ICLR 2025 Papers — Page 8
International Conference on Learning Representations · 3704 papers
CS-Bench: A Comprehensive Benchmark for Large Language Models towards Computer Science Mastery
Xiaoshuai Song (Beijing University of Posts and Telecommunications), Weiran Xu (Beijing University of Posts and Telecommunications)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: A CS-Bench benchmark has been constructed to evaluate the knowledge and reasoning abilities of LLMs in the field of computer science, with zero-shot evaluations conducted on over 30 mainstream models.
CSA: Data-efficient Mapping of Unimodal Features to Multimodal Features
Po-han Li (University of Texas at Austin), ufuk topcu
ClassificationRetrievalContrastive LearningImageMultimodality
🎯 What it does: Utilize a pre-trained unimodal encoder to map to a multimodal space through CCA, achieving the reproduction of CLIP similarity;
CtD: Composition through Decomposition in Emergent Communication
Boaz Carmeli (Technion - Israel Institute of Technology), Yonatan Belinkov (Technion - Israel Institute of Technology)
GenerationData SynthesisAuto EncoderImage
🎯 What it does: A two-stage training framework (Decompose–Compose) is proposed, which first learns a discrete concept codebook in multi-objective coordination games, and then uses this codebook to generate descriptions of new images from combinations of basic concepts, achieving perfect compositional generalization in a zero-shot setting.
Ctrl-Adapter: An Efficient and Versatile Framework for Adapting Diverse Controls to Any Diffusion Model
Han Lin (University of North Carolina at Chapel Hill), Mohit Bansal (University of North Carolina at Chapel Hill)
Image TranslationGenerationData SynthesisMixture of ExpertsDiffusion modelImageVideo
🎯 What it does: A framework named CTRL-Adapter is proposed, which utilizes the pre-trained ControlNet to transfer various control methods (depth maps, edges, poses, etc.) to any image/video diffusion model without retraining any parameters, supporting single/multiple conditions, sparse frame control, as well as downstream tasks such as video editing and style transfer.
Ctrl-U: Robust Conditional Image Generation via Uncertainty-aware Reward Modeling
Guiyu Zhang (University of Macau), Zhedong Zheng (Tsinghua University)
SegmentationGenerationReinforcement LearningDiffusion modelImage
🎯 What it does: The Ctrl-U method is proposed to enhance the controllability and quality of conditional image generation through uncertainty-aware reward modeling.
CtrLoRA: An Extensible and Efficient Framework for Controllable Image Generation
Yifeng Xu (Institute of Computing Technology, Chinese Academy of Sciences), Xilin Chen (Institute of Computing Technology, Chinese Academy of Sciences)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: This paper proposes a scalable, low-resource image-to-image control generation framework called CtrLoRA, which first trains a shared Base ControlNet and learns LoRA for different conditions, then quickly adapts to new conditions with a small amount of data.
CTSyn: A Foundation Model for Cross Tabular Data Generation
Xiaofeng Lin (University of California), Guang Cheng (University of California)
GenerationData SynthesisTransformerDiffusion modelAuto EncoderTabular
🎯 What it does: This paper proposes a cross-table synthesizer called CTSyn, which utilizes an autoencoder to map heterogeneous tables into a unified latent space and generates new tables using a conditional latent diffusion model.
CubeDiff: Repurposing Diffusion-Based Image Models for Panorama Generation
Nikolai Kalischek (ETH Zurich), Federico Tombari (Google)
GenerationData SynthesisDiffusion modelAuto EncoderImage
🎯 What it does: By using a pre-trained diffusion model in the cube mapping space, six perspectives are jointly generated and stitched into a complete 360° panoramic image;
CURIE: Evaluating LLMs on Multitask Scientific Long-Context Understanding and Reasoning
Hao Cui (Google), Subhashini Venugopalan (Google)
TransformerLarge Language ModelPrompt EngineeringTextMultimodalityBenchmarkPhysics Related
🎯 What it does: The CURIE benchmark is proposed to evaluate the long-context understanding, reasoning, and information extraction capabilities of large language models across six scientific fields (materials science, theoretical condensed matter physics, quantum computing, geospatial analysis, biodiversity, and protein structure);
Curriculum-aware Training for Discriminating Molecular Property Prediction Models
Hansi Yang (Hong Kong University of Science and Technology), James Kwok (Hong Kong University of Science and Technology)
Drug DiscoveryGraph Neural NetworkGraphTabularBiomedical Data
🎯 What it does: A training algorithm based on curriculum learning, LAC, is proposed, reformulating molecular property prediction as a node classification problem, and incorporating node-level weighted curriculum learning and edge-level pairing loss to better learn structurally similar but differently behaving molecules (activity cliffs).
