ICLR 2026 Papers — Page 12
International Conference on Learning Representations · 5356 papers
Designing Rules to Pick a Rule: Aggregation by Consistency
Ratip Emin Berker (Carnegie Mellon University), Nihar B Shah (Carnegie Mellon University)
Recommendation SystemOptimizationTabularTime SeriesSequential
🎯 What it does: Designed a data-driven rule selection method called AbC, which selects the most suitable ranking aggregation rule by randomly splitting votes and maximizing the consistency of aggregated results after splitting.
Designing Time Series Experiments in A/B Testing with Transformer Reinforcement Learning
Xiangkun Wu (Zhejiang University), Chengchun Shi (London School of Economics and Political Science)
OptimizationTransformerReinforcement LearningTime Series
🎯 What it does: Proposes a method for time series A/B test experiment design using Transformer-based reinforcement learning, aiming to minimize the mean squared error (MSE) of the average treatment effect (ATE) estimation.
Detect, Decide, Unlearn: A Transfer-Aware Framework for Continual Learning
Yiwen Wang (University of Auckland), Yun Sing Koh (University of Auckland)
Domain AdaptationConvolutional Neural NetworkTransformerImage
🎯 What it does: Proposed a continuous learning framework called DEDUCE, which detects and mitigates negative transfer by selectively 'forgetting' interfering old knowledge to enhance the stability and plasticity of learning new tasks.
Detecting and Mitigating Memorization in Diffusion Models through Anisotropy of the Log-Probability
Rohan Asthana (Friedrich-Alexander-Universität Erlangen-Nürnberg), Vasileios Belagiannis (Friedrich-Alexander-Universität Erlangen-Nürnberg)
GenerationPrompt EngineeringDiffusion modelScore-based ModelImageTextBenchmark
🎯 What it does: Designed a memory detection and mitigation framework without denoising, leveraging high-noise isotropic and low-noise anisotropic features to detect memory in diffusion models, and achieving mitigation during inference through prompt embedding gradient optimization.
Detecting Data Contamination from Reinforcement Learning Post-training for Large Language Models
Yongding Tao (Peking University), Ge Li (Peking University)
Anomaly DetectionData-Centric LearningReinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: This paper addresses the data pollution problem in large language models (LLM) during the reinforcement learning (RL) post-training phase, proposing a self-critique-based entropy similarity detection method called Self-Critique, and constructing the RL-MIA benchmark for systematic evaluation.
Detecting Data Contamination in LLMs via In-Context Learning
Michał Zawalski (NVIDIA), Pablo Ribalta (NVIDIA)
Anomaly DetectionData-Centric LearningTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose the CoDeC method, which utilizes adding same-dataset samples in the context to detect training data contamination in large language models and quantifies the extent of contamination.
Detecting Invariant Manifolds in ReLU-Based RNNs
Lukas Eisenmann (Heidelberg University), Daniel Durstewitz (Heidelberg University)
Recurrent Neural NetworkTime Series
🎯 What it does: Propose a semi-analytical algorithm based on ReLU RNN (PLRNN) to compute stable and unstable invariant manifolds, which can be used for partitioning basins of attraction, identifying similar/dissimilar orbits, and detecting chaos.
Detecting Misbehaviors of Large Vision-Language Models by Evidential Uncertainty Quantification
Tao Huang (State Key Laboratory of Advanced Rail Autonomous Operation), Liping Jing (Beijing Key Laboratory of Security and Privacy in Intelligent Transportation)
Anomaly DetectionVision Language ModelMultimodalityBenchmark
🎯 What it does: By analyzing errors generated by large vision-language models (LVLM), we attribute them to internal conflicts and missing information, and propose a training-free evidence uncertainty quantification method based on the Dempster-Shafer theory.
Detecting Temporal Misalignment Attacks in Multimodal Fusion for Autonomous Driving
Md Hasan Shahriar (Virginia Tech), Wenjing Lou (Virginia Tech)
Anomaly DetectionAutonomous DrivingContrastive LearningMultimodalityPoint Cloud
🎯 What it does: To address the performance degradation in autonomous driving multi-modal fusion caused by temporal misalignment, this paper proposes a lightweight defense framework called AION to detect and suppress temporal misalignment attacks (TMA).
Detection of unknown unknowns in autonomous systems
Ayan Banerjee (Arizona State University), Sandeep Gupta (Arizona State University)
Anomaly DetectionRecurrent Neural NetworkTime SeriesBenchmarkOrdinary Differential Equation
🎯 What it does: Propose a zero-shot multivariate time series anomaly detection method SPIE-AD based on physical dynamics model recovery and consistency reasoning, designed to identify unknown unknowns (U2) errors in autonomous systems.
Detective SAM: Adaptive AI-Image Forgery Localization
Gert Lek (University of Neuchâtel), Lydia Y. Chen (University of Neuchâtel)
SegmentationAnomaly DetectionTransformerVision Language ModelDiffusion modelImage
🎯 What it does: Developed an adaptable image forgery localization framework called Detective SAM based on SAM2, and proposed AutoEditForge to automatically generate training data.
