NeurIPS 2025 Papers — Page 13
Conference on Neural Information Processing Systems · 5275 papers
Ditch the Denoiser: Emergence of Noise Robustness in Self-Supervised Learning from Data Curriculum
Wenquan Lu (Brown University), Randall Balestriero (Brown University)
RetrievalRepresentation LearningContrastive LearningImage
🎯 What it does: A completely self-supervised framework is proposed, capable of learning noise-robust visual representations without using a denoiser during the inference phase.
DitHub: A Modular Framework for Incremental Open-Vocabulary Object Detection
Chiara Cappellino (University of Modena and Reggio Emilia), Rita Cucchiara (University of Modena and Reggio Emilia)
Object DetectionSupervised Fine-TuningImage
🎯 What it does: Proposes the DitHub framework, which utilizes modular and incremental adapter management to achieve scalable and fine-tuned adaptation for open vocabulary object detection.
Diverse Influence Component Analysis: A Geometric Approach to Nonlinear Mixture Identifiability
Hoang-Son Nguyen (Oregon State University), Xiao Fu (Oregon State University)
OptimizationRepresentation LearningAuto EncoderTabularBiomedical Data
🎯 What it does: By utilizing the convex geometric properties of the Jacobian matrix in nonlinear mixture models, the J-VolMax criterion is proposed to implement the DICA method, thereby achieving identifiable learning of latent variables.
Diversifying Parallel Ergodic Search: A Signature Kernel Evolution Strategy
Sreevardhan Sirigiri (University of Sydney), Fabio Ramos (Nvidia)
OptimizationRobotic IntelligenceReinforcement LearningSequential
🎯 What it does: This paper proposes a parallel Stein Variational Euclidean Search method using signature kernels and evolutionary strategies to generate diverse and high-quality robot exploration trajectories.
Diversity as a Reward: Fine-Tuning LLMs on a Mixture of Domain-Undetermined Data
Zhenqing Ling (Sun Yat-sen University), Ying Shen (Alibaba Group)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes a self-supervised data selection framework DAAR based on diversity rewards, which can perform high-quality fine-tuning of large language models in the absence of domain labels.
Diversity Is All You Need for Contrastive Learning: Spectral Bounds on Gradient Magnitudes
Peter Ochieng (University of Cambridge)
OptimizationRepresentation LearningContrastive LearningImage
🎯 What it does: A non-asymptotic spectral framework is proposed to estimate the upper and lower bounds of the InfoNCE gradient norm, and based on this, intervention methods such as batch diversity sampling and batch whitening are designed.
Diversity-Aware Policy Optimization for Large Language Model Reasoning
Jian Yao (Hong Kong Polytechnic University), KC Tan
OptimizationTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper studies the role of diversity in the reinforcement learning (RL) training process for large language model (LLM) inference tasks and proposes a diversity-aware strategy optimization method that applies token-level diversity regularization on positive samples.
Diversity-oriented Deep Multi-modal Clustering
Yanzheng Wang (Zhengzhou University), Mingliang Xu (Zhengzhou University)
Representation LearningAuto EncoderContrastive LearningMultimodality
🎯 What it does: This paper proposes a multimodal clustering method, where the core idea is to first select the modality with the highest information quality as the dominant modality, and then extract differential information from other modalities through a two-level diversity learning process and concatenate it to the dominant modality, ultimately performing clustering on the enhanced dominant modality.
DKDR: Dynamic Knowledge Distillation for Reliability in Federated Learning
Yueyang Yuan (Wuhan University), Mang Ye (Wuhan University)
Federated LearningKnowledge DistillationImage
🎯 What it does: This paper proposes the DKDR framework to address the reliability issues of knowledge distillation caused by heterogeneous multi-domain data in federated learning.
dKV-Cache: The Cache for Diffusion Language Models
Xinyin Ma (National University of Singapore), Xinchao Wang (National University of Singapore)
OptimizationComputational EfficiencyTransformerLarge Language ModelDiffusion modelText
🎯 What it does: A KV-Cache mechanism for diffusion language models (DLM), called dKV-Cache, is proposed to accelerate inference.
DLoFT: Gradient-Decoupled Fine-Tuning for Generalizable Long Chain-of-Thought Reasoning
Sitong Wu (Chinese University of Hong Kong), Jiaya Jia (Chinese University of Hong Kong)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
🎯 What it does: This paper proposes a gradient decoupling fine-tuning method called DLoFT, which allows large language models to learn only the general LongChain-of-Thought reasoning paradigm during training, avoiding overfitting to problem-specific content.
DMol: A Highly Efficient and Chemical Motif-Preserving Molecule Generation Platform
Peizhi Niu (University of Illinois Urbana-Champaign), Olgica Milenkovic (University of Illinois Urbana-Champaign)
GenerationDrug DiscoveryGraph Neural NetworkDiffusion modelGraph
🎯 What it does: A novel graph diffusion model DMol is proposed for efficiently generating drug molecules that retain key molecular substructures.
DMWM: Dual-Mind World Model with Long-Term Imagination
Lingyi Wang (Virginia Tech), Naren Ramakrishnan (Virginia Tech)
Robotic IntelligenceReinforcement Learning from Human FeedbackRecurrent Neural NetworkReinforcement LearningWorld ModelSequential
🎯 What it does: A dual-process-based world model (DMWM) is proposed, which combines the fast intuitive RSSM (System 1) with a logical reasoning neural network (System 2) and achieves long-term, logically consistent imagination and planning through bidirectional feedback.
