ICLR 2026 Papers — Page 13
International Conference on Learning Representations · 5356 papers
Disentangling the Factors of Convergence between Brains and DINOv3
Joséphine Raugel (Meta AI), Jean-Remi King
Representation LearningTransformerContrastive LearningImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Systematically train the self-supervised vision Transformer DINOv3 under variations in three factors—model scale, training volume, and image types—and evaluate its similarity to human brain visual representations using fMRI and MEG data.
DISK: Differentiable Sparse Kernel Complex for Efficient Spatially-Variant Convolution
Zhizhen Wu (Zhejiang University), Yuchi Huo (Zhejiang University)
Computational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: The paper proposes a differentiable sparse kernel complex to efficiently approximate any dense convolution kernel and achieve high-performance filtering with spatially varying convolutions.
Displacement-Resistant Extensions of DPO with Nonconvex $f$-Divergences
Idan Pipano (Technion - Israel Institute of Technology), Mohammad Ghavamzadeh (Qualcomm AI Research)
OptimizationReinforcement Learning from Human FeedbackLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper investigates the application of f-divergence in direct preference optimization (DPO) and proposes a new SQUAREDPO loss function to alleviate the problem of probabilistic displacement.
DispViT: Direct Stereo Disparity Regression with a Single-Stream Vision Transformer
Tongfan Guan (Chinese University of Hong Kong), Yun-Hui Liu (Chinese University of Hong Kong)
Depth EstimationTransformerImage
🎯 What it does: Propose DispViT, a direct disparity regression framework based on a single-stream Vision Transformer, which moves away from traditional explicit matching and cost volume construction. It employs lightweight designs such as shift-embedding tokenizer, hetero-initialization, and Disparity-Aware RoPE, and adds a refinement module after regression to achieve fine results.
DiSRouter: Distributed Self-Routing for LLM Selections
Hang Zheng, Kai Yu (Shanghai Innovation Institution)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Proposed a distributed self-routing framework called DiSRouter, enabling self-assessment and hierarchical routing among LLMs without the need for a centralized external router.
Disrupting Hierarchical Reasoning: Adversarial Protection for Geographic Privacy in Multimodal Reasoning Models
Jiaming Zhang (Nanyang Technological University), Wei Yang Bryan Lim (Nanyang Technological University)
Adversarial AttackVision Language ModelGenerative Adversarial NetworkImageMultimodality
🎯 What it does: Proposes the ReasonBreak framework, which protects the geographic privacy of personal images by disrupting the hierarchical geographic reasoning chains of multimodal large-scale reasoning models (MLRMs) through concept-aware adversarial perturbations.
Dissecting Representation Misalignment in Contrastive Learning via Influence Function
Huanyi Xie (King Abdullah University of Science and Technology), Di Wang (King Abdullah University of Science and Technology)
Representation LearningContrastive LearningImage
🎯 What it does: Proposed the Extended Contrastive Influence Function (ECIF) to evaluate the impact of positive and negative samples in contrastive learning models, enabling efficient model editing and data cleaning.
DisTaC: Conditioning Task Vectors via Distillation for Robust Model Merging
Kotaro Yoshida (Institute of Science Tokyo), Hiroki Naganuma (Mila)
Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerImage
🎯 What it does: Propose a task vector preprocessing method called DisTaC based on knowledge distillation to address the issues of task vector norm mismatch and low confidence in source models during model fusion.
DistDF: Time-series Forecasting Needs Joint-distribution Wasserstein Alignment
Eric Wang, Zhouchen Lin (Peking University)
Time Series
🎯 What it does: Propose the DistDF framework, which aligns the conditional distributions of the predicted sequence and the label sequence by minimizing the joint distribution Wasserstein distance, replacing the traditional MSE training objective.
Distillation of Large Language Models via Concrete Score Matching
Yeongmin Kim (Korea Advanced Institute of Science and Technology), Il-chul Moon
Knowledge DistillationTransformerLarge Language ModelScore-based ModelTextBenchmark
🎯 What it does: Propose a new large language model knowledge distillation method called Concrete Score Distillation (CSD), which directly aligns the difference between student and teacher logits through discrete score matching, solving the traditional issues of softmax smoothing and logit identity bias;
Distilled Pretraining: A modern lens of Data, In-Context Learning and Test-Time Scaling
Sachin Goyal (FAIR at Meta), Kartik Ahuja (FAIR at Meta)
Knowledge DistillationRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: By introducing distilled pretraining (DPT) during the pretraining phase of large language models, systematically evaluate its impact on the model's test-time scalability and in-context learning ability, analyze its internal mechanisms in a simplified bigram model (bigram sandbox), and propose a token entropy-based distillation loss pruning strategy (token routing) to mitigate performance degradation.
Distilling and Adapting: A Topology-Aware Framework for Zero-Shot Interaction Prediction in Multiplex Biological Networks
Alana Deng (Western University), Pingzhao Hu (Western University)
Knowledge DistillationRepresentation LearningDrug DiscoveryGraph Neural NetworkTransformerLarge Language ModelMixture of ExpertsContrastive LearningGraphBiomedical Data
🎯 What it does: Proposed and implemented a framework named CAZI-MBN for zero-shot interaction prediction in multi-layer biological networks, capable of simultaneously handling multi-type interactions and unseen entities.