Deterministic Bounds and Random Estimates of Metric Tensors on Neuromanifolds
Ke Sun (CSIRO Data61)
ClassificationConvolutional Neural NetworkTransformerImageTextAudio
🎯 What it does: This paper studies the Fisher information matrix (FIM) on the parameter space (neuromanifold) of deep classification networks, provides its spectral properties and bounds in a low-dimensional core space, and proposes an unbiased random estimator based on the Hutchinson method;
DETR-ViP: Detection Transformer with Robust Discriminative Visual Prompts
Bo Qian (Xi'an Jiaotong University), Xing Wei (Xi'an Jiaotong University)
Object DetectionTransformerPrompt EngineeringContrastive LearningImageText
🎯 What it does: Built upon Grounding DINO, the VIS-GDINO model was developed to support visual prompts, and further improved to DETR-ViP, enabling more accurate utilization of visual prompts for open-vocabulary object detection.
Developmental Federated Tuning: A Cognitive-Inspired Paradigm for Efficient LLM Adaptation
Yebo Wu (University of Macau), Li Li (University of Macau)
Federated LearningComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelText
🎯 What it does: Proposes DEVFT, a federated fine-tuning framework that gradually expands sub-models in stages, enabling large language models to perform task adaptation more efficiently on edge devices.
DevOps-Gym: Benchmarking AI Agents in Software DevOps Cycle
Yuheng Tang (UC Santa Barbara), Wenbo Guo (UC Santa Barbara)
AI Code AssistantLarge Language ModelAgentic AITextBenchmark
🎯 What it does: Propose the DEVOPS-GYM benchmark, covering four DevOps stages: build and configuration, monitoring, problem solving, and test generation, providing over 700 real tasks and 14 end-to-end pipeline tasks; evaluate AI agents through standardized tool invocation interfaces and dynamic execution environments.
DexMove: Learning Tactile-Guided Non-Prehensile Manipulation with Dexterous Hands
Pei Lin (ShanghaiTech University), Ziyuan Jiao (Beijing Institute for General Artificial Intelligence)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerReinforcement LearningFlow-based ModelMultimodalityTime Series
🎯 What it does: Studied the non-grasping object repositioning task for a tactile-guided multi-finger dexterous hand, and proposed the DexMove framework.
DexNDM: Closing the Reality Gap for Dexterous In-Hand Rotation via Joint-Wise Neural Dynamics Model
Xueyi Liu (Tsinghua University), Li Yi (Tsinghua University)
Domain AdaptationRobotic IntelligenceReinforcement Learning
🎯 What it does: Achieved in-hand rotation of objects with various shapes, sizes, and high aspect ratios on real robots using a sim-to-real framework based on joint-level neural dynamics models and automated data collection.
DGNet: Discrete Green Networks for Data-Efficient Learning of Spatiotemporal PDEs
Yingjie Tan (Tsinghua University), Yaqing Wang (Beijing Institute of Mathematical Sciences and Applications)
Computational EfficiencyGraph Neural NetworkMeshTime SeriesPhysics RelatedOrdinary Differential Equation
🎯 What it does: Proposed DGNet, a graph neural network leveraging the discrete Green's formula, for efficiently learning spatiotemporal PDE solutions under data-scarce conditions and achieving zero-shot generalization on unseen source terms.
DHG-Bench: A Comprehensive Benchmark for Deep Hypergraph Learning
Fan Li (University of New South Wales), Xuemin Lin (Shanghai Jiao Tong University)
Hyperparameter SearchGraph Neural NetworkGraphBenchmark
🎯 What it does: Designed and released DHG-Bench, a comprehensive benchmark that evaluates the performance of 17 HNNs on 22 hypergraph datasets using a unified experimental protocol, covering node-level, edge-level, and graph-level tasks, and exploring effectiveness, efficiency, robustness, and fairness.
DiaBlo: Diagonal Blocks Are Sufficient For Finetuning
Selcuk Gurses (University at Albany SUNY), Zi Yang (University at Albany SUNY)
OptimizationComputational EfficiencySupervised Fine-TuningText
🎯 What it does: This paper proposes the DiaBlo method, which achieves parameter-efficient fine-tuning by only fine-tuning the diagonal blocks of the pre-trained model's weight matrix;
Diagnosing and Improving Diffusion Models by Estimating the Optimal Loss Value
Yixian Xu (Peking University), Chang Liu (Zhongguancun Academy)
GenerationOptimizationDiffusion modelScore-based ModelImage
🎯 What it does: Proposes a theoretical and methodological framework for estimating the optimal loss of diffusion models, utilizing this estimation to diagnose and improve the training process and construct more efficient training schedules; simultaneously, re-examines the scaling laws of diffusion models using the optimal loss as a benchmark.
Diagnosing and Remedying Knowledge Deficiencies in LLMs via Label-free Curricular Meaningful Learning
Kai Xiong (Research Center for Social Computing and Interactive Robotics), Ting Liu (Research Center for Social Computing and Interactive Robotics)
Data SynthesisKnowledge DistillationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Automatically diagnose LLM knowledge defects through unlabeled user queries and improve them using curriculum-based meaningful learning
Diagnosing Generalization Failures from Representational Geometry Markers
Chi-Ning Chou (Flatiron Institute), SueYeon Chung (Harvard University)
Explainability and InterpretabilityRepresentation LearningHyperparameter SearchImage
🎯 What it does: Propose a diagnostic framework based on representative geometric metrics, which predicts model generalization failure by leveraging the intrinsic feature geometry that performs poorly on out-of-distribution (OOD) tasks.
DiCache: Let Diffusion Model Determine Its Own Cache
Jiazi Bu (Shanghai Jiao Tong University), Jiaqi Wang (Shanghai AI Laboratory)
GenerationComputational EfficiencyTransformerDiffusion modelImageVideoText
🎯 What it does: Designed and implemented an adaptive caching strategy called DiCache based on online shallow-layer probing to accelerate Diffusion model inference.