DNA-DetectLLM: Unveiling AI-Generated Text via a DNA-Inspired Mutation-Repair Paradigm
Xiaowei Zhu (Institute of Information Engineering Chinese Academy of Sciences), Yanan Cao (Institute of Information Engineering Chinese Academy of Sciences)
ClassificationAnomaly DetectionTransformerLarge Language ModelText
🎯 What it does: A zero-shot AI-generated text detection method based on the DNA mutation-repair paradigm, DNA-DetectLLM, is proposed, which utilizes an ideal AI sequence for stepwise repair and quantifies the difficulty of repair as a detection signal.
DNAEdit: Direct Noise Alignment for Text-Guided Rectified Flow Editing
Chenxi Xie (Hong Kong Polytechnic University), Lei Zhang (Hong Kong Polytechnic University)
Flow-based ModelRectified FlowImageTextBenchmark
🎯 What it does: A text-guided image editing method named DNAEdit is proposed, which combines Direct Noise Alignment (DNA) and Motion Velocity Guidance (MVG) techniques to achieve high-quality editing under the Rectified Flow framework.
Do Automatic Factuality Metrics Measure Factuality? A Critical Evaluation
Sanjana Ramprasad (Northeastern University), Byron C Wallace
Large Language ModelPrompt EngineeringText
🎯 What it does: A systematic evaluation and stress testing of existing automated factual consistency assessment metrics to explore whether they truly measure the factual consistency between the text and the source.
Do different prompting methods yield a common task representation in language models?
Guy Davidson (New York University), Adina Williams (Meta)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper studies the task representation of large language models and explores whether demonstration prompts and text instructions can activate the same task representation.
Do Language Models Use Their Depth Efficiently?
Róbert Csordás (Stanford University), Christopher Potts (Stanford University)
TransformerLarge Language ModelText
🎯 What it does: This paper explores whether the depth of large pre-trained language models is effectively utilized through causal interventions, cosine similarity, Logitlens, integrated gradients, and other analyses of residual flow, inter-layer interactions, and layer contributions.
Do LLMs Really Forget? Evaluating Unlearning with Knowledge Correlation and Confidence Awareness
Rongzhe Wei (Georgia Institute of Technology), Pan Li (Georgia Institute of Technology)
TransformerLarge Language ModelReinforcement LearningGraph
🎯 What it does: A machine learning model (LLM) forgetfulness evaluation framework based on knowledge graphs and confidence is proposed, and reasoning evaluation is conducted through a powerful LLM discriminator.
Do LVLMs Truly Understand Video Anomalies? Revealing Hallucination via Co-Occurrence Patterns
Menghao Zhang (Beijing University of Posts and Telecommunications), Jingyu Wang (Beijing University of Posts and Telecommunications)
Anomaly DetectionOptimizationTransformerVision Language ModelContrastive LearningVideo
🎯 What it does: This paper reveals the hallucination phenomenon caused by the reliance on statistical shortcuts of large visual-language models (LVLM) in video anomaly detection (VAD) through the analysis of visual-text co-occurrence patterns, and proposes a training framework based on Contrastive Preference Optimization (VAD-DPO) that forces the model to focus on scene semantics rather than co-occurrence biases.
Do Neural Networks Need Gradient Descent to Generalize? A Theoretical Study
Yotam Alexander (Tel Aviv University), Nadav Cohen (Tel Aviv University)
OptimizationTabular
🎯 What it does: This paper theoretically investigates whether gradient descent is a necessary condition for achieving good generalization in over-parameterized neural networks, focusing on matrix factorization models and analyzing the impact of width and depth on the validity of the 'volume hypothesis'.
Do-PFN: In-Context Learning for Causal Effect Estimation
Jake Robertson (Prior Labs), Bernhard Schölkopf (Max Planck Institute for Intelligent Systems)
Meta LearningTransformerTabular
🎯 What it does: This paper proposes Do-PFN, a pre-trained transformer that predicts causal effects from observational data through context learning.
Doctor Approved: Generating Medically Accurate Skin Disease Images through AI-Expert Feedback
Janet Wang (Tulane University), Jihun Hamm (Tulane University)
GenerationData SynthesisLarge Language ModelReinforcement LearningDiffusion modelImage
🎯 What it does: This paper proposes a framework called MAGIC based on AI-expert collaboration, which utilizes large language models to evaluate medical checklists to guide diffusion models in generating skin disease images with high clinical accuracy, and uses these synthetic images for data augmentation.
Document Summarization with Conformal Importance Guarantees
Bruce Kuwahara (Signal 1 AI), Jesse C. Cresswell (Layer 6 AI)
Large Language ModelPrompt EngineeringText
🎯 What it does: Proposes the Conformal Importance Summarization framework, which uses conformal prediction to provide distribution-independent, finite sample coverage guarantees for the retention rate of key sentences in extractive summaries;
DoDo-Code: an Efficient Levenshtein Distance Embedding-based Code for 4-ary IDS Channel
Alan J.X. Guo (Tianjin University), Xin Chen (Tianjin University)
Convolutional Neural NetworkSequential
🎯 What it does: A high-rate single IDS error-correcting 4-ary code called DoDo-Code is designed.
Does Object Binding Naturally Emerge in Large Pretrained Vision Transformers?
