ICLR 2026 Papers — Page 11
International Conference on Learning Representations · 5356 papers
DanceTogether: Generating Interactive Multi-Person Video without Identity Drifting
Junhao Chen (Tsinghua University), Ruqi Huang (Nanjing University)
GenerationData SynthesisPose EstimationVision Language ModelDiffusion modelImageVideoMultimodality
🎯 What it does: Proposed the DanceTogether framework, capable of generating multi-actor interaction videos from a single reference image and independent pose-mask sequences, maintaining identity consistency and action coherence.
Dancing in Chains: Strategic Persuasion in Academic Rebuttal via Theory of Mind
Zhitao He (Hong Kong University of Science and Technology), Yi R. Fung (Hong Kong University of Science and Technology)
TransformerLarge Language ModelReinforcement LearningTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes a rebuttal framework called RebuttalAgent based on Theory of Mind, achieving hierarchical modeling of reviewer intentions, strategy formulation, and evidence-driven response generation.
DARE-bench: Evaluating Modeling and Instruction Fidelity of LLMs in Data Science
Fan Shu (University of Houston), Feng Yan (University of Houston)
Data SynthesisData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTabularTime SeriesBenchmark
🎯 What it does: Built and released DARE-Bench, an executable and trainable multi-task data science benchmark for evaluating LLMs' performance in instruction following and machine learning modeling;
Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents
Jenny Zhang (University of British Columbia), Jeff Clune (University of British Columbia)
Meta LearningAI Code AssistantTransformerLarge Language ModelAgentic AITextBenchmark
🎯 What it does: Developed the Darwin Gödel Machine (DGM), a self-improving system based on large language models that enhances its performance in coding tasks through iterative self-modification and open exploration.
DASH: Deterministic Attention Scheduling for High-throughput Reproducible LLM Training
Xinwei Qiang (Shanghai Jiao Tong University), Minyi Guo (ByteDance Seed)
Computational EfficiencyTransformerLarge Language ModelTextMultimodality
🎯 What it does: In large-scale language model training, to ensure reproducibility, a deterministic mode for the backward pass of FlashAttention-3 is implemented, and the DASH framework along with two scheduling strategies is proposed.
Data Provenance for Image Auto-Regressive Generation
Bihe Zhao (CISPA Helmholtz Center for Information Security), Adam Dziedzic (CISPA Helmholtz Center for Information Security)
GenerationData SynthesisDiffusion modelAuto EncoderImage
🎯 What it does: Propose a non-intrusive, post-processing source tracing framework for image autoregressive models (IAR) that can identify the generating model on published, watermark-free images.
Data Selection for LLM Alignment Using Fine-Grained Preferences
Jia Zhang (National Key Laboratory for Novel Software Technology, Nanjing University), Yu-Feng Li (National Key Laboratory for Novel Software Technology, Nanjing University)
Data-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningContrastive LearningText
🎯 What it does: Propose a data selection method based on the fine-grained preference divergence measure (Preference Divergence, PD), achieving efficient alignment by selecting negative PD samples;
Data-Aware and Scalable Sensitivity Analysis for Decision Tree Ensembles
Namrita Varshney (Indian Institute of Technology Bombay), S. Akshay (Indian Institute of Technology Bombay)
ClassificationOptimizationExplainability and InterpretabilityImageTabularBenchmark
🎯 What it does: For decision tree ensemble models, this paper proposes a data-aware sensitivity analysis method to find feature subset examples near the training distribution that can alter model predictions.
Data-Centric Lessons To Improve Speech-Language Pretraining
Vishaal Udandarao (Apple), Chung-Cheng Chiu (Apple)
Data SynthesisData-Centric LearningTransformerLarge Language ModelTextMultimodalityBenchmarkAudio
🎯 What it does: This paper improves the performance of speech question answering (SQA) through systematic data governance of the pre-training data of SpeechLM;
Data-to-Energy Stochastic Dynamics
Kirill Tamogashev (University of Edinburgh), Nikolay Malkin (University of Edinburgh)
Image TranslationGenerationData SynthesisReinforcement LearningImageStochastic Differential Equation
🎯 What it does: Proposed a generic data-to-energy IPF algorithm for solving optimal stochastic bridges when one or two distributions are unnormalized energy functions.
Dataless Weight Disentanglement in Task Arithmetic via Kronecker-Factored Approximate Curvature
Angelo Porrello (University of Modena and Reggio Emilia), Simone Calderara (University of Modena and Reggio Emilia)
ClassificationOptimizationRepresentation LearningTransformerImageTextBenchmark
🎯 What it does: Propose a data-free task arithmetic regularization method (TAK), which transforms representation drift regularization into curvature matrix approximation, leveraging Kronecker-factored approximate curvature (KFAC) to achieve weight decoupling in model linearization and nonlinear fine-tuning.
DataMIL: Selecting Data for Robot Imitation Learning with Datamodels
Shivin Dass (University of Texas Austin), Roberto Martín-Martín (University of Texas Austin)
Robotic Intelligence
🎯 What it does: Develop the DataMIL framework, utilizing datamodels methods to achieve data selection without rollout in robot imitation learning, and efficiently select data from large-scale heterogeneous datasets through proxy loss and clustering techniques.
Dataset Color Quantization: A Training-Oriented Framework for Dataset-Level Compression
YU CHENYUE, Yang He (Agency for Science, Technology and Research)
CompressionData-Centric LearningImage
🎯 What it does: This paper proposes a dataset-level color quantization framework (DCQ), which significantly reduces storage requirements by compressing image datasets through shared palettes in color space while maintaining training effectiveness.
