompressionOptimizationTransformerImageSequential
🎯 What it does: This paper derives the predictive mean and kernel function expressions of deep multi-head self-attention networks (Transformer approximation models) under the thermodynamic limit of finite width based on the framework of Bayesian statistical mechanics, revealing the task-related kernel combination mechanism among attention paths.
DistillNeRF: Perceiving 3D Scenes from Single-Glance Images by Distilling Neural Fields and Foundation Model Features
Letian Wang (NVIDIA Research), Peter Karkus (NVIDIA Research)
Depth EstimationAutonomous DrivingKnowledge DistillationNeural Radiance FieldContrastive LearningImagePoint Cloud
🎯 What it does: This paper presents DistillNeRF, a generalizable 3D scene representation framework that can perform scene reconstruction, viewpoint synthesis, depth estimation, and zero-shot semantic occupancy prediction using only single-frame multi-view camera images.
Distributed Least Squares in Small Space via Sketching and Bias Reduction
Sachin Garg (University of Michigan), Michal Derezinski
OptimizationTabular
🎯 What it does: This paper proposes a sparse matrix sketching method (called LESS / LESSUniform) in streaming and distributed computing environments, which constructs an almost unbiased least squares estimator by minimizing the bias of the estimator rather than the error. The method is implemented in one or two passes over the data, with a space complexity of only O(d·log(nd)) bits and a time complexity of the current matrix multiplication time O(d^ω).
Distributed-Order Fractional Graph Operating Network
Kai Zhao (Nanyang Technological University), Wee Peng Tay (Nanyang Technological University)
Graph Neural NetworkGraphBiomedical DataStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: A continuous graph neural network framework called DRAGON is proposed, which incorporates learnable distributed fractional-order operators to replace traditional single fractional or integer-order models, achieving more flexible feature updates.
Distribution Guidance Network for Weakly Supervised Point Cloud Semantic Segmentation
Zhiyi Pan (Peking University), Ge Li (Peking University)
SegmentationPoint Cloud
🎯 What it does: The paper proposes a distribution-guided network (DGNet) for weakly supervised point cloud semantic segmentation, which improves performance by modeling the feature space as a mixture of von Mises-Fisher distributions and dynamically aligning them.
Distribution Learning with Valid Outputs Beyond the Worst-Case
Nicholas Rittler, Kamalika Chaudhuri (University of California - San Diego)
GenerationOptimization
🎯 What it does: The study addresses the problem of ensuring the validity of outputs in generative models through distribution learning, proposing that under specific assumptions, learning can be completed with fewer validity queries by minimizing log-loss or utilizing VC class validity functions.
Distribution-Aware Data Expansion with Diffusion Models
haoweiz, Bin Wang (Tsinghua University)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: This paper proposes DistDiff, a training-free data augmentation framework that generates new samples consistent with the real distribution and diverse by using energy guidance and residual multiplicative transformations at intermediate steps of the diffusion model.
Distributional Preference Alignment of LLMs via Optimal Transport
Igor Melnyk (IBM Research), Jarret Ross (IBM Research)
Recommendation SystemOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This study proposes a distributed preference alignment method based on optimal transport (AOT), achieving distribution-level preference alignment for LLMs through a closed-form solution of one-dimensional convex OT.
Distributional regression: CRPS-error bounds for model fitting, model selection and convex aggregation
Dombry Clement, Ahmed Zaoui (Universite de Franche-Comte)
Tabular
🎯 What it does: In the framework of distributed regression, model fitting, model selection, and convex aggregation are performed using Continuous Ranked Probability Score (CRPS), and non-asymptotic convergence properties and upper bounds on expected errors are provided.
Distributional Reinforcement Learning with Regularized Wasserstein Loss
Ke Sun (University of Alberta), Linglong Kong (University of Alberta)
Reinforcement LearningSequential
🎯 What it does: A distributed reinforcement learning algorithm called SinkhornDRL based on Sinkhorn divergence is proposed, which uses regularized Wasserstein loss to measure the difference between the current and target return distributions.
Distributional Successor Features Enable Zero-Shot Policy Optimization
Chuning Zhu (University of Washington), Abhishek Gupta (University of Washington)
OptimizationRobotic IntelligenceReinforcement LearningDiffusion modelSequential
🎯 What it does: A distributed successor feature (DiSPO) framework is proposed, which learns the distribution of successor features and the corresponding reading strategy from offline data, achieving zero-shot policy optimization for arbitrary rewards.
Distributionally Robust Performative Prediction
Songkai Xue (University of Michigan), Yuekai Sun (University of Michigan)
OptimizationTabularFinance Related
🎯 What it does: A distributionally robust performative prediction framework is proposed, and a solution method for the Distributionally Robust Performative Optimum (DRPO) is defined.
Distributionally Robust Reinforcement Learning with Interactive Data Collection: Fundamental Hardness and Near-Optimal Algorithms
Miao Lu (Stanford University), Jose Blanchet (Stanford University)
Reinforcement Learning
🎯 What it does: This paper studies how to learn a policy that performs well in all possible testing environments through robust reinforcement learning (RMDP) under interactive data collection, and theoretically explores the feasibility and difficulties of this problem.