Cut the Crap: An Economical Communication Pipeline for LLM-based Multi-Agent Systems
Guibin Zhang (Tongji University), Tianlong Chen (University of North Carolina at Chapel Hill)
TransformerLarge Language ModelAgentic AIText
🎯 What it does: This paper proposes AgentPrune, which utilizes a trainable low-rank graph mask to prune the space-time communication graph in LLM-based multi-agent systems in a one-time manner, significantly reducing token consumption while maintaining or even improving performance.
Cut Your Losses in Large-Vocabulary Language Models
Erik Wijmans (Apple), Philipp Kraehenbuehl
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposes the Cut Cross-Entropy (CCE) scheme, significantly reducing the memory usage of the cross-entropy layer in training large vocabulary LLMs.
CViT: Continuous Vision Transformer for Operator Learning
Sifan Wang (Yale University), Paris Perdikaris (University of Pennsylvania)
TransformerMeshGraphPhysics Related
🎯 What it does: This paper proposes the Continuous Vision Transformer (CViT), a neural operator that integrates visual Transformers with continuous coordinate embeddings to learn the input-output mapping of PDEs.
Cybench: A Framework for Evaluating Cybersecurity Capabilities and Risks of Language Models
Andy K Zhang, Percy Liang (Stanford University)
TransformerLarge Language ModelAgentic AITextBenchmark
🎯 What it does: The Cybench framework is proposed to evaluate the capabilities of language models in cybersecurity tasks, including 40 professional-level CTF tasks and subtasks.
CyberHost: A One-stage Diffusion Framework for Audio-driven Talking Body Generation
Gaojie Lin (ByteDance), Yanbo Zheng (ByteDance)
GenerationData SynthesisDiffusion modelVideoAudio
🎯 What it does: A one-stage audio-driven human animation framework called CyberHost is proposed, which can generate complete human videos from a single image and audio;
CycleResearcher: Improving Automated Research via Automated Review
Yixuan Weng (Westlake University), Linyi Yang (University College London)
TransformerLarge Language ModelReinforcement LearningAgentic AIText
🎯 What it does: A complete automated framework for the research lifecycle based on open-source LLMs has been constructed—CycleResearcher (autonomous research) and CycleReviewer (simulated peer review), achieving closed-loop improvements in research and review through iterative SimPO training.
Cyclic Contrastive Knowledge Transfer for Open-Vocabulary Object Detection
Chuhan ZHANG, Dong Zhang (Hong Kong University of Science and Technology)
Object DetectionKnowledge DistillationTransformerVision Language ModelContrastive LearningImage
🎯 What it does: Designed and implemented the CCKT-Det framework, achieving open vocabulary object detection by aligning language priors with visual region features in a cyclic manner without using additional annotations or pseudo-labels.
D-FINE: Redefine Regression Task of DETRs as Fine-grained Distribution Refinement
Yansong Peng (University of Science and Technology of China), Feng Wu (University of Science and Technology of China)
Object DetectionKnowledge DistillationTransformerImage
🎯 What it does: We propose D-FINE, a real-time object detector that achieves high-precision localization by redefining the bounding box regression task of DETR.
DailyDilemmas: Revealing Value Preferences of LLMs with Quandaries of Daily Life
Yu Ying Chiu (University of Washington), Yejin Choi (University of Washington)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: A dataset called DAILYDILEMMAS is proposed and constructed, containing 1,360 moral dilemmas in daily life with binary choices, and the value conflicts in these dilemmas are annotated based on the theory of five major values.
DAMO: Decoding by Accumulating Activations Momentum for Mitigating Hallucinations in Vision-Language Models
Kaishen Wang (University of Maryland), Kaixiong Zhou (North Carolina State University)
TransformerVision Language ModelMultimodality
🎯 What it does: A decoding method based on activation momentum, DAMO, is proposed, which utilizes visual information from early layers to correct the activations of later layers during inference, significantly reducing the hallucination outputs of large-scale visual language models (LVLMs).