Dichotomous Diffusion Policy Optimization
Ruiming Liang (Chinese Academy of Sciences), Xianyuan Zhan (Tsinghua University)
Autonomous DrivingOptimizationReinforcement LearningDiffusion modelBenchmark
🎯 What it does: Proposes DIPOLE, a reinforcement learning framework that decomposes into positive and negative diffusion strategies using greedy KL regularization, enabling stable training and control of strategy greediness during offline and online fine-tuning.
DiffAdapt: Difficulty-Adaptive Reasoning for Token-Efficient LLM Inference
Xiang Liu (Hong Kong University of Science and Technology), Eunsol Choi (New York University)
Computational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: Analyze the overthinking phenomenon in LLM inference and propose DiffAdapt, a difficulty-based adaptive reasoning strategy framework, using a lightweight detector to dynamically select easy/normal/hard reasoning modes, reducing token consumption while improving accuracy
DiffBED: Scaling Bayesian Experimental Design to High-Dimensions
Adhi Saravanan, Tom Rainforth (University of Oxford)
OptimizationDiffusion modelImageStochastic Differential Equation
🎯 What it does: Propose the DiffBED method, combining diffusion models with information-guided design sampling to achieve Bayesian experimental design in high-dimensional spaces.
Difference Predictive Coding for Training Spiking Neural Networks
Ville Karlsson (Tampere University), Joni-Kristian Kämäräinen (Tampere University)
ClassificationSpiking Neural NetworkImage
🎯 What it does: Proposed a differential predictive coding (DiffPC) framework, transforming traditional predictive coding (PC) into a local, event-driven spiking implementation;
Difference-Aware Retrieval Policies for Imitation Learning
Quinn Pfeifer, Abhishek Gupta (University Of Washington)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerImageMultimodalityRetrieval-Augmented Generation
🎯 What it does: This paper proposes a semi-parametric retrieval-augmented imitation learning framework called DARP, which reparameterizes decisions using k-nearest neighbors and their relative difference vectors from expert demonstrations, enabling the model to aggregate local action predictions through neighborhood information during inference.
Differentiable JPEG-based Input Perturbation for Knowledge Distillation Amplification via Conditional Mutual Information Maximization
SIYU CHEN, EN-HUI YANG
Knowledge DistillationConvolutional Neural NetworkTransformerImage
🎯 What it does: Propose a differentiable JPEG preprocessing layer (DJIP) that enhances knowledge distillation effects by maximizing the teacher's conditional mutual information (CMI) through input perturbation, without modifying the teacher model's weights.
Differentiable Lifting for Topological Neural Networks
Jorge Luiz Franco (University of São Paulo), Amauri H Souza
ClassificationRepresentation LearningGraph Neural NetworkGraph
🎯 What it does: This paper proposes a differentiable graph lifting framework, DiffLift, which can automatically learn the optimal lifting to transform graphs into higher-order topological structures (such as hypergraphs, cell complexes, or simplicial complexes) during end-to-end training.
Differentiable Model Predictive Control on the GPU
Emre Adabag (Toyota Research Institute), Thomas Jonathan Lew (Toyota Research Institute)
Autonomous DrivingOptimization
🎯 What it does: Developed a GPU-based differentiable model predictive control solver called DiffMPC, and verified its performance in tasks such as reinforcement learning, imitation learning, and car drifting.
Differentiable Simulation of Hard Contacts with Soft Gradients for Learning and Control
Anselm Paulus (University of Tübingen), Georg Martius (University of Tübingen)
OptimizationTime SeriesPhysics RelatedOrdinary Differential Equation
🎯 What it does: Propose DiffMJX and CFD two techniques, using differentiable simulation to solve the problem of inaccurate gradients under hard contact and zero gradients for non-contact objects, and verify them in parameter identification and model predictive control (MPC).
Differential Fine-Tuning Large Language Models Towards Better Diverse Reasoning Abilities
Xiaosong Yuan (Jilin University), Jieping Ye (Alibaba Cloud Computing)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Differential fine-tuning (DiFT) for large language models' reasoning capabilities, which identifies task-related parameters and selectively updates them to enhance reasoning performance in multi-task and continual learning scenarios.
Differentially Private Domain Discovery
Vinod Raman (University of Michigan), Matthew Joseph (Google Research)
Domain AdaptationSafty and PrivacyText
🎯 What it does: This paper addresses three core problems in differential privacy domain discovery: set union, topk, and k-hitting, providing an absolute utility upper bound under missing mass and proving its near-optimality;
Differentially Private Equilibrium Finding in Polymatrix Games
Mingyang Liu (Massachusetts Institute of Technology), Asuman E. Ozdaglar (Massachusetts Institute of Technology)
Optimization
🎯 What it does: Studied the equilibrium finding problem in multi-matrix games under differential privacy constraints, proposing a new distributed algorithm that achieves vanishing Nash gap and privacy budget simultaneously as the number of players increases.
Differentially Private Two-Stage Gradient Descent for Instrumental Variable Regression
Haodong Liang (University of California, Davis), Lifeng Lai (University of California, Davis)
OptimizationSafty and PrivacyTabular
🎯 What it does: Proposed a two-stage gradient descent algorithm, DP-2S-GD, for solving instrumental variable regression (IVaR) under differential privacy constraints, and provided its ρ-zero concentrated differential privacy guarantee as well as finite sample convergence rate.