Yihao Li (University of Pennsylvania), Konrad Kording
SegmentationRepresentation LearningTransformerContrastive LearningImage
🎯 What it does: This study investigates whether object binding capabilities naturally emerge in large pre-trained Vision Transformers (ViT), defines the 'IsSameObject' metric, and validates its decodability through linear/quadratic detectors; further analyzes the low-dimensional subspace of binding information in the feature space and examines its impact on attention allocation and downstream segmentation tasks.
Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?
Yang Yue (LeapLab, Tsinghua University), Gao Huang (Tsinghua University)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: A systematic evaluation was conducted on whether RLVR (Reinforcement Learning with Verifiable Rewards) can enable LLMs to acquire reasoning capabilities that surpass those of the base model, using pass@k curves to measure reasoning boundaries.
Does Representation Guarantee Welfare?
Jakob de Raaij (Harvard University), Alexandros Psomas (Purdue University)
Tabular
🎯 What it does: This study investigates the role of descriptive representation in citizen assembly decision-making and explores the impact of different characteristic intersections (m-representation) on the accuracy of social welfare estimates.
Does Stochastic Gradient really succeed for bandits?
Dorian Baudry (Univ. Grenoble Alpes), Patrick Rebeschini (University of Oxford)
OptimizationReinforcement Learning
🎯 What it does: This paper conducts a theoretical study on the convergence and parameter tuning of the Stochastic Gradient Ascent strategy (SGB) in the multi-armed bandit problem, providing insights into the impact of the learning rate η on cumulative rewards (or regret), and presents upper and lower bounds for the two-armed and K-armed cases.
Does Thinking More Always Help? Mirage of Test-Time Scaling in Reasoning Models
Soumya Suvra Ghosal (University of Maryland), Amrit Singh Bedi (University of Central Florida)
Large Language ModelReinforcement LearningTextChain-of-Thought
🎯 What it does: This paper conducts systematic experiments and analyses on the effect of 'thinking more' to extend reasoning chains in large inference language models during testing, finding that performance first increases and then decreases, indicating the phenomenon of 'overthinking';
Domain Adaptive Hashing Retrieval via VLM Assisted Pseudo-Labeling and Dual Space Adaptation
Jingyao Li (Jilin University), Shuai Lü (Jilin University)
RetrievalDomain AdaptationVision Language ModelImage
🎯 What it does: This paper proposes an unsupervised domain adaptation hashing retrieval framework assisted by visual-language model for pseudo-label generation and dual-space adaptation (VPDS), aimed at improving cross-domain and single-domain retrieval performance.
Domain-RAG: Retrieval-Guided Compositional Image Generation for Cross-Domain Few-Shot Object Detection
Yu Li (Fudan University), Yu-Gang Jiang (Fudan University)
Object DetectionGenerationData SynthesisDiffusion modelImageRetrieval-Augmented Generation
🎯 What it does: A training-free, retrieval-guided combined image generation framework called Domain-RAG is proposed to generate synthetic images that conform to domain distribution while preserving the original foreground in cross-domain few-shot object detection.
Domain-Specific Pruning of Large Mixture-of-Experts Models with Few-shot Demonstrations
zican Dong, Zhifeng Wang (EBTech Co. Ltd)
CompressionDomain AdaptationComputational EfficiencyMixture of ExpertsTabular
🎯 What it does: This paper proposes a domain-specific pruning method called EASY-EP, which performs high compression rate expert pruning on large mixture of experts (MoE) models, addressing the memory bottleneck issue of large-scale MoE models.
Don't be lazy: CompleteP enables compute-efficient deep transformers
Nolan Simran Dey (Cerebras Systems), Joel Hestness (Cerebras Systems)
OptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: The study investigates the parameterization strategy when simultaneously expanding depth and width in Transformers, proposing and validating that CompleteP (α=1) can achieve deep transferability of hyperparameters (learning rate, initialization standard deviation, etc.).
Don’t Forget the Enjoin: FocalLoRA for Instruction Hierarchical Alignment in Large Language Models
Zitong Shi (Microsoft Research), Wei Wang (Stanford University)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: FocalLoRA is proposed to enhance the compliance rate of large language models with system instructions in scenarios of instruction-level conflicts by identifying and fine-tuning only a small number of focal heads.
Don't Just Chase “Highlighted Tokens” in MLLMs: Revisiting Visual Holistic Context Retention
Xin Zou (Hong Kong University of Science and Technology), Xuming Hu (Shanghai Jiao Tong University)
RecognitionOptimizationComputational EfficiencyTransformerLarge Language ModelVision Language ModelImageMultimodality
🎯 What it does: Developed HoloV, a global context-preserving visual token cropping framework for multimodal large language models.
Don’t Let It Fade: Preserving Edits in Diffusion Language Models via Token Timestep Allocation
Woojin Kim (Seoul National University), Jaeyoung Do (Seoul National University)
GenerationData SynthesisOptimizationTransformerLarge Language ModelDiffusion modelText
🎯 What it does: This paper proposes a method called Token Timestep Allocation (TTA-Diffusion) during the inference phase of diffusion language models, which assigns different denoising time steps to each token to achieve a softened token ordering, thereby addressing the issues of control failure and decreased fluency caused by 'update-forgetting'.