DARE the Extreme: Revisiting Delta-Parameter Pruning For Fine-Tuned Models
Wenlong Deng (University of British Columbia), Christos Thrampoulidis (University of British Columbia)
TransformerSupervised Fine-TuningText
🎯 What it does: This study improves the Delta-parameter pruning (DPP) technique by addressing the limitations of the original random Drop-Rescale (DARE) method. It proposes a tunable rescaling factor DAREx-q and incorporates AdamR regularization during the fine-tuning phase in the DAREx-L2 scheme, achieving extreme pruning (up to 99%) while maintaining nearly original performance.
DarkBench: Benchmarking Dark Patterns in Large Language Models
Esben Kran (Apart Research), Mateusz Maria Jurewicz
TransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: This paper proposes the DarkBench benchmark, which systematically evaluates six types of dark patterns that occur in conversations with large language models, quantifying the manipulative behaviors of the models.
DartControl: A Diffusion-Based Autoregressive Motion Model for Real-Time Text-Driven Motion Control
Kaifeng Zhao (ETH Zurich), Siyu Tang (ETH Zurich)
GenerationReinforcement LearningDiffusion modelAuto EncoderVideoText
🎯 What it does: A diffusion-based autoregressive motion primitive model DART is proposed, achieving real-time text-driven long-term human motion generation and spatial control.
Data Center Cooling System Optimization Using Offline Reinforcement Learning
Xianyuan Zhan (Institute for AI Industry Research, Tsinghua University), Feng Zhao (Institute for AI Industry Research, Tsinghua University)
OptimizationGraph Neural NetworkReinforcement LearningTime SeriesOrdinary Differential Equation
🎯 What it does: This paper proposes a physics-aware framework based on offline reinforcement learning for energy consumption optimization in data center cooling systems, achieving closed-loop control and validation in real environments.
Data Distillation for extrapolative protein design through exact preference optimization
Mostafa Karimi (Amazon), Ron Benson (Amazon)
OptimizationKnowledge DistillationProtein Structure PredictionTransformerReinforcement LearningBiomedical Data
🎯 What it does: A deep data distillation framework (EXO) is proposed for protein extrapolation design using model-related hard triplet preference learning, achieving high fitness sequence generation through reinforced gradient direction.
Data Mixing Laws: Optimizing Data Mixtures by Predicting Language Modeling Performance
Jiasheng Ye (Fudan University), Xipeng Qiu (Fudan University)
OptimizationData-Centric LearningTransformerLarge Language ModelText
🎯 What it does: A data mixing law has been developed that can predict the impact of different training data ratios on language model performance without training the full model. By nesting this law with scaling laws for training steps and model size, it allows for the prediction of the final performance of large-scale models at various mixing ratios using small-scale experiments.
Data Pruning by Information Maximization
Haoru Tan (University of Hong Kong), XIAOJUAN QI
OptimizationData-Centric LearningImageMultimodality
🎯 What it does: This paper proposes InfoMax, a secondary optimization method that maximizes sample information while minimizing redundancy, aimed at efficiently selecting high-information sparse subsets (coresets) from large-scale datasets.
Data Scaling Laws in Imitation Learning for Robotic Manipulation
Fanqi Lin (Tsinghua University), Yang Gao (Tsinghua University)
Robotic IntelligenceDiffusion modelSimultaneous Localization and MappingTime Series
🎯 What it does: A systematic study of the data scaling laws in robot manipulation was conducted, revealing that the diversity of environments and objects has the greatest impact on generalization, and an efficient data collection strategy was proposed;
Data Selection via Optimal Control for Language Models
Yuxian Gu (Tsinghua University), Minlie Huang (Tsinghua University)
OptimizationData-Centric LearningLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes a pre-training data selection method based on optimal control theory. By deriving the necessary conditions from Pontryagin's maximum principle, it constructs a PDS framework to solve the data quality score on proxy corpora and uses a small model to predict the scores for the entire corpus to complete offline data screening.