Distilling Causal Signals for One-Shot Directed Evolution of Antibodies
Sai Pooja Mahajan (AIDD Genentech), Rajesh Ranganath (Courant Institute of Mathematical Sciences NYU)
Knowledge DistillationDrug DiscoveryGraph Neural NetworkTransformerContrastive LearningGraphBiomedical Data
🎯 What it does: Proposed a framework named AFFINITYENHANCER that improves antibody affinity using a single antibody sequence without prior information about unseen antigens;
Distilling the Thought, Watermarking the Answer: A Principle Semantic Guided Watermark for Reasoning Large Language Models
Shuliang Liu (Hong Kong University of Science and Technology (Guangzhou)), Xuming Hu (Hong Kong University of Science and Technology (Guangzhou))
TransformerLarge Language ModelText
🎯 What it does: This paper proposes ReasonMark, a two-phase watermark framework designed for inference-oriented LLMs, which first preserves the model's internal reasoning process and then uses key semantic vectors to guide watermark embedding during the answering phase, maintaining logical coherence.
Distilling to Hybrid Attention Models via KL-Guided Layer Selection
Yanhong Li (Allen Institute for Ai), Yoon Kim (Allen Institute for Ai)
Knowledge DistillationRepresentation LearningTransformerText
🎯 What it does: The paper proposes a layer selection method based on KL-divergence, distilling a pre-trained softmax Transformer into an efficient hybrid attention (softmax + linear) architecture.
DistillKac: Few-Step Image Generation via Damped Wave Equations
Weiqiao Han (MIT), Stefano Ermon (Stanford)
GenerationDiffusion modelScore-based ModelImagePhysics RelatedStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Propose DistillKac, which achieves finite-speed image generation via Kac dynamics from the damped wave equation, and enables few-step sampling through endpoint distillation;
DistMLIP: A Distributed Inference Platform for Machine Learning Interatomic Potentials
Kevin Han (Carnegie Mellon University), Gerbrand Ceder (UC Berkeley)
Graph Neural NetworkGraphBenchmarkPhysics Related
🎯 What it does: Proposed a distributed inference platform named DistMLIP, which performs multi-GPU parallel inference for machine learning interatomic potentials (MLIP) using graph-level non-redundant partitioning.
Distractor-free Generalizable 3D Gaussian Splatting
Yanqi Bao (Nanjing University), Yang Gao (Nanjing University)
RestorationGaussian SplattingImage
🎯 What it does: Proposes the DGGS framework to eliminate artifacts in 3D Gaussian Splatting (3DGS), supporting generalized training and inference.
Distributed Algorithms for Euclidean Clustering
Vincent Cohen-Addad (Google Research), Samson Zhou (Texas A&M University)
OptimizationImagePoint Cloud
🎯 What it does: Studied the (1+ε)-coreset construction for distributed Euclidean (k,z)-clustering, and provided optimal communication protocols under the coordinator model and blackboard model.
Distribution-Aware Multi-Granularity Phase Coding: Towards Lower Conversion Error for Spike-Driven Large Language Models
Hanyuan Zheng (Jilin University), Bin Gu (Jilin University)
Computational EfficiencyKnowledge DistillationSpiking Neural NetworkTransformerLarge Language ModelText
🎯 What it does: Developed a distribution-aware multi-granularity phase encoding and an ANN-to-SNN conversion paradigm based on this encoding for low-error, low-energy spike-driven large language models (LLMs).
Distribution-informed Online Conformal Prediction
Dongjian Hu (Nankai University), Changliang Zou (Nankai University)
OptimizationTime SeriesBenchmarkFinance Related
🎯 What it does: Proposed an online adaptive conformal prediction method called Conformal Optimistic Prediction (COP), which dynamically adjusts prediction interval thresholds by estimating the cumulative distribution function (CDF) of nonconformity scores.
Distributional Consistency Loss: Beyond Pointwise Data Terms in Inverse Problems
George Webber (King's College London), Andrew J. Reader (King's College London)
RestorationBiomedical DataPositron Emission Tomography
🎯 What it does: This paper proposes a distribution consistency loss (DC Loss), which evaluates the residual of inverse problems using statistical consistency rather than point-wise matching, and verifies its effectiveness in unsupervised depth image prior (DIP) denoising and PET reconstruction tasks.
Distributional Equivalence in Linear Non-Gaussian Latent-Variable Cyclic Causal Models: Characterization and Learning
Haoyue Dai, Kun Zhang
Explainability and InterpretabilityComputational EfficiencyTabularTime SeriesFinance Related
🎯 What it does: Proposed a learning algorithm for determining distributional equivalence graphs and structure-free assumptions in linear non-Gaussian latent variable cyclic causal models
Distributional Machine Unlearning via Selective Data Removal
Youssef Allouah (Stanford University), Sanmi Koyejo (EPFL)
Safty and PrivacyImageText
🎯 What it does: Proposes a distribution-level machine learning forgetting framework that uses selective data deletion to eliminate the impact of a specific subgroup while preserving the statistical properties of other subgroups.
Distributional value gradients for stochastic environments
Baptiste Debes (KU Leuven), Tinne Tuytelaars (KU Leuven)
Reinforcement LearningAuto Encoder
🎯 What it does: Propose the Distributional Sobolev Deterministic Policy Gradient (DSDPG) algorithm, which can simultaneously learn the return distribution and its gradient in continuous action spaces, and achieve gradient-level Temporal Difference training through Sobolev Bellman backup.