DON’T NEED RETRAINING: A Mixture of DETR and Vision Foundation Models for Cross-Domain Few-Shot Object Detection
Chang-Han Liu, Yang Gao (Nanjing University)
Object DetectionDomain AdaptationTransformerMixture of ExpertsImage
🎯 What it does: A cross-domain few-shot object detection method based on the Mixture-of-Experts (MoE) architecture is proposed, which integrates visual foundation models (VFM) as expert features into the pre-trained DETR detector, achieving cross-domain inference without re-training on base classes.
Don’t Think Longer, Think Wisely: Optimizing Thinking Dynamics for Large Reasoning Models
Sohyun An (University of California), Cho-Jui Hsieh (University of California)
OptimizationComputational EfficiencyTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: A DTO framework is proposed, which optimizes the inference path of large-scale reasoning models by dynamically identifying and pruning thinking patterns, reducing lengthy reasoning;
Don’t Trade Off Safety: Diffusion Regularization for Constrained Offline RL
Junyu Guo (University of California), Javad Lavaei (University of California)
Safty and PrivacyReinforcement LearningDiffusion modelTabularBenchmark
🎯 What it does: In offline safe reinforcement learning, a diffusion model is first used to approximate the behavior policy, and then this model is used to perform inverse KL regularization on the target policy, dynamically balancing rewards and costs through gradient manipulation.
Doodle to Detect: A Goofy but Powerful Approach to Skeleton-based Hand Gesture Recognition
Sang Hoon Han (Sogang University), Sung In Cho (Sogang University)
RecognitionPose EstimationTransformerImageVideo
🎯 What it does: Proposes the SKETCH framework, which transforms the original four-dimensional skeleton sequences into images and uses a visual Transformer for online gesture recognition.
DoseSurv: Predicting Personalized Survival Outcomes under Continuous-Valued Treatments
Moritz Gögl (University of Oxford), Tingting Zhu (University of Oxford)
TabularTime Series
🎯 What it does: A neural network model named DoseSurv is proposed to estimate the heterogeneous treatment effects between continuous dose and survival time (TTE), supporting both single-arm and multi-arm treatment schemes.
DOTA: Distributional Test-time Adaptation of Vision-Language Models
Zongbo Han (Beijing University of Posts and Telecommunications), Changqing Zhang (Tianjin University)
ClassificationDomain AdaptationTransformerVision Language ModelImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A distributed testing adaptive method named Dota is proposed, which achieves dynamic adaptation of visual-language models by continuously estimating the Gaussian distribution of test data and using Bayesian inference without performing gradient backpropagation.
Double Descent Meets Out-of-Distribution Detection: Theoretical Insights and Empirical Analysis on the Role of Model Complexity
Mouin Ben Ammar, Gianni Franchi
Anomaly DetectionConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper explores the relationship between model capacity and OOD detection performance under post-hoc methods, and verifies the existence of the double descent phenomenon.
Doubly Robust Alignment for Large Language Models
Erhan Xu (London School of Economics), Chengchun Shi (London School of Economics)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes a double-robust preference optimization method (DRPO) for fine-tuning large language models within the RLHF framework, achieving a more robust alignment with human preferences.
Doubly-Robust Estimation of Counterfactual Policy Mean Embeddings
Houssam Zenati (University College London), Arthur Gretton (Google Deepmind)
Drug DiscoveryImageTabular
🎯 What it does: A causal distribution estimation framework for policy based on kernel mean embedding (CPME) is proposed, providing both interpolation and double-robust estimators, and designing efficient distribution difference tests and sample generation methods.
DOVE: Efficient One-Step Diffusion Model for Real-World Video Super-Resolution
Zheng Chen (Shanghai Jiao Tong University), Yulun Zhang (Shanghai Jiao Tong University)
RestorationSuper ResolutionTransformerDiffusion modelVideo
🎯 What it does: DOVE is proposed, a single-step diffusion model for real-world video super-resolution;
DOVTrack: Data-Efficient Open-Vocabulary Tracking
Zekun Qian (Tianjin University), Wei Feng (Tianjin University)
Object DetectionObject TrackingConvolutional Neural NetworkDiffusion modelContrastive LearningVideo
🎯 What it does: A data-efficient open vocabulary multi-object tracking method is proposed on sparsely annotated video data, utilizing diffusion feature generation, dynamic group contrastive learning, and adaptive localization loss, significantly improving association, classification, and localization performance.
DP-LLM: Runtime Model Adaptation with Dynamic Layer-wise Precision Assignment
Sangwoo Kwon (Seoul National University), Yeonhong Park (Seoul National University)
TransformerLarge Language ModelText
🎯 What it does: Proposes DP-LLM, a mechanism that dynamically allocates quantization precision layer by layer on the device according to runtime requirements;
DP²O-SR: Direct Perceptual Preference Optimization for Real-World Image Super-Resolution
Rongyuan Wu (Hong Kong Polytechnic University), Lei Zhang (Hong Kong Polytechnic University)
Super ResolutionOptimizationDiffusion modelImage
🎯 What it does: Utilizing the randomness of pre-trained text-to-image diffusion models, we directly optimize the perceptual quality of real-world image super-resolution models by constructing perceptual rewards and preference comparisons.
DPA: A one-stop metric to measure bias amplification in classification datasets
Bhanu Tokas (Arizona State University), Hannah Kerner (Arizona State University)
ClassificationConvolutional Neural NetworkImageMultimodalityTabular
🎯 What it does: Proposes and implements Directional Predictability Amplification (DPA), a metric that quantifies the amplification of bias in classification datasets and distinguishes between positive and negative amplification directions.