Data Shapley in One Training Run
Jiachen T. Wang (Princeton University), Ruoxi Jia (Virginia Tech)
OptimizationComputational EfficiencyData-Centric LearningTransformerLarge Language ModelText
🎯 What it does: Proposes In-Run Data Shapley, which evaluates the contribution of each training data point to model performance using gradient updates during a single training process, without the need to retrain on different subsets.
Data Taggants: Dataset Ownership Verification Via Harmless Targeted Data Poisoning
Wassim Bouaziz (Meta AI), El-Mahdi El-Mhamdi
Anomaly DetectionData-Centric LearningTransformerImage
🎯 What it does: A new dataset ownership verification method called data taggants is proposed, which utilizes clean label targeted data poisoning and gradient matching to generate specific labels for a set of discrete key samples after model training, thereby enabling black-box detection of whether the model has used the labeled dataset.
Data Unlearning in Diffusion Models
Silas Alberti (Stanford University), Stefano Ermon (Stanford University)
GenerationComputational EfficiencyData-Centric LearningDiffusion modelImage
🎯 What it does: A data unlearning method for diffusion models called SISS is proposed, which can effectively delete the memory of specified training samples while maintaining model quality.
Data-adaptive Differentially Private Prompt Synthesis for In-Context Learning
Fengyu Gao (University of Virginia), Jing Yang (University of Virginia)
Data SynthesisSafty and PrivacyPrompt EngineeringText
🎯 What it does: A new data-adaptive differential privacy algorithm, AdaDPSyn, is proposed for generating synthetic examples from private datasets and utilizing these examples for in-context learning (ICL) to protect private information.
Data-centric Prediction Explanation via Kernelized Stein Discrepancy
Mahtab Sarvmaili (Dalhousie University), Ga Wu (Dalhousie University)
Explainability and InterpretabilityData-Centric LearningConvolutional Neural NetworkImageBiomedical Data
🎯 What it does: This paper proposes an example-based prediction explanation method called HD-Explain, which utilizes kernelized Stein discrepancy (KSD) to accurately extract training samples that support a given test point from a trained model.
DataEnvGym: Data Generation Agents in Teacher Environments with Student Feedback
Zaid Khan (University of North Carolina at Chapel Hill), Mohit Bansal (University of North Carolina at Chapel Hill)
Data SynthesisReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextMultimodality
🎯 What it does: This paper presents DATAENVGYM, a testing platform for data generation agents (teachers) that simulates the process of student models iteratively improving performance by generating training data through teachers in a feedback-driven loop.
DataGen: Unified Synthetic Dataset Generation via Large Language Models
Yue Huang (University of Notre Dame), Xiangliang Zhang (University of Notre Dame)
GenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: This paper presents DATAGEN, a unified LLM-driven dataset generation framework that supports high-quality, diverse, and controllable synthesis of any text dataset, and can be used for dynamic benchmarking and data augmentation.
DataMan: Data Manager for Pre-training Large Language Models
Ru Peng (Zhejiang University), Junbo Zhao (Zhejiang University)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Designed and trained a data manager named DataMan, which performs 14 quality assessments and 15 domain labels on pre-trained corpora. Subsequently, data sampled based on these labels is used for LLM pre-training, significantly improving the model's perplexity, contextual learning, and instruction-following performance.
Dataset Distillation via Knowledge Distillation: Towards Efficient Self-Supervised Pre-training of Deep Networks
Siddharth Joshi (University of California), Baharan Mirzasoleiman (University of California)
Knowledge DistillationRepresentation LearningContrastive LearningImage
🎯 What it does: The MKDT method is proposed, achieving dataset distillation for self-supervised pre-training on unlabeled data.
Dataset Ownership Verification in Contrastive Pre-trained Models
Yuechen Xie (Zhejiang University), Mingli Song (NingboTech University)
Anomaly DetectionData-Centric LearningContrastive LearningImage
🎯 What it does: A contrastive relationship gap based on contrastive pre-training models is proposed to verify dataset ownership, targeting black-box scenarios to detect potential dataset theft.