Dataset Distillation as Pushforward Optimal Quantization
Hong Ye Tan (University of California, Los Angeles), Emma Slade (Tangram Therapeutics)
Data SynthesisKnowledge DistillationDiffusion modelImage
🎯 What it does: Propose a discretized dataset distillation method called DDOQ based on optimal quantization, which utilizes the latent space of diffusion models for clustering and generating synthetic training samples.
Dataset Distillation for Memorized Data: Soft Labels can Leak Held-Out Teacher Knowledge
Freya Behrens (École polytechnique fédérale de Lausanne), Lenka Zdeborová (École polytechnique fédérale de Lausanne)
Knowledge DistillationTransformer
🎯 What it does: Investigated how data memorized by the teacher model leaks to the student model when using soft labels in model distillation, i.e., whether the student can achieve non-random accuracy on unseen memorized samples.
DAVE: A VLM Vision Encoder for Document Understanding and Web Agents
Brandon Huang (MIT-IBM Watson AI Lab), Roei Herzig (MIT-IBM Watson AI Lab)
RecognitionSegmentationTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelVision-Language-Action ModelAuto EncoderImageMultimodalityBenchmark
🎯 What it does: DAVE is a visual encoder specifically designed for document and web understanding, employing a two-stage self-supervised and supervised autoregressive pre-training.
DaVinci: Reinforcing Visual-Structural Syntax in MLLMs for Generalized Scientific Diagram Parsing
Xingchen ZENG, Wei Zeng (Central South University)
Image TranslationLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelImageMultimodality
🎯 What it does: This paper proposes DaVinci, a multimodal large language model based on a two-stage framework (supervised learning visual primitives + reinforcement learning structural relationships), for parsing scientific chart images into editable TikZ code.
DCFold: Efficient Protein Structure Generation with Single Forward Pass
Zhe Zhang (Tsinghua University), Wei-Ying Ma (Tsinghua University)
Protein Structure PredictionTransformerDiffusion modelBiomedical Data
🎯 What it does: Compress AlphaFold3 into a single-step high-precision protein folding model DCFold via the Dual Consistency framework and Temporal Geodesic Matching (TGM) scheduling, achieving a 15× speedup in inference time.
DeAltHDR: Learning HDR Video Reconstruction from Degraded Alternating Exposure Sequences
Shuohao Zhang (Harbin Institute of Technology), Wangmeng Zuo (Harbin Institute of Technology)
RestorationConvolutional Neural NetworkOptical FlowVideoBenchmark
🎯 What it does: A new HDR video reconstruction framework named DeAltHDR is proposed to recover high-quality HDR videos from damaged LDR frames, specifically targeting noise reduction and deblurring in alternating exposure sequences.
DEAS: DEtached value learning with Action Sequence for Scalable Offline RL
Changyeon Kim (KAIST), Yuke Zhu (University of Texas at Austin)
Robotic IntelligenceReinforcement LearningBenchmark
🎯 What it does: Propose a novel offline reinforcement learning framework named DEAS that performs value learning using fixed-length action sequences;
Death of the Novel(ty): Beyond N-Gram Novelty as a Metric for Textual Creativity
Arkadiy Saakyan (Columbia University), Tuhin Chakrabarty (Stony Brook University)
GenerationLarge Language ModelSupervised Fine-TuningText
🎯 What it does: By comparing text generated by humans and large language models (LLMs), this study constructs 8,618 annotated expressions based on fine-grained 'close reading' annotations provided by professional writers, evaluating the 'perceived novelty,' 'comprehensibility,' and 'contextual adaptability' of individual sentences.
Debiased and Denoised Representation Learning for Incomplete Multi-view Clustering
Qianqian Wang (Xidian University), Quanxue Gao (Xidian University)
Representation LearningAuto EncoderContrastive LearningImageBenchmark
🎯 What it does: Propose a DDR-IMVC framework based on unbiased and denoised representation learning to address distribution shift and noise caused by missing views.
Debiased Front-Door Learners for Heterogeneous Effects
Yonghan Jung (University of Illinois Urbana-Champaign)
Tabular
🎯 What it does: Propose two estimators for heterogeneous treatment effect (HTE) under the front-door structure: FD-DR-Learner and FD-R-Learner, which can estimate individualized causal effects in the presence of unmeasured confounding and observable mediator variables.
Debugging Concept Bottleneck Models through Removal and Retraining
Eric Enouen (Cornell University), sainyam galhotra
Explainability and InterpretabilityLarge Language ModelImage
🎯 What it does: Proposes an interpretable debugging framework divided into two steps: removing harmful concepts and retraining, to eliminate bias and errors in CBM models;
DecAlign: Hierarchical Cross-Modal Alignment for Decoupled Multimodal Representation Learning
Chengxuan Qian (Texas A&M University), Zhengzhong Tu (Texas A&M University)
ClassificationRepresentation LearningTransformerMultimodality
🎯 What it does: Propose a hierarchical cross-modal alignment framework named DecAlign, which first disentangles multimodal features into modality-specific (heterogeneous) and modality-shared (homogeneous) categories, then aligns them separately;
Decentralized Attention Fails Centralized Signals: Rethinking Transformers for Medical Time Series
Guoqi Yu (Hong Kong Polytechnic University), Shujun Wang (Hong Kong Polytechnic University)
Computational EfficiencyRepresentation LearningTransformerTime SeriesBiomedical DataElectrocardiogram
🎯 What it does: Proposed a centralized core token aggregation-redistribution (CoTAR) module to replace the decentralized attention in Transformers, aiming to better capture channel dependencies in medical time series signals.