Distributional Vision-Language Alignment by Cauchy-Schwarz Divergence
Wenzhe Yin (University of Amsterdam), Stratis Gavves
RetrievalVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Propose CS-Aligner, a distributed vision-language alignment framework that combines Cauchy-Schwarz (CS) divergence with mutual information.
Distributionally Robust Classification for Multi-source Unsupervised Domain Adaptation
Seonghwi Kim (Pohang University of Science and Technology), Minwoo Chae (Pohang University of Science and Technology)
ClassificationDomain AdaptationImage
🎯 What it does: Proposes a multi-source unsupervised domain adaptation framework based on distributionally robust optimization, achieving modeling of target distribution uncertainty by constructing an ambiguity set through conditional mixing and target input perturbation.
Distributionally Robust Cooperative Multi-agent Reinforcement Learning with Value Factorization
Chengrui Qu (California Institute of Technology), Adam Wierman (California Institute of Technology)
Recurrent Neural NetworkReinforcement LearningBenchmark
🎯 What it does: Proposed a distributed robust IGM (DrIGM) and implemented a robust value factorization algorithm, enhancing the performance of CTDE collaborative RL under environmental uncertainty.
Distributionally Robust Linear Regression with Block Lewis Weights
Naren Sarayu Manoj (Toyota Technological Institute), Kumar Kshitij Patel (Yale University)
OptimizationTabular
🎯 What it does: Proposed a new algorithm for solving the empirical distribution robust (GDR) least squares problem with multiple data groups, achieving balanced control of errors in high-dimensional linear regression.
Distributionally Robust Optimization via Generative Ambiguity Modeling
JIAQI WEN, Jianyi Yang (University of Houston)
OptimizationReinforcement LearningDiffusion modelAuto EncoderImageTime Series
🎯 What it does: This paper proposes a generative adversarial distribution robust optimization (GAS-DRO) framework based on generative models, utilizing diffusion models or VAEs to generate adversarial distributions and constructing a solvable ambiguity set through constrained reconstruction loss, providing complete algorithms and convergence proofs.
Distributions as Actions: A Unified Framework for Diverse Action Spaces
Jiamin He (University of Alberta), Martha White (University of Alberta)
Reinforcement Learning
🎯 What it does: The study treats action distribution parameters as actions, unifying a reinforcement learning (RL) framework for any action space.
DIVA-GRPO: Enhancing Multimodal Reasoning through Difficulty-Adaptive Variant Advantage
Haowen Gao (State Key Laboratory of AI Safety Institute of Computing Technology Chinese Academy of Sciences), Xueqi Cheng (State Key Laboratory of AI Safety Institute of Computing Technology Chinese Academy of Sciences)
OptimizationTransformerLarge Language ModelReinforcement LearningMultimodalityBenchmark
🎯 What it does: Developed a GRPO framework named DIVA-GRPO based on dynamic difficulty adaptive variant advantage to enhance the long-chain reasoning capability of multimodal large language models.
DiVE-k: DIFFERENTIAL VISUAL REASONING FOR FINE-GRAINED IMAGE RECOGNITION
Raja Kumar (University of Southern California), Ram Nevatia (University of Southern California)
ClassificationRecognitionSupervised Fine-TuningReinforcement LearningVision Language ModelImage
🎯 What it does: Propose the DiVE-k framework, which utilizes the top-k generation results from large vision-language models as training signals for multiple-choice questions. The model is trained via reinforcement learning to perform differentiated visual reasoning, thereby enhancing fine-grained image recognition performance.
DiVeQ: Differentiable Vector Quantization Using the Reparameterization Trick
Mohammad Hassan Vali, Arno Solin (Aalto University)
GenerationCompressionAuto EncoderGenerative Adversarial NetworkImageAudio
🎯 What it does: Proposed two differentiable vector quantization methods, DiVeQ and SF-DiVeQ, and evaluated them on multiple tasks including image compression, image generation, and speech decoding.
Divergence-Free Neural Networks with Application to Image Denoising
Sébastien Herbreteau (Univ Rennes), Etienne Meunier (Inria)
RestorationConvolutional Neural NetworkImage
🎯 What it does: Propose a zero-divergence neural network architecture (CENSURE) achieved through structured conservative field combinations, which can be directly applied to unsupervised image denoising;
Diverse and Sparse Mixture-of-Experts for Causal Subgraph–Based Out-of-Distribution Graph Learning
Jerry Sun (University of Toronto), Chi-Guhn Lee (University of Toronto)
Domain AdaptationRepresentation LearningGraph Neural NetworkMixture of ExpertsGraphBenchmark
🎯 What it does: Proposes DiSCO, a graph neural network framework based on Mixture-of-Experts (MoE), which directly models instance-level heterogeneous causal subgraphs to achieve out-of-distribution (OOD) generalization on graph structures without relying on environment labels or causal assumptions.
Diverse Dictionary Learning
Yujia Zheng (Carnegie Mellon University), Kun Zhang (Carnegie Mellon University)
Representation LearningDiffusion modelAuto EncoderGenerative Adversarial NetworkImageBenchmark
🎯 What it does: Propose a 'diversity dictionary learning' framework to study the identifiability of latent variable information under extremely loose assumptions, and provide theoretical identifiability for set operations such as intersection, complement, and symmetric difference. Further elaborate that element-level identifiability can be achieved when sufficient diversity conditions are met.