DaWin: Training-free Dynamic Weight Interpolation for Robust Adaptation
Changdae Oh (University of Wisconsin-Madison), Dongyoon Han (NAVER AI Lab)
Domain AdaptationComputational EfficiencyTransformerMixture of ExpertsImage
🎯 What it does: A training-free, dynamic weight interpolation method called DaWin is proposed, which estimates the interpolation coefficient for each sample using the prediction entropy ratio of unlabeled test samples and reduces inference costs through clustering with a Beta mixture model.
DAWN: Dynamic Frame Avatar with Non-autoregressive Diffusion Framework for Talking head Video Generation
Hanbo Cheng (University of Science and Technology of China), Jia Pan (iFLYTEK Research)
GenerationData SynthesisTransformerDiffusion modelVideoAudio
🎯 What it does: The DAWN framework is proposed, which utilizes a non-autoregressive diffusion model to generate dynamic-length speaker videos in one go, creating realistic, lip-synchronized talking head videos from a single portrait and audio input.
DCT-CryptoNets: Scaling Private Inference in the Frequency Domain
Arjun Roy (Purdue University), Kaushik Roy (Purdue University)
Safty and PrivacyComputational EfficiencyImage
🎯 What it does: A framework for fully homomorphic encrypted neural network inference in the frequency domain (DCT) called DCT-CryptoNets is proposed to reduce the computational overhead of nonlinear activation and homomorphic guiding switches.
DebGCD: Debiased Learning with Distribution Guidance for Generalized Category Discovery
Yuanpei Liu (Visual AI Lab, University of Hong Kong), Kai Han (Visual AI Lab, University of Hong Kong)
ClassificationRecognitionKnowledge DistillationTransformerImage
🎯 What it does: Proposes the DebGCD framework, which combines auxiliary debiasing learning, semantic distribution detection, and distribution-guided curriculum learning in the generalized category discovery (GCD) task, significantly improving the model's recognition and clustering performance for both known and unknown categories.
Debiasing Federated Learning with Correlated Client Participation
Zhenyu Sun (Northwestern University), Ermin Wei (Northwestern University)
Federated LearningImageTabular
🎯 What it does: This paper studies the non-uniform and correlated client participation patterns in cross-device federated learning, proposing a high-order Markov chain model to characterize client participation under minimum interval constraints, and analyzes the asymptotic bias of the FedAvg algorithm based on this model.
Debiasing Mini-Batch Quadratics for Applications in Deep Learning
Lukas Tatzel (University of Tübingen), Philipp Hennig (University of Tübingen)
OptimizationConvolutional Neural NetworkTransformerImage
🎯 What it does: Research and correct the systematic bias of second-order approximations in small-batch sample calculations in deep learning, proposing two debiasing strategies;
dEBORA: Efficient Bilevel Optimization-based low-Rank Adaptation
Emanuele Zangrando (Gran Sasso Science Institute), Francesco Tudisco (Miniml.AI Ltd)
OptimizationTransformerSupervised Fine-TuningImageText
🎯 What it does: A dual-layer optimization framework named dEBORA is proposed and implemented for low-rank adapter fine-tuning on large pre-trained models, capable of dynamically selecting the optimal rank for each layer during training, achieving a balance between parameter efficiency and performance.
Decentralized Optimization with Coupled Constraints
Demyan Yarmoshik (Moscow Institute of Physics and Technology), Dmitry Kovalev (Yandex)
OptimizationFederated LearningTabular
🎯 What it does: An optimal first-order algorithm for decentralized optimization with coupled affine constraints is proposed, along with a matching lower bound.
Decentralized Sporadic Federated Learning: A Unified Algorithmic Framework with Convergence Guarantees
Shahryar Zehtabi (Purdue University), Christopher Brinton
Federated LearningConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a framework called DSpodFL that introduces intermittency in both local gradient computation and model aggregation in decentralized federated learning, and provides convergence analysis under both convex and non-convex objectives.
DeciMamba: Exploring the Length Extrapolation Potential of Mamba
Assaf Ben-Kish (Tel Aviv University), Raja Giryes (Tel Aviv University)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper systematically analyzes the bottlenecks of the Mamba model in length-extrapolation through visualization, measurement, and experiments, finding that its effective receptive field (ERF) is limited by the length of the training sequence. It then proposes a dynamic pooling mechanism (DeciMamba) based on the importance scores of the 'delta_t' from the S6 layer, achieving context compression and expansion for long sequences, thereby significantly enhancing Mamba's long sequence inference capability without retraining.