Decentralized Nonconvex Optimization under Heavy-Tailed Noise: Normalization and Optimal Convergence
Shuhua Yu (Carnegie Mellon University), Soummya Kar (Carnegie Mellon University)
OptimizationFederated LearningText
🎯 What it does: Proposed a decentralized non-convex optimization algorithm GT-NSGDm, specifically designed for gradient fields with heavy-tailed noise.
Decision Aggregation under Quantal Response
Zhihuan Huang (Peking University), Yuqing Kong (Peking University)
OptimizationLarge Language ModelText
🎯 What it does: This paper studies the decision aggregation problem under bounded rationality (quantized response) experts, proving that when experts' rationality levels fall below a threshold, simple majority voting is the optimal robust aggregator, revealing the phenomenon that bounded rationality in groups can outperform fully rational experts, and empirically validating this on large language models.
Decision-Theoretic Approaches for Improved Learning-Augmented Algorithms
Spyros Angelopoulos (CNRS), Georgii Melidi (Sorbonne University)
Optimization
🎯 What it does: This paper proposes a decision theory framework to evaluate and optimize the performance of online algorithms that utilize machine learning predictions, and applies it to three classical problems: ski rental, one-max search, and contract scheduling.
DeCo-DETR: Decoupled Cognition DETR for efficient Open-Vocabulary Object Detection
Siheng Wang (Jiangsu University), Qiang Sun (MBZUAI)
Object DetectionComputational EfficiencyKnowledge DistillationTransformerVision Language ModelImageBenchmark
🎯 What it does: Proposes a novel open-vocabulary object detection framework called DeCo-DETR, which achieves efficient detection without relying on text encoders.
Decoding Dynamic Visual Experience from Calcium Imaging via Cell-Pattern-Aware Pretraining
Sangyoon Bae (Seoul National University), Jiook Cha (McGill University)
RestorationRepresentation LearningData-Centric LearningConvolutional Neural NetworkTransformerSupervised Fine-TuningAuto EncoderTime SeriesBiomedical Data
🎯 What it does: This paper proposes the POYO-CAP framework, which first uses statistical predictability (skewness, kurtosis) to screen predictable neurons, performs masking reconstruction + lightweight auxiliary supervised self-supervised pre-training on them, and then fine-tunes the remaining more random neurons to achieve end-to-end decoding from neural activity to visual frames;
Decoding Open-Ended Information Seeking Goals from Eye Movements in Reading
Cfir Avraham Hadar (Technion Israel Institute of Technology), Yevgeni Berzak (Technion Israel Institute of Technology)
RetrievalExplainability and InterpretabilityRepresentation LearningData-Centric LearningRecurrent Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningTextMultimodality
🎯 What it does: The study automatically infers the specific information retrieval goals readers are pursuing while reading paragraphs by analyzing their eye movement trajectories, and proposes two decoding tasks: goal selection and goal reconstruction.
DecompGAIL: Learning Realistic Traffic Behaviors with Decomposed Multi-Agent Generative Adversarial Imitation Learning
Ke Guo (Nanyang Technological University), Chen Lv (Nanyang Technological University)
Autonomous DrivingTransformerReinforcement LearningGenerative Adversarial NetworkSequential
🎯 What it does: Proposed DecompGAIL, a multi-agent generative adversarial imitation learning framework that decomposes realism into two components: ego-map and ego-neighbor, for generating more realistic traffic simulations;
Decomposed Attention Fusion in MLLMs for Training-free Video Reasoning Segmentation
Su Ho Han (Yonsei University), Seon Joo Kim (Inha University)
SegmentationTransformerPrompt EngineeringVision Language ModelVideoMultimodality
🎯 What it does: This paper proposes a training-free multi-modal large language model (MLLM) video reasoning segmentation framework named DecAF, which achieves fine-grained mask generation through attention convolution fusion and SAM2-guided refinement;
Decomposing LLM Computation with Jets
Yihong Chen (University of Oxford), Luca Franceschi
Explainability and InterpretabilityLarge Language ModelText
🎯 What it does: This paper proposes the Jet Expansions framework, which utilizes jets (generalized Taylor expansions) to recursively decompose the residual networks of large language models, breaking down complex computations into interpretable input→output paths and remainders, thereby enhancing model interpretability and maintainability.
Decomposing Representation Space into Interpretable Subspaces with Unsupervised Learning
Xinting Huang, Michael Hahn (Saarland University)
Explainability and InterpretabilityRepresentation LearningTransformerText
🎯 What it does: Propose an unsupervised neighborhood distance minimization method to decompose the representation space of neural networks into interpretable multidimensional subspaces.
Decomposition of Concept-Level Rules in Visual Scenes
Fan Shi (Fudan University), Bin Li (Fudan University)
RecognitionExplainability and InterpretabilityRepresentation LearningVision Language ModelImageMultimodality
🎯 What it does: This paper proposes a general framework called Concept-Rule Decomposition (CRD), which leverages large vision-language models (LVLMs) to automatically extract visual concepts from images and learn their spatial rules, subsequently obtaining the most explanatory concept set and rules through sampling;
Deconstructing Guidance: A Semantic Hierarchy for Precise Diffusion Model Editing
Wootaek Jeong (Korea University), Heung-Il Suk (Korea University)
GenerationDiffusion modelImage
🎯 What it does: Propose the 'semantic scale hypothesis' and design a no-training, plug-and-play Prism-Edit module based on this hypothesis. By hierarchically amplifying the weak semantic signals in the guidance difference vector of diffusion models, precise and controllable editing of the background and subject is achieved.