Diverse Text Decoding via Iterative Reweighting
Ruiqi Shi (Chinese University of Hong Kong), Sinno Jialin Pan (Chinese University of Hong Kong)
GenerationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposes OverRIDE—a decoding method based on iterative reweighting—which significantly enhances LLM output diversity by real-time capturing and suppressing already generated semantic patterns through fine-tuning a low-rank output head adapter during inference.
Diverse Text-to-Image Generation via Contrastive Noise Optimization
Byungjun Kim (KAIST), Jong Chul Ye (KAIST)
GenerationDiffusion modelContrastive LearningImageText
🎯 What it does: Proposed a method called Contrastive Noise Optimization (CNO) for optimizing the initial noise in text-to-image diffusion models, enhancing generation diversity by applying InfoNCE loss in Tweedie space.
DIVERSE: Disagreement-Inducing Vector Evolution for Rashomon Set Exploration
Gilles Eerlings (Hasselt University), Kris Luyten (Hasselt University)
OptimizationComputational EfficiencyHyperparameter SearchConvolutional Neural NetworkImage
🎯 What it does: Propose the DIVERSE framework, which inserts FiLM layers into a pre-trained network and uses CMA-ES to search in a low-dimensional modulation space, generating a diverse Rashomon set without requiring retraining.
Diversified Multinomial Logit Contextual Bandits
Heesang Ann (Seoul National University), Min-hwan Oh (Seoul National University)
OptimizationReinforcement LearningTabular
🎯 What it does: Proposed a hierarchical contextual compromise model combining the Polynomial Logit (MNL) model with a submodular diversity function, and designed the OFU-DMNL algorithm to address its online learning and combinatorial optimization problems.
Diversity-Enhanced Reasoning for Subjective Questions
Yumeng Wang (Hong Kong University of Science and Technology), Yi R. Fung (Hong Kong University of Science and Technology)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
🎯 What it does: Designed and trained the MultiRole-R1 framework, enhancing the performance of large reasoning models on subjective reasoning tasks through self-supervised synthesis of multi-role reasoning chains and diversity reward shaping.
Diversity-Incentivized Exploration for Versatile Reasoning
Zican Hu (Nanjing University), Zhi Wang (Nanjing University)
Reinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: Propose a reinforcement learning framework called DIVER based on global sequence diversity to enhance LLM reasoning capabilities.
Divid: Disentangled Spatial-Temporal Modeling within LLMs for Temporally Grounded Video Understanding
Yepeng Tang (Beijing Jiaotong University), Jing Liu (Chinese Academy of Sciences)
RecognitionTransformerLarge Language ModelVision Language ModelVideoTextMultimodality
🎯 What it does: Proposed the Divid framework, which explicitly separates spatial and temporal modeling in the LLM decoder, achieving efficient video understanding through time-guided keyframe selection and soft routing.
Divide and Abstract: Autoformalization via Decomposition and Abstraction Learning
Marcus J. Min (University of Pennsylvania), Osbert Bastani (University of Pennsylvania)
AI Code AssistantLarge Language ModelTextBenchmark
🎯 What it does: Proposes a zero-training automatic formalization framework DNA, which first learns generic abstractions from the corpus and then hierarchically decomposes sentences to generate formalizations.
Divide, Harmonize, Then Conquer It: Shooting Multi-Commodity Flow Problems with Multimodal Language Models
Xinyu Yuan (Zhejiang University), Wenzhi CHEN
OptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelTextMultimodalityGraph
🎯 What it does: The paper proposes a multi-commodity traffic allocation method called PRAM based on a multimodal large language model, achieving fast and high-quality allocation through partitioning and multi-agent reinforcement learning.
DM4CT: Benchmarking Diffusion Models for Computed Tomography Reconstruction
Jiayang Shi (Leiden University), Joost Batenburg
RestorationDiffusion modelImageBiomedical DataComputed TomographyBenchmark
🎯 What it does: Proposed and implemented the DM4CT Benchmark, systematically evaluated the performance of diffusion models in medical, industrial, and synchrotron CT reconstruction, and released a high-resolution synchrotron CT dataset.
DMAP: A Distribution Map for Text
Tom Kempton (University of Manchester), Stuart Burrell (Visa Inc.)
Anomaly DetectionExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose the DMAP (Distribution Map for Text) method, which maps text through a language model to sampling points within the 0~1 interval, jointly encoding word ranking and probability information for text statistics and analysis.
DND: Boosting Large Language Models with Dynamic Nested Depth
Tieyuan Chen (Shanghai Jiao Tong University), Jianguo Li (Inclusion AI, Ant Group)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose Dynamic Nested Depth (DND), which dynamically selects key tokens in Transformer layers using routers and applies additional nested processing to enhance LLM performance.
DNT: a Deeply Normalized Transformer that can be trained by Momentum SGD
Xianbiao Qi (Intellifusion Inc), Zhouchen Lin (Peking University)
ClassificationGenerationOptimizationComputational EfficiencyTransformerLarge Language ModelImageText
🎯 What it does: Propose Deeply Normalized Transformer (DNT), which achieves performance comparable to AdamW by appropriately adding and positioning normalization layers in the Transformer to concentrate gradient distribution, enabling training with standard momentum SGD (mSGDW).
Do 3D Large Language Models Really Understand 3D Spatial Relationships?
Xianzheng Ma (University of Oxford), Victor Adrian Prisacariu (University of Oxford)
Explainability and InterpretabilityLarge Language ModelSupervised Fine-TuningImageMultimodalityPoint CloudBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes a more rigorous 3D question-answering benchmark, Real-3DQA, and introduces perspective rotation evaluation and 3D reweighted fine-tuning methods
Do Large Language Models Know What They Are Capable Of?