Decision Information Meets Large Language Models: The Future of Explainable Operations Research
Yansen Zhang (City University of Hong Kong), Chen Ma (City University of Hong Kong)
OptimizationExplainability and InterpretabilityTransformerLarge Language ModelTextTabularBenchmark
🎯 What it does: Proposed the Explainable Operations Research (EOR) framework, which utilizes large language models (LLM) to provide interpretable decisions and explanations for operational optimization problems, and introduces the concept of 'decision information' along with a bipartite graph-based quantification method;
Decision Tree Induction Through LLMs via Semantically-Aware Evolution
Tennison Liu (University of Cambridge), Mihaela van der Schaar (University of Cambridge)
ClassificationOptimizationLarge Language ModelPrompt EngineeringTabular
🎯 What it does: This paper proposes a genetic programming algorithm guided by large language models (LLM), called LLEGO, to achieve efficient evolution of decision trees.
DECO: Unleashing the Potential of ConvNets for Query-based Detection and Segmentation
Xinghao Chen (Huawei Noah's Ark Lab), Yunhe Wang (Huawei Noah's Ark Lab)
Object DetectionSegmentationConvolutional Neural NetworkImage
🎯 What it does: A query-based object detection and segmentation framework called DECO is proposed, which achieves competitive detection performance on the COCO dataset.
Decoding Game: On Minimax Optimality of Heuristic Text Generation Strategies
Sijin Chen (Princeton University), Jason Matthew Klusowski
GenerationOptimizationTransformerLarge Language ModelText
🎯 What it does: A two-player zero-sum game framework (Decoding Game) is proposed to study the theoretical optimality of decoding strategies in text generation.
Decomposition Polyhedra of Piecewise Linear Functions
Marie-Charlotte Brandenburg (Ruhr-Universität Bochum), Christoph Hertrich (University of Technology Nuremberg)
Optimization
🎯 What it does: This paper studies the differential convex function decomposition of continuous piecewise linear (CPWL) functions from the perspective of polyhedral geometry, proposing the concept of 'decomposed polyhedra' and providing its geometric structure and the polyhedral properties of optimal decomposition.
Deconstructing Denoising Diffusion Models for Self-Supervised Learning
Xinlei Chen (Meta), Kaiming He (Massachusetts Institute of Technology)
GenerationRepresentation LearningTransformerDiffusion modelAuto EncoderImage
🎯 What it does: A systematic breakdown of the denoising diffusion model (DDM) for generative tasks is conducted, ultimately resulting in a simplified low-dimensional denoising autoencoder (l-DAE) for self-supervised representation learning.
Deconstructing What Makes a Good Optimizer for Autoregressive Language Models
Rosie Zhao (Harvard University), Sham M. Kakade
OptimizationTransformerLarge Language ModelText
🎯 What it does: A systematic comparison and in-depth experiments were conducted on various optimizers (SGD, Adam, Adafactor, Lion, Signum, Adalayer, etc.) under different scales, architectures, and hyperparameter settings of autoregressive language models. The performance and stability of simplified optimizers were explored, and experiments verified that using adaptive preconditioning only on the last layer and LayerNorm can achieve results comparable to global adaptive optimizers.
Decoupled Finetuning for Domain Generalizable Semantic Segmentation
Jaehyun Pahk (POSTECH), Suha Kwak (Universität Tübingen)
SegmentationDomain AdaptationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: This study proposes a decoupled fine-tuning framework for domain generalization semantic segmentation, named DeFT, aimed at reducing the overfitting problem caused by joint fine-tuning of the pre-trained encoder and decoder.
Decoupled Graph Energy-based Model for Node Out-of-Distribution Detection on Heterophilic Graphs
Yuhan Chen (Sun Yat-sen University), Xiaochun Cao (Sun Yat-sen University)
Anomaly DetectionGraph Neural NetworkContrastive LearningGraph
🎯 What it does: Under training without OOD samples, DeGEM is proposed to achieve out-of-distribution (OOD) detection for graph nodes.