Deconstructing Positional Information: From Attention Logits to Training Biases
Zihan Gu (Chinese Academy of Sciences), Yue Hu (Chinese Academy of Sciences)
Explainability and InterpretabilityRepresentation LearningTransformerText
🎯 What it does: Investigated the mechanism of position encoding (PE) in Transformers, constructing a unified analytical framework based on Toeplitz matrices, distinguishing additive and multiplicative PE, and revealing RoPE's single-head deposition pattern through synthetic tasks and head ablation experiments.
Decoupled DMD: CFG Augmentation as the Spear, Distribution Matching as the Shield
Dongyang Liu (Alibaba Group), Hongsheng Li (Chinese University of Hong Kong)
GenerationTransformerDiffusion modelGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: Propose decomposing the DMD training objective into two parts: CFG Enhancement (CA) as the core engine for few-step generation, and Distribution Matching (DM) as a regularization shield for stable training, and design a decoupled denoising schedule to improve performance.
Decoupled MeanFlow: Turning Flow Models into Flow Maps for Accelerated Sampling
Kyungmin Lee (Korea Advanced Institute Of Science And Technology), Jinwoo Shin (Korea Advanced Institute Of Science And Technology)
GenerationFlow-based ModelImage
🎯 What it does: Convert pre-trained flow models to flow graph models without modification, enabling high-quality image generation in just 1-4 steps
Decoupled Q-Chunking
Qiyang Li (University of California Berkeley), Sergey Levine (University of California Berkeley)
OptimizationKnowledge DistillationReinforcement LearningBenchmark
🎯 What it does: Proposed the Decoupled Q-Chunking (DQC) algorithm, which decouples the block size of the value function from the block size of the policy, using distilled sub-value functions from partial action blocks to train the policy
Decoupling Dynamical Richness from Representation Learning: Towards Practical Measurement
Yoonsoo Nam (Oxford University), Ard A. Louis (Oxford University)
Representation LearningConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a dynamic richness metric D_LR that is independent of model performance, along with a visualization tool, for analyzing feature learning and expressive capacity during training.
Decoupling Positional and Symbolic Attention in Transformers
Felipe Urrutia (CENIA), Cristobal Rojas (IMC UC)
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: This paper theoretically defines the 'positionality' and 'symbolicity' behaviors of attention heads under RoPE position encoding in Transformers, proves their mutual exclusivity, and proposes quantitative metrics. Subsequently, it evaluates these behaviors on LLMs such as Gemma-2, Qwen-2, and LLaMA-3, designs benchmark tasks to verify the incompatibility between the two behaviors, and demonstrates their causal impact on model performance through frequency control.
Decoupling Primitive with Experts: Dynamic Feature Alignment for Compositional Zero-Shot Learning
Xiao Zhang (Xi'an Jiaotong University), Nanning Zheng (Xi'an Jiaotong University)
ClassificationRecognitionMixture of ExpertsVision Language ModelContrastive LearningMultimodality
🎯 What it does: Propose the EVA framework, combining Mixture-of-Experts domain experts with semantic variant alignment to achieve zero-shot reasoning for composite state-object combinations.
Decoupling the Class Label and the Target Concept in Machine Unlearning
Jianing Zhu (Hong Kong Baptist University), Masashi Sugiyama (RIKEN Center for Advanced Intelligence Project)
OptimizationRepresentation LearningConvolutional Neural NetworkImage
🎯 What it does: Proposes a theoretical framework for decoupling class labels from target concepts in machine learning, and defines three new class-level forgetting tasks (target mismatch, model mismatch, data mismatch) based on this framework.
Deep FlexQP: Accelerated Nonlinear Programming via Deep Unfolding
Alex Oshin (Georgia Institute of Technology), Evangelos Theodorou
OptimizationRecurrent Neural NetworkFinance Related
🎯 What it does: Propose FlexQP and its deep unfolding version Deep FlexQP, which can handle infeasibility in QP subproblems and accelerate solving through learning-based parameterization.
Deep Global-sense Hard-negative Discriminative Generation Hashing for Cross-modal Retrieval
Kun Cheng (Qufu Normal University), Jie Nie (Ocean University of China)
RetrievalComputational EfficiencyGraph Neural NetworkTransformerContrastive LearningMultimodality
🎯 What it does: Propose a global-aware hard negative sample generation framework DGHDGH to enhance the discriminative ability of cross-modal hashing retrieval;
Deep Hierarchical Learning with Nested Subspace Networks for Large Language Models
Paulius Rauba (University of Cambridge), Mihaela van der Schaar (University of Cambridge)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose Nested Subspace Networks (NSNs), which use adjustable-rank low-rank decomposition within linear layers, allowing a single model to dynamically switch between different ranks during inference based on computational budget, thus achieving a continuous performance-computation trade-off.