DPAIL: Training Diffusion Policy for Adversarial Imitation Learning without Policy Optimization
Yunseon Choi (KAIST), Kee-Eung Kim (KAIST)
Adversarial AttackRobotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningDiffusion modelMultimodality
🎯 What it does: Proposes a diffusion strategy adversarial imitation learning framework (DPAIL) that does not require policy optimization.
Dr. RAW: Towards General High-Level Vision from RAW with Efficient Task Conditioning
Wenjun Huang (University of California), Mohsen Imani (University of California)
Object DetectionSegmentationPose EstimationSupervised Fine-TuningImage
🎯 What it does: A unified framework named Dr. RAW is proposed, which performs high-level visual tasks directly on camera RAW data and achieves parameter-efficient tuning through preprocessing, sensor prior prompts (SPP), and LoRA.
Drag-and-Drop LLMs: Zero-Shot Prompt-to-Weights
Zhiyuan Liang (National University of Singapore), Kai Wang (National University of Singapore)
TransformerLarge Language ModelPrompt EngineeringTextMultimodality
🎯 What it does: This paper proposes Drag-and-Drop LLMs (DnD), a gradient-free, zero-shot model adaptation framework that maps a small number of unlabeled task prompts to LoRA weight updates;
DreamLight: Towards Harmonious and Consistent Image Relighting
Yong Liu (Tsinghua University), Yansong Tang (Tsinghua University)
Image TranslationImage HarmonizationGenerationDiffusion modelImageText
🎯 What it does: The DreamLight model is proposed to achieve unified image relighting, allowing seamless composition of subjects into new backgrounds under image or text prompts while maintaining consistent lighting and tone.
DreamPRM: Domain-reweighted Process Reward Model for Multimodal Reasoning
Qi Cao (University of California San Diego), Pengtao Xie (University of California San Diego)
OptimizationData-Centric LearningReinforcement LearningMultimodalityChain-of-Thought
🎯 What it does: This paper proposes DreamPRM, a dual-layer optimization-based multimodal process reward model training framework that can adaptively reweight different multimodal reasoning datasets to enhance the model's generalization ability.
DreamVLA: A Vision-Language-Action Model Dreamed with Comprehensive World Knowledge
Wenyao Zhang, Xin Jin
Depth EstimationRobotic IntelligenceTransformerLarge Language ModelVision-Language-Action ModelDiffusion modelOptical FlowImageVideoMultimodality
🎯 What it does: The DreamVLA framework is proposed, which uses future world knowledge prediction (dynamic regions, depth, semantics) as an intermediate reasoning step, integrating vision, language, and action to achieve a perception-prediction-action loop for robotic manipulation.
DRIFT: Dynamic Rule-Based Defense with Injection Isolation for Securing LLM Agents
Hao Li (Washington University in St. Louis), Chaowei Xiao (Johns Hopkins University)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: A dynamic rule isolation framework called DRIFT is proposed to defend against prompt injection attacks in LLM agent systems.
DriveDPO: Policy Learning via Safety DPO For End-to-End Autonomous Driving
ShuYao Shang, Zhaoxiang Zhang (Institute of Automation, Chinese Academy of Sciences)
Autonomous DrivingKnowledge DistillationTransformerReinforcement LearningContrastive LearningMultimodality
🎯 What it does: The DriveDPO framework is proposed, which first integrates human imitation similarity and rule-based safety scores into a single supervised distribution through unified policy distillation, and then fine-tunes the policy using safe DPO to learn safe and human-like driving trajectories.
DrivingRecon: Large 4D Gaussian Reconstruction Model For Autonomous Driving
Hao LU, Ying-Cong Chen
Autonomous DrivingGaussian SplattingOptical FlowImagePoint Cloud
🎯 What it does: A driving scene reconstruction model called DrivingRecon based on a 4D Gaussian distribution is proposed, which can quickly generate high-quality dynamic four-dimensional scenes in a single forward inference.
DSAS: A Universal Plug-and-Play Framework for Attention Optimization in Multi-Document Question Answering
Jiakai Li (University of Electronic Science and Technology of China), Ke Qin (University of Electronic Science and Technology of China)
OptimizationTransformerLarge Language ModelText
🎯 What it does: This paper proposes Dual-Stage Adaptive Sharpening (DSAS), a training-free plugin that optimizes the attention of Transformers in multi-document question answering tasks, addressing long-distance dependencies and the 'intermediate defocus' problem.
DSCS: Fast CPDAG-Based Verification of Collapsible Submodels in High-Dimensional Bayesian Networks
Wentao Wu (Northeast Normal University), Jianhua Guo (Beijing Technology and Business University)
Graph Neural NetworkGraph
🎯 What it does: This study investigates how to quickly verify the estimable collapsibility of submodels in high-dimensional Bayesian networks and proposes the DSCS algorithm based on CPDAG.
DSRF: A Dynamic and Scalable Reasoning Framework for Solving RPMs
Chengtai Li (University of Nottingham Ningbo China), Xudong Jiang (Nanyang Technological University)
ImageVideo
🎯 What it does: A dynamic scalable reasoning framework (DSRF) is proposed to address the issues of rule scalability and adaptability in abstract visual reasoning tasks.
Dual Alignment Framework for Few-shot Learning with Inter-Set and Intra-Set Shifts
Siyang Jiang (Chinese University of Hong Kong), Ming-syan Chen
ClassificationMeta LearningGenerative Adversarial NetworkImage
🎯 What it does: To address the distribution shift issues between the support set and the query set, as well as within each of them in few-shot learning, this paper proposes the Dual Support-Query Shift challenge and presents the DUAL framework for robust feature extraction and dual regularization optimal transport alignment.