Decoupled Subgraph Federated Learning
Javad Aliakbari (Chalmers University of Technology), Alexandre Graell i Amat (Chalmers University of Technology)
Federated LearningSafty and PrivacyGraph Neural NetworkGraph
🎯 What it does: Proposes the FEDSTRUCT framework, which utilizes structural information in distributed subgraph federated learning without sharing node features to complete semi-supervised node classification tasks.
Decoupling Angles and Strength in Low-rank Adaptation
Massimo Bini (University of Tübingen), Zeynep Akata
SegmentationGenerationTransformerSupervised Fine-TuningImageText
🎯 What it does: A new parameter-efficient fine-tuning method called DeLoRA is proposed, which enhances robustness without significantly increasing computational costs by adding normalization and a learnable scale factor λ to the low-rank matrices of LoRA, decoupling angle learning from adaptation strength.
Decoupling Layout from Glyph in Online Chinese Handwriting Generation
Minsi Ren, yi chen
GenerationRecurrent Neural NetworkDiffusion modelContrastive LearningText
🎯 What it does: This paper proposes a hierarchical method that first generates the layout of running script using text content and style references, and then generates online Chinese characters in accordance with the style character by character through a diffusion model, thus completing the generation of a complete written text line.
DEEM: Diffusion models serve as the eyes of large language models for image perception
Run Luo (Shenzhen Key Laboratory for High Performance Data Mining), Min Yang (Shenzhen Key Laboratory for High Performance Data Mining)
GenerationData SynthesisRetrievalTransformerLarge Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: The DEEM method is proposed, which provides self-supervised visual consistency regularization through a diffusion model to correct the semantic distribution of the visual encoder, thereby enhancing the robustness and accuracy of large language models in image perception, generation, and cross-modal interaction.
Deep Compression Autoencoder for Efficient High-Resolution Diffusion Models
Junyu Chen (Massachusetts Institute of Technology), Song Han (Massachusetts Institute of Technology)
GenerationCompressionDiffusion modelAuto EncoderImage
🎯 What it does: This paper proposes the Deep Compression Autoencoder (DC-AE), which accelerates the training and inference of high-resolution diffusion models by improving the spatial compression ratio.
Deep Distributed Optimization for Large-Scale Quadratic Programming
Augustinos D Saravanos, Evangelos Theodorou
OptimizationTabularFinance Related
🎯 What it does: A distributed optimization architecture based on deep learning, DeepDistributedQP, is designed to solve large-scale quadratic programming problems, and a corresponding centralized version, DeepQP, is provided.
Deep Incomplete Multi-view Learning via Cyclic Permutation of VAEs
Xin Gao (Fudan University), Jian Pu (Fudan University)
GenerationRepresentation LearningAuto EncoderMultimodality
🎯 What it does: This study focuses on representation learning for missing multi-view data, proposing the MVP framework that establishes view correspondences in the latent space of VAE to infer missing views and achieve unified representation learning.
Deep Kernel Posterior Learning under Infinite Variance Prior Weights
Jorge Loria (Aalto University), Anindya Bhadra (Purdue University)
Tabular
🎯 What it does: This paper proposes the Deep α-KP, a deep α-stable kernel process, as the limit of infinitely wide Bayesian neural networks with infinitely variance weights, and provides its conditional Gaussian mixture representation and recursive kernel formula, based on which posterior inference and prediction are achieved.
Deep Kernel Relative Test for Machine-generated Text Detection
Yiliao Song (University of Adelaide), Feng Liu (University of Technology Sydney)
ClassificationGenerationTransformerLarge Language ModelText
🎯 What it does: A non-parametric machine-generated text detection method based on kernel relative testing, R-Detect, is proposed. It utilizes text vector representation and kernel maximum mean discrepancy (MMD) for statistical significance testing, achieving zero-shot detection.
Deep Learning Alternatives Of The Kolmogorov Superposition Theorem
Leonardo Ferreira Guilhoto (University of Pennsylvania), Paris Perdikaris (University of Pennsylvania)
TabularTime SeriesPhysics Related
🎯 What it does: This study explores an alternative formulation of the Kolmogorov Superposition Theorem (KST), proposes the ActNet network architecture, and validates it on various Physics-Informed Neural Network (PINN) tasks.