Deep Ignorance: Filtering Pretraining Data Builds Tamper-Resistant Safeguards into Open-Weight LLMs
Kyle O'Brien (EleutherAI), Stella Biderman (EleutherAI)
Adversarial AttackData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningTextBiomedical Data
🎯 What it does: This paper constructs a series of 6.9B parameter open-weight LLMs by adopting multi-stage text filtering during the pre-training phase of open-source large language models to remove knowledge related to biological threat agents, and evaluates their security and tamperability.
Deep Latent Variable Model based Vertical Federated Learning with Flexible Alignment and Labeling Scenarios
Kihun Hong (Korea Advanced Institute of Science and Technology), Ganguk Hwang (Korea Advanced Institute of Science and Technology)
Federated LearningImagePoint CloudBenchmarkAudio
🎯 What it does: Proposed the FALSE-VFL framework, which supports utilizing unaligned and unlabeled data in multi-party VFL while addressing MCAR, MAR, and MNAR missing data mechanisms.
Deep Learning for Subspace Regression
Vladimir Fanaskov, Ivan Oseledets
OptimizationRepresentation LearningPhysics Related
🎯 What it does: Proposed a 'subspace regression' framework that learns mappings from parameters to linear subspaces, enabling reduced-order modeling for high-dimensional parametric problems, feature subspace approximation, iterative method acceleration, and optimal control.
Deep Learning with Learnable Product-Structured Activations
Saanjali Maharaj (University of Toronto), Prasanth B. Nair (University of Toronto)
RestorationSuper ResolutionImageBiomedical DataComputed TomographyPhysics RelatedAudio
🎯 What it does: Propose a deep low-rank separable neural network (LRNN), replacing traditional fixed activation functions with a learnable product-structured activation function;
Deep SPI: Safe Policy Improvement via World Models
Florent Delgrange (Vrije Universiteit Brussel), Willem Röpke (Vrije Universiteit Brussel)
Convolutional Neural NetworkReinforcement LearningWorld ModelImageBenchmark
🎯 What it does: In online deep reinforcement learning, combining world models with representation learning, the Safe Policy Improvement (DeepSPI) algorithm is proposed to ensure safe policy updates, convergence, and improved performance in high-dimensional environments.
Deep Think with Confidence
Yichao Fu (University Of California San Diego), Jiawei Zhao (Meta Ai)
Computational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose the DeepConf method, which utilizes the internal confidence of LLMs to dynamically filter low-quality reasoning trajectories, thereby enhancing reasoning efficiency and accuracy.
Deep-ICE: The first globally optimal algorithm for empirical risk minimization of two-layer maxout and ReLU networks
Xi He (Peking University), Max A Little
ClassificationOptimizationTabularBenchmark
🎯 What it does: Developed and implemented the first globally optimal algorithm, Deep-ICE, for minimizing 0-1 loss under two-layer ReLU/Maxout networks, with a scalable implementation provided.
DeepAFL: Deep Analytic Federated Learning
Jianheng Tang (Peking University), Yunhuai Liu (Peking University)
Federated LearningConvolutional Neural NetworkImage
🎯 What it does: Proposes a completely gradient-agnostic deep analytical federated learning framework, DeepAFL, which utilizes residual blocks to achieve multi-layer representation learning.
DeepCompress: A Dual Reward Strategy for Dynamically Exploring and Compressing Reasoning Chains
Tian Liang (Tencent AI Lab), Dong Yu (Tencent AI Lab)
OptimizationComputational EfficiencyReinforcement LearningTextBenchmarkChain-of-Thought
🎯 What it does: Developed the DeepCompress framework, dynamically adjusting the Chain-of-Thought length of large-scale inference models, leveraging dual length rewards and model-aware difficulty to achieve higher accuracy and more efficient inference.
DeepEyes: Incentivizing "Thinking with Images" via Reinforcement Learning
Ziwei Zheng (Xiaohongshu Inc), XingYu
Reinforcement LearningVision Language ModelMultimodalityChain-of-Thought
🎯 What it does: Train a visual-language model called DeepEyes that can proactively 'look at images' and make cropping decisions during reasoning, using end-to-end reinforcement learning without requiring pre-collected reasoning data.
DeepEyesV2: Toward Agentic Multimodal Model
Jack Hong, XingYu (Xiaohongshu Inc)
AI Code AssistantTransformerSupervised Fine-TuningReinforcement LearningAgentic AIVision Language ModelMultimodalityRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes DeepEyesV2, an agent-based multi-modal large language model capable of actively invoking external tools (such as code execution environments and web search) and integrating them into the multi-modal reasoning process.
DeepFRC: An End-to-End Deep Learning Model for Functional Registration and Classification
Siyuan Jiang (ShanghaiTech University), Pengcheng Zeng (ShanghaiTech University)
ClassificationConvolutional Neural NetworkContrastive LearningTime Series
🎯 What it does: Proposed an end-to-end deep learning framework called DeepFRC for simultaneously accomplishing functional registration (alignment) and classification;
DeepMath-103K: A Large-Scale, Challenging, Decontaminated, and Verifiable Mathematical Dataset for Advancing Reasoning
Zhiwei He (Tencent), Dong Yu (Tencent)
Data-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Proposed and released the DeepMath-103K dataset, and trained a series of LLMs (DeepMath series models) on this dataset, which can be further enhanced through RL and SFT, achieving state-of-the-art performance on mathematical and cross-domain reasoning tasks.