Dual Data Alignment Makes AI-Generated Image Detector Easier Generalizable
Ruoxin Chen (Tencent YouTu Lab), Shouhong Ding (Tencent YouTu Lab)
Object DetectionData-Centric LearningAuto EncoderImageBenchmark
🎯 What it does: This study investigates a dual domain data alignment method (Dual Data Alignment, DDA) for detecting AI-generated images, synchronously aligning real images and synthetic images at both the pixel and frequency levels to reduce bias and enhance detection generalization capabilities.
Dual Prototype-Enhanced Contrastive Framework for Class-Imbalanced Graph Domain Adaptation
Xin Ma (Sichuan University), Jiancheng Lv
Domain AdaptationGraph Neural NetworkContrastive LearningGraph
🎯 What it does: A dual-prototype enhanced contrastive framework (ImGDA) is proposed for unsupervised graph domain adaptation under severe label imbalance in the source domain.
Dual-Comb Ghost Imaging with Transformer-Based Reconstruction for Optical Fiber Endomicroscopy
David Dang (University of California), Ho Wai Howard Lee (University of California)
CompressionTransformerImage
🎯 What it does: Under the hardware constraints of single-core optical fibers and single-pixel detectors, snapshot compressed Ghost Imaging is achieved by combining dual-frequency comb interferometry, and a Transformer architecture (Optical Ghost-GPT) is utilized for high-speed, high-quality reconstruction of bucket values at extremely low sampling rates, completing real-time, low-size optical endoscopic imaging.
Dual-Flow: Transferable Multi-Target, Instance-Agnostic Attacks via $\textit{In-the-wild}$ Cascading Flow Optimization
Yixiao Chen (Tsinghua University), Junliang Xing (Tsinghua University)
Adversarial AttackDiffusion modelImageStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Proposes a Dual-Flow framework that generates initial perturbations using a pre-trained diffusion ODE flow and fine-tunes the reverse velocity function through LoRA, forming a multi-target instance-independent adversarial attack;
Dual-Path Temporal Decoder for End-to-End Multi-Object Tracking
Hyunseop Kim (Chungnam National University), Yeong Jun Koh (Chungnam National University)
Object TrackingTransformerVideo
🎯 What it does: A dual-path temporal decoder is proposed, combining detection and tracking in an end-to-end Transformer framework to achieve multi-object tracking.
Dual-Res Tandem Mamba-3D: Bilateral Breast Lesion Detection and Classification on Non-contrast Chest CT
Jiaheng Zhou (DAMO Academy, Alibaba Group), Yuxing Tang
ClassificationObject DetectionSegmentationConvolutional Neural NetworkImageBiomedical DataComputed Tomography
🎯 What it does: A joint multi-task framework for breast lesion detection and malignancy classification on non-contrast chest CT is proposed.
Dual-Space Semantic Synergy Distillation for Continual Learning of Unlabeled Streams
Donghao Sun (Xidian University), Cheng Deng (Xidian University)
Knowledge DistillationTransformerLarge Language ModelPrompt EngineeringImageTextMultimodality
🎯 What it does: A dual-space semantic collaborative distillation method is proposed, which utilizes a small number of representative samples to generate hierarchical text descriptions through a multimodal model, and integrates visual and textual pseudo-labels to enhance the performance of unlabeled incremental learning.
Dual-Stage Value-Guided Inference with Margin-Based Reward Adjustment for Fast and Faithful VLM Captioning
Ankan Deria (Mohamed bin Zayed University of AI), Imran Razzak
GenerationComputational EfficiencyTransformerReinforcement LearningVision Language ModelImageText
🎯 What it does: A two-stage value-guided reasoning framework, ViMaR, is proposed, significantly enhancing the credibility and detail of image descriptions generated by VLM.
DUAL: Learning Diverse Kernels for Aggregated Two-sample and Independence Testing
Zhijian Zhou (University of Melbourne), Feng Liu (University of Melbourne)
OptimizationImage
🎯 What it does: A method named DUAL is proposed, which improves kernel-based two-sample and independence tests using multi-kernel learning and diversity metrics, significantly enhancing test power while maintaining control over Type I error.
DualCnst: Enhancing Zero-Shot Out-of-Distribution Detection via Text-Image Consistency in Vision-Language Models
Fayi Le (Fujian University of Technology), Zhuo-Xu Cui (Shenzhen Institutes of Advanced Technology)
GenerationAnomaly DetectionTransformerVision Language ModelDiffusion modelImageText
🎯 What it does: The DualCnst framework is proposed to achieve zero-shot OOD detection by generating synthetic images of ID and OOD categories.
DualEqui: A Dual-Space Hierarchical Equivariant Network for Large Biomolecules
Junjie Xu (Pennsylvania State University), Suhang Wang (Pennsylvania State University)
Protein Structure PredictionGraph Neural NetworkBiomedical DataBenchmark
🎯 What it does: This paper proposes DualEquiNet, a dual-space hierarchical equivariant network for the property prediction of large-scale biomolecules (RNA, proteins) in three-dimensional structures.