Difficult Examples Hurt Unsupervised Contrastive Learning: A Theoretical Perspective
Yi-Ge Zhang (Peking University), Yisen Wang (Peking University)
Representation LearningContrastive LearningImage
🎯 What it does: Propose the negative impact of hard samples on unsupervised contrastive learning, and provide a theoretical framework along with experimental validation.
Difficulty–Diversity Collaborative Filtering for Data-Efficient LLM Fine-Tuning
Long P. Hoang (Singapore University of Technology and Design), Wei Lu (Nanyang Technological University)
Data-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: This paper proposes an automated method to select high-quality training subsets tailored for target LLMs, significantly reducing annotation costs while maintaining fine-tuning performance comparable to full-data training.
DiffInk: Glyph- and Style-Aware Latent Diffusion Transformer for Text to Online Handwriting Generation
Wei Pan (South China University of Technology), Lianwen Jin (South China University of Technology)
GenerationTransformerDiffusion modelAuto EncoderImageText
🎯 What it does: Propose DiffInk, an end-to-end method for generating complete online handwritten text lines by combining InkVAE and InkDiT.
DiffPBR: Point-Based Rendering via Spatial-Aware Residual Diffusion
Yiping Xie (Zhejiang University), Qi Ye (Zhejiang Key Laboratory of Airspace Awareness and Autonomous Unmanned Systems)
GenerationDiffusion modelGaussian SplattingPoint Cloud
🎯 What it does: Propose a point cloud rendering framework DiffPBR based on diffusion models, capable of generating high-quality, view-consistent images;
DiffSDA: Unsupervised Diffusion Sequential Disentanglement Across Modalities
Hedi Zisling (Ben Gurion University), Omri Azencot (Ben Gurion University)
Representation LearningDiffusion modelAuto EncoderVideoMultimodalityTime SeriesAudio
🎯 What it does: Propose an unsupervised cross-modal sequence separation method called DiffSDA that can separate static and dynamic factors;
DiffSparse: Accelerating Diffusion Transformers with Learned Token Sparsity
Haowei Zhu (Advanced Micro Devices, Inc.), Emad Barsoum (Advanced Micro Devices, Inc.)
GenerationComputational EfficiencyTransformerDiffusion modelImageVideo
🎯 What it does: Propose DiffSparse, a learnable hierarchical sparse allocation framework to accelerate diffusion Transformer inference.
DiffTrans: Differentiable Geometry-Materials Decomposition for Reconstructing Transparent Objects
Changpu Li (Harbin Institute of Technology), Wenjie Pei (Harbin Institute of Technology)
OptimizationComputational EfficiencyNeural Radiance FieldImage
🎯 What it does: Proposed a differentiable rendering framework called DiffTrans for simultaneously reconstructing the geometry and material properties (refractive index and absorption rate) of transparent objects from multi-view images, supporting complex topology and internal textures;
DiffuCoder: Understanding and Improving Masked Diffusion Models for Code Generation
Shansan Gong (Apple), Yizhe Zhang (Apple)
AI Code AssistantTransformerLarge Language ModelReinforcement LearningDiffusion modelText
🎯 What it does: Trained a 7B-parameter Masked Diffusion Language Model (DiffuCoder) for code generation, and conducted a systematic analysis of its decoding behavior;
DiffuDETR: Rethinking Detection Transformers with Denoising Diffusion Process
Youssef Ahmed Nawar (Alexandria University), Marwan Torki (Alexandria University)
Object DetectionTransformerDiffusion modelContrastive LearningImage
🎯 What it does: Modify the object query initialization and training process of DETR to a denoising diffusion process, proposing DiffuDETR and DiffuDINO to achieve generation of precise boxes from noise.
DiffuGuard: How Intrinsic Safety is Lost and Found in Diffusion Large Language Models
Zherui Li (Beijing University of Posts and Telecommunications), Jiaheng Zhang (National University of Singapore)
Safty and PrivacyLarge Language ModelDiffusion modelText
🎯 What it does: This paper conducts a two-dimensional (intra-step and inter-step) safety analysis of the iterative reasoning process in diffusion large language models (dLLMs), revealing safety biases caused by greedy remasking and the 'denoising path dependency' phenomenon. Based on this, it proposes an untrained defense framework called DIFFUGUARD, which includes two modules: stochastic annealing remasking and block-level audit and repair.
Diffusion & Adversarial Schrödinger Bridges via Iterative Proportional Markovian Fitting
Sergei Kholkin, Alexander Korotin
GenerationDiffusion modelImageMultimodalityTabularBenchmarkPhysics RelatedStochastic Differential Equation
🎯 What it does: Proposed and theoretically analyzed a unified iterative proportional Markov fitting (IPMF) framework, revealing that bidirectional IMF is equivalent to alternating between IPF and IMF projection, and proving its exponential/weak convergence under Gaussian distribution and bounded support scenarios;
Diffusion Alignment as Variational Expectation-Maximization
Jaewoo Lee (KAIST), Jinkyoo Park (Omelet)
GenerationReinforcement LearningDiffusion modelImageBiomedical Data
🎯 What it does: Designed and verified a diffusion model alignment framework DAV based on variational expectation maximization, aiming to maximize rewards while maintaining diversity in image generation and DNA sequence design.