Casey O. Barkan (RAND Corporation), Oliver Sourbut (Future of Life Foundation)
AI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper evaluates the performance of large language models (LLMs) in predicting their own capabilities, learning from experience, and updating confidence in multi-step tasks through three experiments.
Do LLM Agents Know How to Ground, Recover, and Assess? Evaluating Epistemic Competence in Information-Seeking Agents
Jiaqi Shao (Duke Kunshan University), Bing Luo (Duke Kunshan University)
RetrievalLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: Propose SeekBench, a process-level evaluation framework for search agents in large language models, which uses expert-annotated trajectory data to measure the agents' knowledge retrieval and reasoning processes.
Do LLMs Forget What They Should? Evaluating In-Context Forgetting in Large Language Models
Yuli Qian (Peking University), Yutao Xie (Microsoft STC Asia)
Large Language ModelPrompt EngineeringTextSequentialBenchmark
🎯 What it does: Proposed and implemented the ICF-Bench benchmark for systematically evaluating the selective forgetting (in-context forgetting) capabilities of large language models during reasoning.
Do Not Let Low-Probability Tokens Over-Dominate in RL for LLMs
Zhihe Yang (Chinese University of Hong Kong), Yunjian Xu (Chinese University of Hong Kong)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Identify and mitigate the dominance of low-probability tokens over gradients in RL training, thereby improving the performance of large language models on reasoning tasks.
Do Vision-Language Models Respect Contextual Integrity in Location Disclosure?
Ruixin Yang (Georgia Institute of Technology), Alan Ritter (Georgia Institute of Technology)
Safty and PrivacyLarge Language ModelPrompt EngineeringVision Language ModelImageMultimodalityBenchmarkChain-of-Thought
🎯 What it does: This paper proposes the VLM-GEOPRIVACY benchmark to evaluate whether visual language models maintain contextual integrity when disclosing image geographical locations.
Do We Need All the Synthetic Data? Targeted Image Augmentation via Diffusion Models
Dang Nguyen (University of California Los Angeles), Baharan Mirzasoleiman (University of California Los Angeles)
Object DetectionGenerationData SynthesisConvolutional Neural NetworkTransformerDiffusion modelImage
🎯 What it does: Propose a directional diffusion model data augmentation framework called TADA, which synthesizes augmented samples only for those that are not quickly learned during training;
Do We Really Need Permutations? Impact of Model Width on Linear Mode Connectivity
Akira Ito (Tohoku University), Atsutoshi Kumagai (NTT Computer and Data Science Laboratories)
ClassificationConvolutional Neural NetworkImage
🎯 What it does: Achieve linear mode connectivity (LMC) by expanding model width without parameter permutation, and propose hierarchical exponential weighted connectivity (LEWC) to explain this phenomenon;
Doctor-R1: Mastering Clinical Inquiry with Experiential Agentic Reinforcement Learning
Yunghwei Lai (Tsinghua University), Yang Liu (Tsinghua University)
Reinforcement Learning from Human FeedbackReinforcement LearningAgentic AITextBiomedical DataBenchmark
🎯 What it does: Proposed and implemented DOCTOR-R1—a doctor agent based on experience-driven agent reinforcement learning, capable of performing strategic questioning in multi-round interactions and making accurate diagnoses.
Does “Do Differentiable Simulators Give Better Policy Gradients?” Give Better Policy Gradients?
Ku Onoda (University of Tokyo), Yutaka Matsuo (University of Tokyo)
Reinforcement Learning from Human FeedbackReinforcement LearningWorld Model
🎯 What it does: Studied gradient estimation methods under differentiable simulators in reinforcement learning, and proposed two new gradient hybrid strategies: breakpoint detection-based hybrid gradient (DDCG) and stepwise inverse variance weighting (IVW-H).
Does FLUX Already Know How to Perform Physically Plausible Image Composition?
Shilin Lu (Nanyang Technological University), Adams Wai-Kin Kong (Nanyang Technological University)
Image HarmonizationGenerationVision Language ModelDiffusion modelImageBenchmark
🎯 What it does: Proposed a training-free image synthesis framework called SHINE, which can seamlessly insert target objects into complex lighting and high-resolution scenes while maintaining background integrity;
Does Higher Interpretability Imply Better Utility? A Pairwise Analysis on Sparse Autoencoders
Xu Wang (University of Hong Kong), Difan Zou (University of Hong Kong)
Explainability and InterpretabilityLarge Language ModelAuto EncoderText
🎯 What it does: Investigate the relationship between interpretability and control performance of sparse autoencoders (SAE) in large language models, systematically evaluating the performance of 90 SAEs across different models, architectures, and sparsity levels; propose selecting efficient control features via ∆Token Confidence and compare the control effectiveness of different feature selection methods; further analyze the association between interpretability and control performance after feature selection.
Does the Data Processing Inequality Reflect Practice? On the Utility of Low-Level Tasks
Roy Turgeman (Bar-Ilan University), Tom Tirer (Bar-Ilan University)
ClassificationRepresentation LearningData-Centric LearningConvolutional Neural NetworkAuto EncoderContrastive LearningImageBenchmark
🎯 What it does: Under limited sample conditions, it is proven that even strong classifiers close to Bayes optimal can still improve classification accuracy after low-level preprocessing of the original data (such as dimensionality reduction, denoising, or encoding).