DeepPrim: a Physics-Driven 3D Short-term Weather Forecaster via Primitive Equation Learning
Jiawei Chen (Zhejiang University), Liang Sun (DAMO Academy, Alibaba Group)
TransformerTime SeriesPhysics RelatedOrdinary Differential Equation
🎯 What it does: Propose DeepPrim, a deep physics-driven model for short-term weather forecasting by learning the primitive equations of the Earth's atmosphere.
DeepRAG: Thinking to Retrieve Step by Step for Large Language Models
Xinyan Guan (Chinese Academy of Sciences), Jie Zhou (Tencent Inc)
RetrievalTransformerReinforcement LearningTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes the DeepRAG framework, modeling retrieval-augmented reasoning as a Markov decision process, enabling demand-based progressive retrieval and reasoning;
DeepResearch Bench: A Comprehensive Benchmark for Deep Research Agents
Mingxuan Du (University of Science and Technology of China), Zhendong Mao (University of Science and Technology of China)
RetrievalLarge Language ModelTextBenchmark
🎯 What it does: Proposed DeepResearch Bench, a benchmark containing 100 PhD-level research tasks across 22 disciplines, to evaluate the report generation and information retrieval capabilities of Deep Research Agents (DRAs).
DeepSADR: Deep Transfer Learning with Subsequence Interaction and Adaptive Readout for Cancer Drug Response Prediction
Yuanpeng Zhang (Central South University), Lei Deng (Institute for Infocomm Research, A*STAR)
Domain AdaptationDrug DiscoveryGraph Neural NetworkTransformerSupervised Fine-TuningAuto EncoderGraphTabularBiomedical Data
🎯 What it does: Propose the DeepSADR model, which constructs an interaction graph using drug substructures and gene function subsequences, and implements deep transfer learning to predict drug efficacy in cancer patients.
DeepScientist: Advancing Frontier-Pushing Scientific Findings Progressively
Yixuan Weng (Westlake University), Yue Zhang (Westlake University)
OptimizationTransformerLarge Language ModelTextBenchmark
🎯 What it does: Built a LLM multi-agent system called DeepScientist, capable of autonomously completing the full scientific discovery process from conception to experimental validation on a monthly time scale, achieving results exceeding human state-of-the-art (SOTA) in three cutting-edge AI tasks.
DeepSearch: Overcome the Bottleneck of Reinforcement Learning with Verifiable Rewards via Tree-based Search
Fang Wu (Stanford University), Yejin Choi (Stanford University)
Reinforcement LearningText
🎯 What it does: Propose the DeepSearch framework, which directly embeds Monte Carlo Tree Search (MCTS) into the training loop of reinforcement learning with verifiable rewards (RLVR) to systematically enhance the model's reasoning capabilities through search.
DeepTRACE: Auditing Deep Research AI Systems for Tracking Reliability Across Citations and Evidence
Pranav Narayanan Venkit (Salesforce AI Research), Chien-Sheng Wu (Salesforce AI Research)
Explainability and InterpretabilityLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Evaluate and audit the end-to-end trustworthiness of public generative search engines and deep research agents, proposing an eight-dimensional evaluation framework named DeepTRACE.
DeepWeightFlow: Re-Basined Flow Matching for Generating Neural Network Weights
Saumya Gupta (Northeastern University), Ayan Paul (Northeastern University)
GenerationData SynthesisConvolutional Neural NetworkTransformerFlow-based ModelImageTextTabularBenchmark
🎯 What it does: Propose and implement a deep generative model called DeepWeightFlow based on Flow Matching, which directly generates complete and high-performance neural network weights in the weight space. It supports multiple architectures (MLP, ResNet, ViT, BERT) and can be scaled up to 100M parameters through Canonicalization (Git Re-Basin, TransFusion) and PCA.
Defending against Backdoor Attacks via Module Switching
Weijun Li (Macquarie University), Qiongkai Xu (Macquarie University)
Safty and PrivacyAdversarial AttackData-Centric LearningNeural Architecture SearchImageText
🎯 What it does: This paper proposes a module switching defense (MSD) mechanism to counter backdoor attacks on deep learning models in post-training environments.
DefensiveKV: Taming the Fragility of KV Cache Eviction in LLM Inference
Yuan Feng (University of Science and Technology of China), Xike Xie (University of Science and Technology of China)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelAgentic AITextBenchmark
🎯 What it does: Proposed and implemented defensive aggregation for KV cache eviction, and built DefensiveKV and its hierarchical version Layer-DefensiveKV to significantly compress the cache while maintaining generation quality;
Deforming Videos to Masks: Flow Matching for Referring Video Segmentation
Zanyi Wang (SGIT AI Lab, State Grid Corporation of China), Jingdong Wang (Baidu)
SegmentationTransformerDiffusion modelFlow-based ModelAuto EncoderVideoTextOrdinary Differential Equation
🎯 What it does: Redefine reference video segmentation (RVOS) as a continuous flow problem conditioned on text, and propose the end-to-end FlowRVS framework that directly transforms the video's implicit representation into masks.
Deft Scheduling of Dynamic Cloud Workflows with Varying Deadlines via Mixture-of-Experts
Ya Shen (Victoria University of Wellington), Mengjie Zhang (Victoria University of Wellington)
OptimizationGraph Neural NetworkReinforcement LearningMixture of ExpertsGraph
🎯 What it does: Propose a DEFT model for dynamic cloud workflow scheduling, integrating Mixture-of-Experts with graph adaptive gating.