Diffusion and Flow-based Copulas: Forgetting and Remembering Dependencies
David Huk (University of Warwick), Theodoros Damoulas (University of Warwick)
GenerationRepresentation LearningDiffusion modelFlow-based ModelImageTabularStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Proposes two copula models based on diffusion and flow—classification diffusion copula and reflection copula. Designs a normal OU diffusion process and reflection process that only forgets dependencies between variables while maintaining uniform margins, and achieves density estimation and sampling by learning dependencies during the memory process.
Diffusion Blend: Inference-Time Multi-Preference Alignment for Diffusion Models
Min Cheng (Texas A&M University), Panganamala Kumar
GenerationComputational EfficiencyReinforcement LearningDiffusion modelImageStochastic Differential Equation
🎯 What it does: Propose the Diffusion Blend framework, enabling multi-preference alignment of diffusion models during inference by linearly combining user-specified rewards and KL regularization weights, without requiring retraining.
Diffusion Bridge Variational Inference for Deep Gaussian Processes
JIAN XU, John Paisley (Columbia University)
ClassificationRestorationDiffusion modelScore-based ModelImageTabularStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Proposed a variational inference method based on diffusion bridges (DBVI) for efficiently inferring the posterior of inducing variables in deep Gaussian processes.
Diffusion Fine-Tuning via Reparameterized Policy Gradient of the Soft Q-Function
Hyeongyu Kang (KAIST), Jinkyoo Park (KAIST)
GenerationReinforcement LearningDiffusion modelImageTextMultimodality
🎯 What it does: This paper proposes a KL-regularized reinforcement learning framework called SQDF, which directly utilizes differentiable rewards to fine-tune diffusion models via reparameterized policy gradients, avoiding over-optimization while maintaining sample diversity.
Diffusion Language Model Knows the Answer Before It Decodes
Pengxiang Li (Hong Kong Polytechnic University), Shiwei Liu (ELLIS Institute)
Computational EfficiencyLarge Language ModelDiffusion modelText
🎯 What it does: Proposed a no-training fast decoding strategy called Prophet, leveraging the phenomenon that diffusion language models (DLMs) can converge early in the decoding process, enabling early submission decoding.
Diffusion Language Models are Provably Optimal Parallel Samplers
Haozhe Jiang (University of California, Berkeley), Lijie Chen (University of California, Berkeley)
Computational EfficiencyDiffusion modelChain-of-Thought
🎯 What it does: Investigate and demonstrate that diffusion language models (DLMs) with extended chain-of-thought (CoT) can achieve minimal sequential steps in parallel sampling tasks, and reach optimal space complexity when supported by re-masking or revisions;
Diffusion LLMs Can Do Faster-Than-AR Inference via Discrete Diffusion Forcing
Xu Wang (Shanghai Jiao Tong University), Zhijie Deng (Shanghai Jiao Tong University)
Computational EfficiencyKnowledge DistillationAI Code AssistantTransformerSupervised Fine-TuningDiffusion modelText
🎯 What it does: Propose a training framework named Discrete Diffusion Forcing (D2F), which transforms discrete diffusion language models (dLLM) into an autoregressive (AR)-diffusion hybrid model that supports KV caching and enables block-level parallel inference, significantly accelerating inference speed.
Diffusion Models as Dataset Distillation Priors
Duo Su (Tsinghua University), Jun Zhu (Tsinghua University)
ClassificationData SynthesisDiffusion modelImage
🎯 What it does: Propose the Diffusion As Priors (DAP) framework, leveraging multiple priors from diffusion models for dataset distillation.
Diffusion Negative Preference Optimization Made Simple
Joshua Tian Jin Tee (Korea Advanced Institute of Science and Technology), Chang D. Yoo (Korea Advanced Institute of Science and Technology)
GenerationDiffusion modelImage
🎯 What it does: Propose a single-model Diff-SNPO for negative preference optimization, addressing computational costs and weakened contrast caused by dual-model training and weight merging.
Diffusion Transformers with Representation Autoencoders
Boyang Zheng (New York University), Saining Xie (New York University)
GenerationData SynthesisTransformerDiffusion modelAuto EncoderImage
🎯 What it does: Built a high-dimensional representation autoencoder (RAE) without compression by combining frozen pre-trained representation encoders (e.g., DINOv2, SigLIP, MAE) with trained ViT decoders, and trained Diffusion Transformers (DiT) in this space
Diffusion-DFL: Decision-focused Diffusion Models for Stochastic Optimization
Zihao Zhao (Georgia Institute of Technology), Kai Wang (Georgia Institute of Technology)
OptimizationDiffusion modelScore-based ModelTabularFinance Related
🎯 What it does: This paper proposes embedding diffusion generative models into decision-focused learning (DFL) to learn the complete uncertain parameter distribution and optimize stochastic decisions in an end-to-end manner.
DiffusionBlocks: Block-wise Neural Network Training via Diffusion Interpretation
Makoto Shing (Sakana AI), Takuya Akiba (Sakana AI)
ClassificationGenerationTransformerDiffusion modelScore-based ModelImageText
🎯 What it does: Propose the DiffusionBlocks framework, which splits residual networks such as Transformers into independently trainable blocks, utilizing diffusion theory to achieve gradient-free dependencies between blocks.
DiffusionNFT: Online Diffusion Reinforcement with Forward Process
Kaiwen Zheng (Tsinghua), Ming-Yu Liu (NVIDIA)
Reinforcement LearningDiffusion modelFlow-based ModelContrastive LearningImage
🎯 What it does: Proposed DiffusionNFT, an online reinforcement learning framework that directly optimizes policies on the forward process of diffusion models;
DiffVax: Optimization-Free Image Immunization Against Diffusion-Based Editing
Tarik Can Ozden, James Matthew Rehg
Safty and PrivacyComputational EfficiencyAdversarial AttackConvolutional Neural NetworkLarge Language ModelDiffusion modelImageVideo
🎯 What it does: Developed DiffVax, an unoptimized image immunization framework that can add imperceptible perturbations to images and videos in milliseconds, preventing diffusion model editing.