DistrictNet: Decision-aware learning for geographical districting
Cheikh Ahmed (Polytechnique Montreal), Thibaut Vidal (Polytechnique Montreal)
OptimizationGraph Neural NetworkGraph
🎯 What it does: The DISTRICTNET framework is proposed, which utilizes graph neural networks to generate edge weights and employs a differentiable capacity minimum spanning tree (CMST) solver to predict and solve the urban districting problem.
DiTFastAttn: Attention Compression for Diffusion Transformer Models
Zhihang Yuan (Tsinghua University), Yu Wang (Tsinghua University)
GenerationData SynthesisCompressionComputational EfficiencyTransformerDiffusion modelImageVideo
🎯 What it does: A post-training compression method called DiTFastAttn is proposed, which significantly reduces the attention computation of the Diffusion Transformer in image and video generation and accelerates inference by utilizing three techniques: window attention residual sharing, temporal attention sharing, and CFG attention sharing.
Divergences between Language Models and Human Brains
Yuchen Zhou (Carnegie Mellon University), Leila Wehbe (Carnegie Mellon University)
TransformerLarge Language ModelSupervised Fine-TuningTextMagnetic Resonance Imaging
🎯 What it does: This paper systematically compares the differences in processing narrative texts between language models (GPT-2 XL) and the human brain, using MEG data to evaluate model prediction errors, automatically generating and validating two major hypotheses regarding social/emotional and physical common sense differences, and enhancing the consistency with brain responses through fine-tuning the model on corresponding datasets.
Diversify, Contextualize, and Adapt: Efficient Entropy Modeling for Neural Image Codec
Jun-Hyuk Kim (Samsung Advanced Institute of Technology), Dokwan Oh (Samsung Advanced Institute of Technology)
CompressionTransformerImage
🎯 What it does: A fast and efficient entropy modeling framework DCA is designed and implemented for neural image encoders; it performs forward adaptation through multi-scale hyperpotential representations (local, regional, global) and uses backward and forward context in a four-step block to achieve adaptive distribution parameterization.
Diversity Is Not All You Need: Training A Robust Cooperative Agent Needs Specialist Partners
Rujikorn Charakorn (Vistec Institute of Technology), Nat Dilokthanakul (King Mongkut's Institute of Technology Ladkrabang)
Knowledge DistillationRobotic IntelligenceReinforcement LearningSequential
🎯 What it does: This paper explores the impact of training partner diversity, specialization, and overfitting (handshaking) on the robustness of general agents in cooperative multi-agent learning. It proposes the SpecTRL/SpecTRLdAgger methods, which combine reinforcement learning and supervised learning for knowledge distillation, allowing the retention of diverse and specialized behaviors from partners generated by XP-min while eliminating overfitting behaviors.
Diversity-Driven Synthesis: Enhancing Dataset Distillation through Directed Weight Adjustment
Jiawei Du (Agency for Science Technology and Research), Joey Tianyi Zhou (Agency for Science Technology and Research)
Data SynthesisKnowledge DistillationImage
🎯 What it does: This paper proposes a new dataset distillation method called Directed Weight Adjustment (DWA), which enhances the diversity and representativeness of the synthetic dataset by dynamically adjusting the weights during the synthesis process.
Divide-and-Conquer Meets Consensus: Unleashing the Power of Functions in Code Generation
Jingchang Chen (Harbin Institute of Technology), Bing Qin (Harbin Institute of Technology)
AI Code AssistantTransformerLarge Language ModelText
🎯 What it does: A code generation framework called FUNCODER is proposed, which combines divide-and-conquer with functional consensus to recursively decompose functions and filter through functional consensus to achieve code generation for complex requirements.
Divide-and-Conquer Posterior Sampling for Denoising Diffusion priors
Yazid Janati (Ecole polytechnique), Jimmy Olsson (KTH Royal Institute of Technology)
RestorationSuper ResolutionDiffusion modelImage
🎯 What it does: In Bayesian inverse problems, a denoising diffusion model (DDM) is used as a prior, and a block decomposition posterior sampling framework (DCPS) is proposed, which gradually approximates the target posterior by constructing a series of intermediate posteriors;
Divide-and-Conquer Predictive Coding: a structured Bayesian inference algorithm
Eli Zachary Sennesh, Tommaso Salvatori (Vienna University of Technology)
GenerationData SynthesisScore-based ModelImage
🎯 What it does: A new divide-and-conquer predictive coding algorithm (DCPC) is proposed, which achieves Bayesian inference for structured generative models through hierarchical sampling and local predictive error.
DMesh: A Differentiable Mesh Representation
Sanghyun Son (University of Maryland), Yi Zhou (Adobe Research)
OptimizationGenerative Adversarial NetworkPoint CloudMesh
🎯 What it does: A differentiable 3D mesh representation called DMesh is proposed, which can simultaneously optimize vertex geometry and connectivity, supporting mesh reconstruction with different topologies.