Does Weak-to-strong Generalization Happen under Spurious Correlations?
Chenruo Liu (New York University), Qi Lei (New York University)
ClassificationKnowledge DistillationConvolutional Neural NetworkTransformerSupervised Fine-TuningImage
🎯 What it does: This paper investigates the generalization behavior of weak-to-strong (W2S) transfer learning under pseudo-correlation (group imbalance), providing theoretical analysis and validation on real-world visual tasks.
DoFlow: Flow-based Generative Models for Interventional and Counterfactual Forecasting on Time Series
Dongze Wu (Georgia Institute of Technology), Yao Xie (Georgia Institute of Technology)
GenerationAnomaly DetectionRecurrent Neural NetworkFlow-based ModelTime SeriesBiomedical DataOrdinary Differential Equation
🎯 What it does: Propose a causal temporal prediction framework DoFlow based on Continuous Normalizing Flows (CNF), which can perform observation, intervention, and counterfactual prediction on known causal DAGs, and provides explicit trajectory likelihood for anomaly detection.
Doloris: Dual Conditional Diffusion Implicit Bridges with Sparsity Masking Strategy for Unpaired Single-Cell Perturbation Estimation
Changxi Chi (Zhejiang University Westlake University), Stan Z. Li (Hong Kong University of Science and Technology)
Drug DiscoveryDiffusion modelBiomedical Data
🎯 What it does: Propose a new framework called Doloris, which uses a dual-conditional diffusion model and a sparse mask strategy to predict cellular responses on unpaired single-cell perturbation data.
Domain Expansion: A Latent Space Construction Framework for Multi-Task Learning
Chi-Yao Huang (Arizona State University), Yezhou Yang (Arizona State University)
ClassificationPose EstimationRepresentation LearningContrastive LearningImagePoint CloudMesh
🎯 What it does: Propose the Domain Expansion framework, which maps multi-task objectives into mutually orthogonal latent subspaces through orthogonal pooling, avoiding latent representation collapse.
Don't Just Fine-tune the Agent, Tune the Environment
Siyuan Lu (Inclusion AI, Ant Group), Tao Lin (Westlake University)
Supervised Fine-TuningReinforcement LearningAgentic AIPrompt EngineeringText
🎯 What it does: Proposed the ENVIRONMENT TUNING training framework, enabling LLM agents to learn multi-round tool usage through environmental interaction rather than trajectory replication, even with minimal training data.
Don’t Pass@k: A Bayesian Framework for Large Language Model Evaluation
Mohsen Hariri (Case Western Reserve University), Vipin Chaudhary (Case Western Reserve University)
Large Language ModelTextBenchmark
🎯 What it does: Propose a Bayesian framework based on Dirichlet priors to replace traditional evaluation metrics such as Pass@k and avg@N, enabling posterior estimation and confidence interval inference for LLM performance.
Don't Settle Too Early: Self-Reflective Remasking for Diffusion Language Models
Zemin Huang (Zhejiang University), Guo-Jun Qi (MAPLE Lab, Westlake University)
GenerationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningDiffusion modelText
🎯 What it does: Proposes a self-reflective Remask mechanism that enables diffusion language models to identify and re-mask erroneous words during the generation process, achieving multi-round text correction.
Don't Shift the Trigger: Robust Gradient Ascent for Backdoor Unlearning
Xingyi Zhao (Utah State University), Shuhan Yuan (Utah State University)
ClassificationSafty and PrivacyTransformerSupervised Fine-TuningText
🎯 What it does: Investigated and addressed the trigger drift problem in backdoor removal methods based on gradient ascent (GA) in text classification tasks, proposing a robust gradient ascent (RGA) framework.
Don't Throw Away Your Beams: Improving Consistency-based Uncertainties in LLMs via Beam Search
Ekaterina Fadeeva (ETH Zurich), Maxim Panov (Mohamed bin Zayed University of Artificial Intelligence)
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: The study uses beam search to generate candidate answers and constructs an importance-weighted consistency uncertainty estimator to replace traditional polynomial sampling methods.
Don't Throw Away Your Pretrained Model
Shangbin Feng (University of Washington), Dong Yu (Tencent AI Seattle Lab)
GenerationComputational EfficiencyTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: Developed a reasoning strategy called SWITCH GENERATION based on model checkpoint collaboration, using a switcher LM to dynamically select pre-trained, fine-tuned, and aligned models to generate text fragments.
DOPPLER: Dual-Policy Learning for Device Assignment in Asynchronous Dataflow Graphs
Xinyu Yao (Rice University), Chris Jermaine
OptimizationGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: In asynchronous work-conserving (WC) systems, the DOPPLER framework is proposed to address the device allocation problem for dataflow graph (Dataflow Graph) computations, using dual strategies (selection strategy SEL and placement strategy PLC) to learn optimal operation sequences and device allocation schemes through reinforcement learning;
Doubly-Regressing Approach for Subgroup Fairness
Kunwoong Kim (KAIST), Yongdai Kim (Seoul National University)
Generative Adversarial NetworkTextTabularBenchmark
🎯 What it does: Proposes the concept of subgroup subset fairness and develops the DRAF (Doubly Regressing Adversarial learning for Fairness) algorithm, achieving distributed subgroup fairness and marginal fairness in scenarios with sparse and high-dimensional sensitive attributes using a single discriminator.