DiffWind: Physics-Informed Differentiable Modeling of Wind-Driven Object Dynamics
Yuanhang Lei (Zhejiang University), Zhaopeng Cui (Zhejiang University)
Data SynthesisOptimizationGaussian SplattingVideoPhysics Related
🎯 What it does: This paper proposes the DiffWind framework, which simultaneously reconstructs invisible wind fields and the four-dimensional motion of objects from sparse perspective videos using differentiable physics simulation, and supports forward simulation and wind relabeling.
Dimension-Free Decision Calibration for Nonlinear Loss Functions
Jingwu Tang (Carnegie Mellon University), Jiahao Zhang (Carnegie Mellon University)
OptimizationExplainability and Interpretability
🎯 What it does: Propose a dimension-agnostic decision calibration framework under a nonlinear loss function, utilizing smooth quantal response to align predictions with decisions.
DiMeR: Disentangled Mesh Reconstruction Model with Normal-only Geometry Training
Lutao Jiang (Hong Kong University of Science and Technology (Guangzhou)), Ying-Cong Chen (Hong Kong University of Science and Technology (Guangzhou))
GenerationTransformerImageMesh
🎯 What it does: Proposes DiMeR, a forward sparse view mesh reconstruction framework that decouples geometry and texture, where the geometry branch takes normal maps as input and is trained via 3D supervision, while the texture branch takes RGB images as input;
Direct Doubly Robust Estimation of Conditional Quantile Contrasts
Josh Givens (University of Bristol), Katarzyna Reluga (Humboldt University of Berlin)
TabularBiomedical Data
🎯 What it does: Proposes the first doubly robust method for directly estimating conditional quantile contrast (CQC), avoiding the traditional steps of first estimating intermediate functions and then inverting them.
Direct Preference Optimization for Primitive-Enabled Hierarchical RL: A Bilevel Approach
Utsav Singh (IIT Kanpur), Amrit Singh Bedi (University of Central Florida)
Reinforcement Learning from Human FeedbackReinforcement Learning
🎯 What it does: Propose a hierarchical reinforcement learning framework called DIPPER, which separates subgoal generation from low-level execution through bi-level optimization.
Direct Reward Fine-Tuning on Poses for Single Image to 3D Human in the Wild
Seunguk Do (Seoul National University), Jaesik Park (Seoul National University)
Image TranslationGenerationPose EstimationReinforcement LearningDiffusion modelImageMeshBenchmark
🎯 What it does: Post-training a multi-view diffusion model using the direct reward fine-tuning method DrPose to enhance the naturalness of poses in dynamic scenarios for single-view 3D human reconstruction.
Directed Semi-Simplicial Learning with Applications to Brain Activity Decoding
Manuel Lecha (Istituto Italiano di Tecnologia), Claudio Battiloro (Harvard University)
ClassificationGraph Neural NetworkGraphBiomedical Data
🎯 What it does: Proposed and implemented Semi-Simplicial Neural Networks (SSNs) for semi-simplicial sets to learn directed high-order network structures.
Directional Convergence, Benign Overfitting of Gradient Descent in leaky ReLU two-layer Neural Networks
Ichiro Hashimoto (University of Toronto)
ClassificationOptimization
🎯 What it does: This paper provides sufficient conditions for benign overfitting in fixed-width leaky ReLU two-layer neural network classifiers trained on mixed data via gradient descent.
Directional Sheaf Hypergraph Networks: Unifying Learning on Directed and Undirected Hypergraphs
Emanuele Mule (Sapienza University of Rome), Fabrizio Silvestri (Sapienza University of Rome)
ClassificationRepresentation LearningGraph Neural NetworkDiffusion modelGraphBenchmark
🎯 What it does: Proposed the Directional Cellular Laminar Tensor Network (DSHN), achieving effective learning of directed multi-dimensional relationships by constructing complex-valued cellular laminar Laplacian operators on directed hypergraphs.
Directional Textual Inversion for Personalized Text-to-Image Generation
Kunhee Kim (KAIST), Hyunjung Shim (KAIST)
GenerationTransformerPrompt EngineeringVision Language ModelDiffusion modelImageText
🎯 What it does: Proposes the Directional Text Inversion (DTI) method to achieve personalization and improve prompt fidelity in text-to-image generation.
DirMoE: Dirichlet-Routed Mixture of Experts
Amirhossein Vahidi (Wellcome Sanger Institute, Wellcome Genome Campus), Mohammad Lotfollahi (Wellcome Sanger Institute, Wellcome Genome Campus)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelMixture of ExpertsAuto EncoderText
🎯 What it does: Introduce a sparse Mixture-of-Experts router based on Dirichlet variational autoencoders, separating expert selection and contribution while maintaining end-to-end differentiability;
Discern Truth from Falsehood: Reducing Over-Refusal via Contrastive Refinement
Yuxiao Lu (Singapore Management University), Jie Shi (Huawei Technologies Co., Ltd)
Safty and PrivacyTransformerSupervised Fine-TuningContrastive LearningText
🎯 What it does: Introduce a contrastive refinement phase before safety alignment, leveraging contrastive learning to reduce LLMs' over-rejection of non-harmful but surface-toxic prompts.