DualFocus: Depth from Focus with Spatio-Focal Dual Variational Constraints
Sungmin Woo (Yonsei University), Sangyoun Lee (Yonsei University)
Depth EstimationConvolutional Neural NetworkImage
🎯 What it does: A depth estimation framework called DualFocus based on focus stacking is proposed.
DualMPNN: Harnessing Structural Alignments for High-Recovery Inverse Protein Folding
Xuhui Liao (Harbin Institute of Technology), Junjie Chen (Harbin Institute of Technology)
Protein Structure PredictionGraph Neural NetworkGraph
🎯 What it does: A dual-stream information transmission neural network, DualMPNN, is designed to guide protein reverse folding using structural alignment templates.
DualOptim: Enhancing Efficacy and Stability in Machine Unlearning with Dual Optimizers
Xuyang Zhong (City University of Hong Kong), Chen Liu (City University of Hong Kong)
GenerationOptimizationLarge Language ModelSupervised Fine-TuningDiffusion modelImageText
🎯 What it does: A Dual Optimizer (DualOptim) is proposed and validated, which uses an adaptive learning rate optimizer (such as Adam) for the forgetting target and the original optimizer (such as SGD) for the retaining target, decoupling momentum between the two to enhance the effectiveness and stability of machine unlearning (MU).
DUET: Dual-Perspective Pseudo Labeling and Uncertainty-aware Exploration & Exploitation Training for Source-Free Domain Adaptation
Jae Yun Lee (Sogang University), Sung In Cho (Sogang University)
Domain AdaptationKnowledge DistillationRepresentation LearningContrastive LearningImage
🎯 What it does: In the source-free domain adaptation (SFDA) task, a dual-view pseudo-label generation and uncertainty-aware visual optimization framework is proposed, utilizing the collaborative predictions of the target model and CLIP to produce more reliable pseudo-labels, while dynamically balancing exploration and exploitation during training.
DuetGraph: Coarse-to-Fine Knowledge Graph Reasoning with Dual-Pathway Global-Local Fusion
Jin Li (University of Science and Technology of China), Xike Xie (University of Science and Technology of China)
Graph Neural NetworkGraph
🎯 What it does: This paper proposes DuetGraph, which constructs a dual-channel global-local fusion and coarse-fine decomposition reasoning framework to alleviate the over-smoothing problem in knowledge graph reasoning.
DUO: No Compromise to Accuracy Degradation
Jinda Jia (Indiana University), Xin Liu (ByteDance Inc.)
CompressionOptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: The DUO framework is proposed, which alleviates accuracy degradation by inserting high-precision gradient synchronization steps during gradient compression training.
DuoGPT: Training-free Dual Sparsity through Activation-aware Pruning in LLMs
Ruokai Yin (Yale University), Priyadarshini Panda (Yale University)
CompressionOptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: The DuoGPT framework is proposed to achieve unsupervised dual sparsification (weight unstructured sparsity + runtime activation sparsity), simultaneously reducing memory and computation during inference.
DuSA: Fast and Accurate Dual-Stage Sparse Attention Mechanism Accelerating Both Training and Inference
Chong Wu (City University of Hong Kong), Hong Yan (City University of Hong Kong)
OptimizationComputational EfficiencyTransformerImageVideo
🎯 What it does: This paper proposes a dual-stage sparse attention mechanism, DuSA, to accelerate the training and inference of Transformers.
DyFlow: Dynamic Workflow Framework for Agentic Reasoning
Yanbo Wang (Mohamed bin Zayed University of Artificial Intelligence), Xiuying Chen (Mohamed bin Zayed University of Artificial Intelligence)
Large Language ModelSupervised Fine-TuningAgentic AITextBiomedical Data
🎯 What it does: We propose DyFlow, a dynamic workflow framework that utilizes LLM designers and executors to adaptively generate and adjust reasoning sub-goals based on real-time feedback during execution.
DyG-Mamba: Continuous State Space Modeling on Dynamic Graphs
Dongyuan Li (University of Tokyo), Renhe Jiang (University of Tokyo)
Recommendation SystemRecurrent Neural NetworkGraph Neural NetworkGraphTime SeriesSequentialStochastic Differential Equation
🎯 What it does: A dynamic graph modeling framework DyG-Mamba based on continuous state space models is proposed, which can capture long-term temporal dependencies and enhance robustness.
DyMoDreamer: World Modeling with Dynamic Modulation
Boxuan Zhang (Beijing Institute of Technology), Gang Wang (Beijing Institute of Technology)
Reinforcement LearningWorld ModelVideo
🎯 What it does: A new world model-based reinforcement learning algorithm DyMoDreamer has been developed, which enhances the extraction of dynamic features and temporal information through a dynamic modulation mechanism, achieving higher sample efficiency.
DyMU: Dynamic Merging and Virtual Unmerging for Efficient Variable-Length VLMs
Zhenhailong Wang (University of Illinois Urbana-Champaign), Ran Xu (Salesforce Research)
CompressionComputational EfficiencyTransformerVision Language ModelImageMultimodality
🎯 What it does: This paper proposes a training-free, adaptive visual-language model compression framework called DYMU, which can dynamically reduce the number of visual tokens based on image complexity and restore complete information through virtual decompression.
Dyn-O: Building Structured World Models with Object-Centric Representations
Zizhao Wang (University of Texas at Austin), Jiang Bian (Microsoft Research Asia)
Representation LearningReinforcement LearningContrastive LearningWorld ModelImageVideo
🎯 What it does: Developed Dyn-O, a world model based on object-centered representation, which learns object-level features and models dynamics in this space while achieving static-dynamic feature decoupling.