Delay Flow Matching
Bolin Zhao (Fudan University), Qunxi Zhu (Fudan University)
GenerationFlow-based ModelImageBiomedical DataOrdinary Differential Equation
🎯 What it does: This paper proposes a continuous-time generative framework based on Neural Delay Differential Equations (Neural DDE) called Delay Flow Matching (DFM), achieving precise modeling of distribution transitions and handling trajectory crossings and distribution heterogeneity.
DeLeaker: Dynamic Inference-Time Reweighting For Semantic Leakage Mitigation in Text-to-Image Models
Mor Ventura (Technion), Roi Reichart (Technion)
GenerationSafty and PrivacyTransformerVision Language ModelDiffusion modelImageText
🎯 What it does: Designed a lightweight, no-training inference-time attention reweighting method called DeLeaker to eliminate semantic leakage in text-to-image generation models;
DeLiVR: Differential Spatiotemporal Lie Bias for Efficient Video Deraining
Shuning Sun (University of Chinese Academy of Sciences), Zhuoran Zheng (Qilu University of Technology)
RestorationTransformerVideo
🎯 What it does: Proposes DeLiVR, a Transformer-based video de-raining method utilizing Lie group differential bias.
Delta-XAI: A Unified Framework for Explaining Prediction Changes in Online Time Series Monitoring
Changhun Kim (AITRICS), Eunho Yang (AITRICS)
Anomaly DetectionExplainability and InterpretabilityConvolutional Neural NetworkRecurrent Neural NetworkTransformerTime SeriesBiomedical DataElectronic Health Records
🎯 What it does: This paper proposes a unified framework Delta-XAI to explain the reasons behind prediction changes in online time series monitoring, and designs new evaluation metrics for this task;
Delving into Spectral Clustering with Vision-Language Representations
Bo Peng (University of Technology Sydney), Zhen Fang (University of Technology Sydney)
Representation LearningTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Proposes a multi-modal spectral clustering method based on Neural Tangent Kernel (NTK), constructing a more discriminative similarity matrix by multiplying visual similarity with semantic overlap, anchored by forward text semantics.
DeMo: Decoupled Momentum Optimization
Bowen Peng (Nous Research), qiang liu
OptimizationComputational EfficiencyText
🎯 What it does: Proposes Decoupled Momentum Optimization (DeMo), a framework that significantly reduces communication volume in distributed training while maintaining the same convergence performance as AdamW.
DemoGrasp: Universal Dexterous Grasping from a Single Demonstration
Haoqi Yuan (Peking University), Zongqing Lu (Peking University)
Robotic IntelligenceTransformerReinforcement LearningFlow-based ModelImagePoint Cloud
🎯 What it does: Learn a general grasping strategy applicable to various multi-fingered robot arms using a single successful demonstration combined with reinforcement learning, achieving zero-shot transfer from simulation to real robots.
Demystifying and Enhancing the Efficiency of Large Language Model Based Search Agents
Tiannuo Yang (University Of Southern California), Willie Neiswanger (Nankai University)
RetrievalOptimizationComputational EfficiencyLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes an efficient inference framework called SearchAgent-X specifically designed for large language model (LLM) search agents, aiming to optimize the interaction process between alternating reasoning and retrieval.
Demystifying Deep Search: A Holistic Evaluation with Hint-free Multi-Hop Questions and Factorised Metrics
Maojia Song (Singapore University of Technology and Design), Jingren Zhou (Alibaba Group)
RetrievalLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes a benchmark called WebDetective to evaluate autonomous reasoning capabilities in prompt-free multi-hop deep search tasks;
Demystifying Robot Diffusion Policies: Action Memorization and a Simple Lookup Table Alternative
Chengyang He (Stanford University), Mac Schwager (Stanford University)
Computational EfficiencyRobotic IntelligenceDiffusion modelContrastive LearningBenchmark
🎯 What it does: Empirical analysis of Diffusion Policy under low-data conditions demonstrates that it achieves efficient control primarily through retrieval/memory of trained actions rather than generalization. Based on this, a simplified Action Lookup Table (ALT) model is proposed, which maintains comparable performance while significantly improving inference speed and providing OOD detection.
Demystifying Supervision Data Generalization in Multimodal LMs
Xuan Qi (University of Pennsylvania), Xingyu Fu (University of Pennsylvania)
Data-Centric LearningTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelMultimodality
🎯 What it does: The study investigates whether the impact of a supervised dataset on a target benchmark can be predicted before training in multimodal large language models and proposes the DATAPROPHET metric.
Demystifying The Mechanisms Behind Emergent Exploration in Goal-Conditioned RL
Mahsa Bastankhah (Princeton University), Benjamin Eysenbach (Princeton University)
Explainability and InterpretabilityRepresentation LearningReinforcement LearningContrastive Learning
🎯 What it does: Theoretically analyze and experimentally investigate single-goal contrastive reinforcement learning (SGCRL), revealing its exploration mechanism: the agent explores implicitly based on rewards derived from representation similarity and switches to exploitation after achieving the goal.
Denoising Neural Reranker for Recommender Systems
Wenyu Mao (University of Science and Technology of China), Xiang Wang (University of Science and Technology of China)
Recommendation SystemGenerative Adversarial Network
🎯 What it does: This paper proposes a denoising neural re-ranker (DNR) for multi-stage recommendation systems, which enhances re-ranking performance by adding noise to and denoising the scores generated by the retriever.