Disco: Densely-overlapping Cell Instance Segmentation via Adjacency-aware Collaborative Coloring
Rui Sun (Shanghai Academy of Artificial Intelligence for Science), Yuan Cheng (Shanghai Academy of Artificial Intelligence for Science)
SegmentationConvolutional Neural NetworkBiomedical Data
🎯 What it does: Proposes Disco, a cell instance segmentation method based on graph coloring, to address the segmentation of overlapping cells in high-density tissues.
DISCO: Diversifying Sample Condensation for Efficient Model Evaluation
Alexander Rubinstein (Tübingen AI Center, University of Tübingen), Seong Joon Oh (Tübingen AI Center, University of Tübingen)
Computational EfficiencyKnowledge DistillationImageTextBenchmark
🎯 What it does: Proposed an efficient model evaluation method called DISCO, which constructs a small evaluation subset by selecting samples that generate the maximum model discrepancy, and directly predicts the performance of the full dataset using model signatures.
Discount Model Search for Quality Diversity Optimization in High-Dimensional Measure Spaces
Bryon Tjanaka (University of Southern California), Stefanos Nikolaidis (University of Southern California)
OptimizationConvolutional Neural NetworkTransformerVision Language ModelImage
🎯 What it does: Developed a quality diversity optimization algorithm called DMS based on discount model search, addressing distortion issues in high-dimensional measure spaces and supporting a new QDDM paradigm where measures are specified by datasets.
Discounted Online Convex Optimization: Uniform Regret Across a Continuous Interval
Wenhao Yang (Nanjing University), Lijun Zhang (Nanjing University)
OptimizationTabular
🎯 What it does: Studied online convex optimization (OCO) with a discount factor λ and proposed an algorithm called SOGD that does not require prior knowledge of λ, achieving a unified discounted equilibrium loss over a continuous interval.
Discovering alternative solutions beyond the simplicity bias in recurrent neural networks
William Qian (Harvard University), Cengiz Pehlevan (Harvard University)
OptimizationExplainability and InterpretabilityRecurrent Neural NetworkTime SeriesSequential
🎯 What it does: Proposed a training method called Iterative Neural Similarity Deflation (INSD), which iteratively generates multiple non-trivial solutions by penalizing the linear predictability of already trained RNNs.
Discovering and Steering Interpretable Concepts in Large Generative Music Models
Nikhil Singh (Dartmouth College), Patricia Maes (MIT)
GenerationExplainability and InterpretabilityTransformerLarge Language ModelAuto EncoderAudio
🎯 What it does: This paper proposes a method to unsupervisedly mine interpretable concepts from the residual flows of large autoregressive music generation models using sparse autoencoders (SAE), and employs these concepts to control the entire generation process.
Discovering heterogeneous synaptic plasticity rules via large-scale neural evolution
Ziyuan Ye (Hong Kong Polytechnic University), Jibin Wu (Hong Kong Polytechnic University)
OptimizationSpiking Neural NetworkImage
🎯 What it does: Explore heterogeneous synaptic plasticity rules in a biologically realistic mouse V1 model using a multi-objective evolutionary algorithm, seeking plasticity schemes that align with experimental observations and achieve cross-domain visual tasks.
Discovering Novel LLM Experts via Task-Capability Coevolution
Andrew Dai, Yujin Tang (Sakana AI)
Data SynthesisOptimizationTransformerLarge Language ModelMixture of ExpertsTextBenchmark
🎯 What it does: Developed an open-source co-evolutionary framework named AC/DC, jointly evolving large language models with synthetic tasks, generating a diverse cluster of expert models through model merging (crossover, noise mutation) and diversity optimization (DNS);
DiscoX: Benchmarking Discourse-Level Translation in Expert Domains
Xiying ZHAO (ByteDance), Wenhao Huang (ByteDance)
Large Language ModelAgentic AITextBenchmark
🎯 What it does: Proposed the DiscoX benchmark for evaluating expert-level Chinese↔English long dialogue-level translation, and designed a reference-free Metric-S evaluation system.
Discrete Adjoint Matching
Oswin So (MIT), Guan-Horng Liu (FAIR at Meta)
GenerationOptimizationDiffusion modelText
🎯 What it does: Proposed Discrete Adjoint Matching (DAM), a discrete generative model entropy-regularized reward optimization method based on CTMC.
Discrete Bayesian Sample Inference for Graph Generation
Ole Petersen (Technical University of Munich), Stephan Günnemann
GenerationData SynthesisDrug DiscoveryTransformerGraphBenchmarkStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Proposed a graph generation model called GraphBSI based on Bayesian Sample Inference (BSI), which can generate discrete graph structures in one go;
Discrete Compositional Generation via General Soft Operators and Robust Reinforcement Learning
Marco Jiralerspong (Université de Montréal), Gauthier Gidel (Université de Montréal)
GenerationTransformerReinforcement LearningFlow-based ModelTextBiomedical Data
🎯 What it does: Studied the discrete combinatorial generation problem in scientific discovery, proposing a general soft reinforcement learning operator (general mellowmax, GM) and a training algorithm based on trajectory consistency (Trajectory General Mellowmax, TGM) to generate diverse and high-quality candidate objects.