DynaAct: Large Language Model Reasoning with Dynamic Action Spaces
Xueliang Zhao (University of Hong Kong), Lingpeng Kong (University of Hong Kong)
OptimizationTransformerLarge Language ModelReinforcement LearningAgentic AITextBenchmark
🎯 What it does: Dynamically constructing a compressed action space through submodular functions enhances the performance of large language models in complex problem reasoning processes.
DynaGuide: Steering Diffusion Polices with Active Dynamic Guidance
Max Du, Shuran Song (Stanford University)
OptimizationRobotic IntelligenceTransformerDiffusion modelImageMultimodality
🎯 What it does: Proposes the DynaGuide method, which utilizes an external dynamics model to guide gradients during the denoising process of diffusion strategies, achieving dynamic steering during inference;
Dynam3D: Dynamic Layered 3D Tokens Empower VLM for Vision-and-Language Navigation
Zihan Wang (National University of Singapore), Gim Hee Lee (National University of Singapore)
Robotic IntelligenceTransformerVision Language ModelContrastive LearningSimultaneous Localization and MappingMultimodalityPoint Cloud
🎯 What it does: A dynamic hierarchical 3D representation framework, Dynam3D, has been developed for monocular vision-language navigation, combining a 3D patch-instance-zone vision-language model to achieve real-time environmental perception and navigation action prediction.
Dynamic Algorithm for Explainable $k$-medians Clustering under $\ell_p$ Norm
Konstantin Makarychev (Northwestern University), Liren Shan (Toyota Technological Institute at Chicago)
OptimizationExplainability and Interpretability
🎯 What it does: This paper proposes an interpretable k-medians clustering algorithm for any finite p ≥ 1, along with a corresponding dynamic version.
Dynamic and Chemical Constraints to Enhance the Molecular Masked Graph Autoencoders
Jiahui Zhang (University of Science and Technology of China), Yang Wang (University of Science and Technology of China)
Representation LearningDrug DiscoveryGraph Neural NetworkAuto EncoderContrastive LearningGraph
🎯 What it does: Introduced a dynamic and law-compliant proxy task framework DyCC for Masked Graph Autoencoders in molecular representation learning.
Dynamic Bundling with Large Language Models for Zero-Shot Inference on Text-Attributed Graphs
Yusheng Zhao (Peking University), Ming Zhang (University of Illinois Chicago)
ClassificationGraph Neural NetworkLarge Language ModelTextGraph
🎯 What it does: The DENSE method is proposed, which packages the texts of nearby nodes into bundles and queries a large language model (LLM) to obtain bundle-level labels. These labels are then used as supervisory signals to train a graph neural network (GNN) for zero-shot node classification.
Dynamic Configuration for Cutting Plane Separators via Reinforcement Learning on Incremental Graph
Mingxuan Ye (University of Science and Technology of China), Feng Wu (University of Science and Technology of China)
OptimizationGraph Neural NetworkTransformerReinforcement LearningGraph
🎯 What it does: A method called DynSep is proposed, which uses reinforcement learning to dynamically decide when to stop pruning plane separators and which separators should be activated on incremental graphs, in order to improve the solving efficiency of Mixed Integer Linear Programming (MILP).
Dynamic Diameter in High-Dimensions against Adaptive Adversary and Beyond
Kiarash Banihashem (University of Maryland), Morteza Monemizadeh (TU Eindhoven)
🎯 What it does: This study investigates the fully dynamic maintenance of the diameter of point sets and k-center clustering in high-dimensional Euclidean space, with algorithms that are robust against random adaptive adversaries (who have complete knowledge of the algorithm's random numbers).
Dynamic Diffusion Schrödinger Bridge in Astrophysical Observational Inversions
Ye Zhu (École Polytechnique), Olga Russakovsky (Princeton University)
GenerationData SynthesisComputational EfficiencyDiffusion modelScore-based ModelTime SeriesPhysics Related
🎯 What it does: The study utilizes a dynamic Schrödinger bridge model for observational inverse prediction of star-forming giant molecular clouds, proposing the Astro-DSB framework.
Dynamic Focused Masking for Autoregressive Embodied Occupancy Prediction
Yuan Sun (Rutgers University), Jorge Ortiz (Rutgers University)
Object DetectionSegmentationComputational EfficiencyGaussian SplattingImageBenchmark
🎯 What it does: This paper proposes a multi-scale autoregressive Gaussian representation framework called DFGauss, based on dynamic focus masking, for 3D occupancy prediction from a monocular RGB perspective.
Dynamic Gaussian Splatting from Defocused and Motion-blurred Monocular Videos
Xuankai Zhang (Sun Yat-sen University), Qing Zhang (Sun Yat-sen University)
RestorationData SynthesisConvolutional Neural NetworkGaussian SplattingVideo
🎯 What it does: A unified framework is proposed to synthesize high-quality new views from monocular videos containing defocus and motion blur using dynamic Gaussian scattering technology.
Dynamic Masking and Auxiliary Hash Learning for Enhanced Cross-Modal Retrieval
Shuang Zhang (Hebei Normal University), Mingying Xu (Anhui University of Science and Technology)
RetrievalTransformerContrastive LearningImageMultimodality
🎯 What it does: A dynamic masking and auxiliary hashing learning framework is proposed to enhance cross-modal retrieval performance for images and text.