Doubly-Robust LLM-as-a-Judge: Externally Valid Estimation with Imperfect Personas
Luke Guerdan (Carnegie Mellon University), Steven Wu
Data-Centric LearningTransformerLarge Language ModelText
🎯 What it does: This paper proposes a doubly robust estimation framework that combines persona scores generated by LLMs with human ratings affected by bias, estimating quality metrics of generative AI systems under the target distribution to address external validity issues.
DoVer: Intervention-Driven Auto Debugging for LLM Multi-Agent Systems
Ming Ma (Institute of Neuroscience, State Key Laboratory of Brain Cognition and Brain-inspired Intelligence Technology, Center for Excellence in Brain Science and Intelligence Technology, Chinese Academy of Sciences), Dongmei Zhang (Microsoft)
Large Language ModelAgentic AIPrompt EngineeringTextSequential
🎯 What it does: Proposes DoVer—a framework for automatically debugging and repairing errors in multi-agent systems by identifying failure points and executing targeted interventions (e.g., editing messages or plans) to verify and fix errors.
Downgrade to Upgrade: Optimizer Simplification Enhances Robustness in LLM Unlearning
Yicheng Lang (Michigan State University), Sijia Liu (Michigan State University)
OptimizationSafty and PrivacyLarge Language ModelTextBenchmark
🎯 What it does: Investigate the impact of optimizer level on the robustness of unlearning in large language models, and propose a hybrid FO-ZO optimization strategy to enhance robustness
Doxing via the Lens: Revealing Location-related Privacy Leakage on Multi-modal Large Reasoning Models
Weidi Luo (University of Georgia), Chaowei Xiao (John Hopkins University)
Safty and PrivacyTransformerLarge Language ModelVision Language ModelImageMultimodalityBenchmarkChain-of-Thought
🎯 What it does: This paper studies the risk of geographical location information leakage in images through multimodal large inference models.
DP-Fusion: Token-Level Differentially Private Inference for Large Language Models
Rushil Thareja (Mohamed bin Zayed University of Artificial Intelligence), Nils Lukas (Mohamed bin Zayed University of Artificial Intelligence)
Safty and PrivacyTransformerLarge Language ModelText
🎯 What it does: Proposed a LLM inference mechanism named DP-FUSION that provides explicit differential privacy (DP) guarantees for sensitive tokens in documents while maintaining readability and text quality.
DPad: Efficient Diffusion Language Models with Suffix Dropout
Xinhua Chen (Duke University), Yiran Chen (Duke University)
Computational EfficiencyAI Code AssistantTransformerLarge Language ModelDiffusion modelText
🎯 What it does: Designed and implemented a training-free inference strategy called DPad, which significantly reduces the computational cost of dLLM's suffix attention by pre-trimming suffix tokens using a sliding window and distance-decay dropout, thereby improving inference efficiency.
dParallel: Learnable Parallel Decoding for dLLMs
Zigeng Chen (National University of Singapore), Xinchao Wang (National University of Singapore)
Computational EfficiencyKnowledge DistillationTransformerSupervised Fine-TuningText
🎯 What it does: Propose the dParallel method, achieving highly parallel decoding for dLLM through certainty-forcing distillation, significantly reducing decoding steps;
DPQuant: Efficient and Private Model Training via Dynamic Quantization Scheduling
Yubo Gao (University of Toronto), Nandita Vijaykumar (University of Toronto)
Safty and PrivacyComputational EfficiencyConvolutional Neural NetworkTransformerImageText
🎯 What it does: Developed DPQUANT, a dynamic quantization framework for differential privacy training;
DR-GGAD: Dual Residual Centering for Mitigating Anomaly Non‑Discriminativity in Generalist Graph Anomaly Detection
Changlong Fu (Yunnan University), Yun Yang (Yunnan University)
Anomaly DetectionGraph Neural NetworkGraph
🎯 What it does: Under the premise of no target fine-tuning, the DR-GGAD model is proposed to alleviate the anomaly indistinguishability problem in cross-domain graph anomaly detection through dual residual centers.
DR-SAC: Distributionally Robust Soft Actor-Critic for Reinforcement Learning under Uncertainty
Mingxuan Cui (New York University), Zhengyuan Zhou (New York University)
Reinforcement LearningAuto Encoder
🎯 What it does: Developed DR-SAC, an offline distributed robust soft actor-critic algorithm for continuous action spaces;
DR-Submodular Maximization with Stochastic Biased Gradients: Classical and Quantum Gradient Algorithms
Shengminjie Chen (State Key Lab of Processors, Institute of Computing Technology, Chinese Academy of Sciences), Zihan Zhao (State Key Lab of Processors, Institute of Computing Technology, Chinese Academy of Sciences)
OptimizationBenchmark
🎯 What it does: Study the algorithm and theoretical analysis for maximizing continuous DR-submodular functions in environments with random biased gradients.
Dr.LLM: Dynamic Layer Routing in LLMs
Ahmed Heakl (Parameter Lab), Seong Joon Oh (Parameter Lab)
Computational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Designed and implemented Dr.LLM, a dynamic layer routing framework that can be attached to frozen LLMs, enabling the model to decide to skip, execute, or repeat individual layers based on input, thereby reducing computational cost while maintaining or improving accuracy.