Discrete Diffusion for Bundle Construction
Teng Tu (National University of Singapore), Tat-Seng Chua (National University of Singapore)
GenerationRecommendation SystemTransformerDiffusion modelSequential
🎯 What it does: Propose a method based on discrete diffusion models for unordered set generation, applied to bundle construction in product bundling;
Discrete Diffusion for Reflective Vision-Language-Action Models in Autonomous Driving
Pengxiang Li (LiAuto), XianPeng Lang
Autonomous DrivingLarge Language ModelVision-Language-Action ModelDiffusion modelMultimodalityBenchmark
🎯 What it does: Propose the ReflectDrive framework, integrating discrete diffusion models with Vision-Language-Action (VLA) models for safe trajectory generation;
Discrete Diffusion Trajectory Alignment via Stepwise Decomposition
Jiaqi Han (Stanford University), Stefano Ermon
GenerationData SynthesisReinforcement Learning from Human FeedbackDiffusion modelTextBiomedical Data
🎯 What it does: This paper proposes an offline preference optimization method for discrete diffusion models (SDPO), which decomposes the overall trajectory alignment problem into posterior matching for each diffusion step through progressive decomposition of diffusion trajectory alignment, thereby achieving efficient and accurate reward evaluation and model fine-tuning.
Discrete Guidance Matching: Exact Guidance for Discrete Flow Matching
Zhengyan Wan (East China Normal University), Guang Cheng (University of California, Los Angeles)
GenerationData SynthesisFlow-based ModelMultimodalityBenchmark
🎯 What it does: Propose a general guidance framework for discrete flow matching models, derive precise posterior-based guidance, train the guidance network using Bregman divergence, and further enhance performance by leveraging target distribution samples through regularization.
Discrete Latent Features Ablate Adversarial Attack: A Robust Prompt Tuning Framework for VLMs
Yang Chen (Southern University of Science and Technology), Yu Zhang (Hong Kong University of Science and Technology)
ClassificationAdversarial AttackPrompt EngineeringVision Language ModelAuto EncoderImageTextMultimodalityBenchmark
🎯 What it does: To address the adversarial robustness issue in vision-language models, this paper proposes DEFEAT, an adversarial training framework based on discrete latent features, which defends against adversarial samples by reconstructing image features through VQ-VAE discretization and combining Prompt training.
Discrete Variational Autoencoding via Policy Search
Michael Drolet (Technical University of Darmstadt), Oleg Arenz (Technical University of Darmstadt)
GenerationRepresentation LearningTransformerReinforcement LearningAuto EncoderImageSequential
🎯 What it does: Proposed the DAPS (Discrete Autoencoding via Policy Search) framework, which trains discrete VAEs using a parameter-free policy search method.
Disentangled Hierarchical VAE for 3D Human-Human Interaction Generation
Zichen Geng (University of Western Australia), Ajmal Saeed Mian (University of Western Australia)
GenerationTransformerDiffusion modelAuto EncoderContrastive LearningText
🎯 What it does: Proposes a framework based on discrete hierarchical variational autoencoders and latent diffusion models to generate realistic and physically plausible 3D human-human interactive motions from text prompts;
Disentangled Representation Learning for Parametric Partial Differential Equations
Ning Liu (Lehigh University), Yue Yu (Lehigh University)
Representation LearningAuto EncoderBiomedical DataPhysics Related
🎯 What it does: Proposes DisentangO, a hyper neural operator architecture that can disentangle physical factors from neural operator parameters and simultaneously solve forward and inverse PDE problems.
Disentangled representation learning through unsupervised symmetry group discovery
Barthélémy Dang-Nhu (Sorbonne Université), Sylvain ARGENTIERI (Sorbonne Université)
Representation LearningAuto EncoderImage
🎯 What it does: This paper proposes a fully unsupervised method that first learns action representations via A-VAE and recovers symmetric group structures through clustering, then builds upon this to learn disentangled representations for linear symmetric gene decomposition.
Disentangled Robot Learning via Separate Forward and Inverse Dynamics Pretraining
Wenyao Zhang, Li Zhang
Representation LearningRobotic IntelligenceTransformerVision-Language-Action ModelDiffusion modelAuto EncoderVideo
🎯 What it does: Propose the DeFi framework, which decomposes robot learning into separated pre-training of forward visual prediction and inverse action inference, followed by end-to-end fine-tuning on downstream tasks.
Disentanglement of Variations with Multimodal Generative Modeling
Yijie Zhang (University of Iowa), Weiran Wang (University of Iowa)
GenerationData SynthesisRepresentation LearningDiffusion modelAuto EncoderContrastive LearningImageTextMultimodalityBiomedical Data
🎯 What it does: Proposed a multi-modal variational autoencoder called IDMVAE to learn disentangled representations of shared and private information
Disentangling Knowledge Representations for Large Language Model Editing
Mengqi Zhang (Shandong University), Pengjie Ren (Shandong University)
Representation LearningTransformerLarge Language ModelContrastive LearningTextBenchmark
🎯 What it does: Proposed a knowledge editing framework for LLMs called DiKE based on knowledge representation decoupling, which can insert new knowledge while preserving fine-grained knowledge unrelated to the editing target;
Disentangling Length Bias in Preference Learning via Response-Conditioned Modeling
Jianfeng Cai (University of Science and Technology of China), Houqiang Li (University of Science and Technology of China)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Propose the Response-conditioned Bradley-Terry (Rc-BT) framework, which trains reward models and Direct Preference Optimization (DPO) using length-enhanced instructions, thereby simultaneously eliminating length bias in RLHF and enhancing LLMs' ability to follow length instructions.