DeNOTS: Stable Deep Neural ODEs for Time Series
Ilya Kuleshov (Applied AI Institute), Alexey Zaytsev (Applied AI Institute)
Recurrent Neural NetworkTime SeriesBenchmarkOrdinary Differential Equation
🎯 What it does: Propose a new continuous-time sequential model called DeNOTS, which increases the network's 'depth' by expanding the integration time window and incorporates anti-phase negative feedback (Anti-NF) to maintain trajectory stability and avoid forgetting.
Dens3R: A Foundation Model for 3D Geometry Prediction
Xianze Fang (Alibaba Group), chengfei lv
Depth EstimationTransformerContrastive LearningImagePoint Cloud
🎯 What it does: Built a unified 3D foundation model called Dens3R, which can regress depth, surface normals, point maps, and image pairing information from single/multi-view images without pose information in one go;
DenseGRPO: From Sparse to Dense Reward for Flow Matching Model Alignment
Haoyou Deng (Huazhong University of Science and Technology), Nong Sang (Huazhong University of Science and Technology)
GenerationReinforcement Learning from Human FeedbackReinforcement LearningScore-based ModelFlow-based ModelImageTextMultimodalityStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Propose DenseGRPO, which utilizes ODE inference to estimate dense rewards at each step and adaptively adjusts the noise intensity in SDE sampling, thereby achieving more precise reinforcement learning alignment in text-to-image flow matching models.
Densemarks: Learning Canonical Embeddings for Human Heads Images via Point Tracks
Dmitrii Pozdeev (Technical University of Munich), Artem Sevastopolsky (Technical University of Munich)
Pose EstimationRepresentation LearningTransformerContrastive LearningImageVideo
🎯 What it does: Learn a low-dimensional (3D) embedding representation (DenseMarks) for facial head images, achieving pixel-level correspondence through point tracking;
Deploying Models to Non-participating Clients in Federated Learning without Fine-tuning: A Hypernetwork-based Approach
Yuhao Zhou (Sichuan University), Jiancheng Lv (Sichuan University)
ClassificationFederated LearningImage
🎯 What it does: This paper proposes the HyperFedZero method, achieving seamless migration of the global model to non-participating clients without fine-tuning in federated learning, enabling zero-shot personalization.
Depth Anything 3: Recovering the Visual Space from Any Views
Haotong Lin (ByteDance Seed), Bingyi Kang (ByteDance Seed)
Depth EstimationKnowledge DistillationTransformerImagePoint CloudBenchmark
🎯 What it does: Propose Depth Anything 3, which can predict dense depth maps and ray maps using a single ordinary Transformer (e.g., DINOv2) under any number of views and whether the camera pose is known, thereby recovering a consistent 3D visual space.
Depth Anything with Any Prior
Zehan Wang (Zhejiang University), Zhou Zhao (Zhejiang University)
Depth EstimationConvolutional Neural NetworkPoint Cloud
🎯 What it does: Propose Prior Depth Anything, a unified framework that integrates sparse/low-resolution/missing metric depth priors with relative depth prediction to generate fine and complete metric depth maps.
DepthLM: Metric Depth from Vision Language Models
Zhipeng Cai (Meta), Yangyang Shi (Meta)
Depth EstimationAutonomous DrivingTransformerSupervised Fine-TuningVision Language ModelImageBenchmark
🎯 What it does: Propose the DepthLM framework, converting VLM into a powerful pixel-level metric depth estimator.
DeRaDiff: Denoising Time Realignment of Diffusion Models
Ratnavibusena Don Shahain Manujith (National University of Singapore), Yang Zhang (National University of Singapore)
GenerationDiffusion modelImage
🎯 What it does: Propose the DeRaDiff method, which enables real-time control of alignment strength and generates models with approximate different regularization strengths without retraining, by adjusting the KL regularization intensity of the pre-aligned diffusion model during inference.
Derandomized Online-to-Non-convex Conversion for Stochastic Weakly Convex Optimization
Fanfan Ji (Nanjing University), Xiaotong Yuan
OptimizationConvolutional Neural NetworkTransformerImage
🎯 What it does: Proposes a non-randomized online-to-non-convex optimization conversion framework (D-O2NC) for weakly convex non-smooth non-convex problems, proving that it achieves the optimal dimension-free upper bound under weak convexity settings; simultaneously designs a periodic restart variant to further accelerate training and improve convergence.
DES-LOC: Desynced Low Communication Adaptive Optimizers for Foundation Models
Alex Iacob (University of Cambridge), Nicholas D. Lane (University of Cambridge)
OptimizationComputational EfficiencyText
🎯 What it does: Proposes DES-LOC, which decouples the synchronization periods of parameters and momentum in distributed foundational model training, significantly reducing communication overhead while ensuring convergence.
DESIGNER: Design-Logic-Guided Multidisciplinary Data Synthesis for LLM Reasoning
Weize Liu (Alibaba Group), Bo Zheng (Alibaba Group)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes DESIGNER, a design logic guided multi-disciplinary data synthesis pipeline that generates millions of high-difficulty, multi-step reasoning exam questions using large-scale raw text.
Designing Affine-Invariant Neural Networks for Photometric Corruption Robustness and Generalization
Mounir Messaoudi (Inria), Charles Kervrann (Inria)
ClassificationObject DetectionConvolutional Neural NetworkImageBiomedical Data
🎯 What it does: Propose a new network architecture—Scale-Equivariant Shift-Invariant (SEqSI)—achieving global illumination invariance by adding a zero-sum convolutional layer before a scale-equivariant backbone;