Draft-based Approximate Inference for LLMs
Kevin Galim (FuriosaAI), Kangwook Lee (University of Wisconsin-Madison)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelText
🎯 What it does: Proposed an approximate inference framework based on a draft model, utilizing lookahead to predict future outputs, thereby enhancing the accuracy of KV cache trimming and prompt compression, and further introduced SpecKV, SpecPC, and their cascaded solution SpecKV-PC.
DragFlow: Unleashing DiT Priors with Region-Based Supervision for Drag Editing
Zihan Zhou (Nanyang Technological University), Adams Wai-Kin Kong (Nanyang Technological University)
Image HarmonizationGenerationTransformerLarge Language ModelDiffusion modelFlow-based ModelImageBenchmark
🎯 What it does: Proposes the DragFlow framework, enabling drag-based image editing using Diffusion Transformer (e.g., FLUX), combined with region-level supervision, hard constraint background preservation, and adapter-enhanced inversion;
Dragging with Geometry: From Pixels to Geometry-Guided Image Editing
Xinyu Pu (Southeast University), Pan Zhou (Singapore Management University)
Depth EstimationDiffusion modelImageBenchmark
🎯 What it does: Proposes GeoDrag, a one-step drag-and-drop image editing framework that integrates 3D geometric information with 2D pixel plane information.
DRAGON: Guard LLM Unlearning in Context via Negative Detection and Reasoning
Yaxuan Wang (University of California, Santa Cruz), Yang Liu (University of California, Santa Cruz)
Safty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Propose a pre-inference LLM forgetting framework (DRAGON) that detects and intervenes through prompting, enabling the forgetting of privacy and harmful knowledge without modifying model weights
Draw-In-Mind: Rebalancing Designer-Painter Roles in Unified Multimodal Models Benefits Image Editing
Ziyun Zeng (Show Lab, National University of Singapore), Mike Zheng Shou (Show Lab, National University of Singapore)
GenerationTransformerVision Language ModelFlow-based ModelImageTextMultimodalityChain-of-Thought
🎯 What it does: Proposed and trained the Draw-In-Mind (DIM) dataset and model, emphasizing shifting design responsibility to the understanding module within a unified multimodal model to enhance image editing performance.
DRBench: A Realistic Benchmark for Enterprise Deep Research
Amirhossein Abaskohi (ServiceNow Research), Issam H. Laradji (ServiceNow Research)
TransformerLarge Language ModelAgentic AIMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes DRBench, a benchmark for enterprise deep research tasks, evaluating AI agents using real-world enterprise environments and multimodal data;
DreamCS: Geometry-Aware Text-to-3D Generation with Unpaired 3D Reward Supervision
Xiandong Zou (Singapore Management University), Pan Zhou (Singapore Management University)
GenerationData SynthesisReinforcement Learning from Human FeedbackTransformerReinforcement LearningVision Language ModelDiffusion modelScore-based ModelContrastive LearningTextMesh
🎯 What it does: Propose the DreamCS framework, which integrates the 3D reward model RewardCS into the text-to-3D generation process, achieving geometry preference-guided 3D generation;
DreamOn: Diffusion Language Models For Code Infilling Beyond Fixed-size Canvas
Zirui Wu (University of Hong Kong), Lingpeng Kong (University of Hong Kong)
GenerationAI Code AssistantSupervised Fine-TuningDiffusion modelText
🎯 What it does: Propose the DREAMON framework, which introduces [expand] and [delete] special states into diffusion language models to achieve dynamic length generation, solving the fixed-length limitation.
DreamPhase: Offline Imagination and Uncertainty-Guided Planning for Large-Language-Model Agents
Shayan Mohajer Hamidi (Stanford University), Konstantinos N. Plataniotis (University of Toronto)
TransformerLarge Language ModelAgentic AIPrompt EngineeringWorld ModelText
🎯 What it does: Develop a planning framework called DREAMPHASE based on offline internal imagination, which injects reflection into LLMs by leveraging a latent world model and uncertainty-guided value assessment to improve decision-making.
DreamSwapV: Mask-guided Subject Swapping for Any Customized Video Editing
Weitao Wang (Tsinghua University), Hao Zhang (ByteDance)
GenerationTransformerDiffusion modelAuto EncoderVideoBenchmark
🎯 What it does: Developed an end-to-end video subject replacement framework called DreamSwapV based on mask-guided approaches.
DRIFT-Net: A Spectral-Coupled Neural Operator for PDEs Learning
Jiayi Li (University of New South Wales), Flora D. Salim (University of New South Wales)
Convolutional Neural NetworkPhysics Related
🎯 What it does: Designed a dual-branch spectral-spatial coupled neural operator, DRIFT-NET, for efficiently learning PDE dynamics and replacing traditional window attention modules;
DRIFT: Decompose, Retrieve, Illustrate, then Formalize Theorems
Meiru Zhang (University of Cambridge), Gerasimos Lampouras (Huawei)
AI Code AssistantTransformerLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed the DRIFT framework, addressing the dual challenges of knowledge and syntax in automatic formalization by decomposing natural language mathematical statements into sub-queries, retrieving dependencies, instantiating theorems, and ultimately achieving automatic formalization.
DRIFT: Divergent Response in Filtered Transformations for Robust Adversarial Defense
Amira Guesmi (New York University Abu Dhabi), Muhammad Shafique (New York University Abu Dhabi)
Adversarial AttackConvolutional Neural NetworkTransformerImage
🎯 What it does: Proposes a lightweight, learnable set of pre-filters (DRIFT) that enhances adversarial robustness by maximizing gradient separation between different filters (gradient consensus disruption).