Deep Linear Probe Generators for Weight Space Learning
Jonathan Kahana (Hebrew University of Jerusalem), Yedid Hoshen (Hebrew University of Jerusalem)
ClassificationOptimizationConvolutional Neural NetworkImage
🎯 What it does: This study focuses on weight space learning and proposes a Deep Linear Probe Generator (ProbeGen) that predicts model properties, such as training dataset or generalization error, by learning structured probe inputs.
Deep MMD Gradient Flow without adversarial training
Alexandre Galashov (University College London), Arthur Gretton (University College London)
GenerationData SynthesisDiffusion modelGenerative Adversarial NetworkImage
🎯 What it does: A noise-conditioned maximum mean discrepancy (MMD) discriminator without adversarial training is proposed, and samples are generated using a noise-adaptive MMD gradient flow (DMMD);
Deep Networks Learn Features From Local Discontinuities in the Label Function
Prithaj Banerjee (Indian Institute of Technology Madras), Chandra Shekar Lakshminarayanan
ClassificationExplainability and InterpretabilityTabularOrdinary Differential Equation
🎯 What it does: This study investigates how deep networks learn features through gradient descent and proposes a Deep Linear Gated Network (DLGN) to capture the local discontinuities of label functions in an enumerable manner.
Deep Random Features for Scalable Interpolation of Spatiotemporal Data
Weibin Chen (PhysicsX), So Takao (California Institute of Technology)
Time Series
🎯 What it does: Using deep neural networks constructed with random features for scalable interpolation of remote sensing spatiotemporal fields;
Deep Signature: Characterization of Large-Scale Molecular Dynamics
Tiexin Qin (City University of Hong Kong), Haoliang Li (City University of Hong Kong)
Protein Structure PredictionRecurrent Neural NetworkGraph Neural NetworkGraphTime SeriesBenchmark
🎯 What it does: Proposes the Deep Signature framework for efficiently describing large-scale protein dynamics trajectories and predicting functional properties.
Deep Weight Factorization: Sparse Learning Through the Lens of Artificial Symmetries
Chris Kolb (Ludwig Maximilian University of Munich), David Rügamer (Ludwig Maximilian University of Munich)
CompressionOptimizationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: A Deep Weight Factorization (DWF) method is proposed, which decomposes network weights into D≥2 factors and achieves differentiable L1 sparsification using L2 regularization, supporting networks of arbitrary depth.
DeeperForward: Enhanced Forward-Forward Training for Deeper and Better Performance
Liang Sun (Shenzhen University), Weicheng Xie (Shenzhen University)
ClassificationRecognitionConvolutional Neural NetworkImage
🎯 What it does: This paper proposes DeeperForward, which extends the Forward-Forward training method to a 17-layer CNN and improves the learning of deep networks through average affinity and layer normalization.
DeepGate4: Efficient and Effective Representation Learning for Circuit Design at Scale
Ziyang Zheng (Chinese University of Hong Kong), Qiang Xu (Shanghai Jiao Tong University)
Computational EfficiencyRepresentation LearningGraph Neural NetworkTransformerGraph
🎯 What it does: DeepGate4 is proposed, a graph Transformer model for representation learning on large-scale circuits (up to millions or even billions of gates), aimed at addressing the over-compression issue of traditional GNNs and the secondary complexity of Transformers.
DeepLTL: Learning to Efficiently Satisfy Complex LTL Specifications for Multi-Task RL
Mathias Jackermeier (University of Oxford), Alessandro Abate (University of Oxford)
Recurrent Neural NetworkReinforcement LearningSequential
🎯 What it does: A representation method based on Büchi automata reach-avoid sequences is proposed, and a general sequence-conditioned policy is trained to achieve zero-shot execution of any LTL specification in multi-task reinforcement learning.
DeepRTL: Bridging Verilog Understanding and Generation with a Unified Representation Model
Yi Liu (National Technology Innovation Center for EDA), Qiang Xu (National Technology Innovation Center for EDA)
GenerationAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkChain-of-Thought
🎯 What it does: This paper presents DeepRTL, a unified representation model for understanding and generating Verilog code, fine-tuned through curriculum learning.
DeepSeek-Prover-V1.5: Harnessing Proof Assistant Feedback for Reinforcement Learning and Monte-Carlo Tree Search
Huajian Xin (DeepSeek-AI), Chong Ruan (DeepSeek-AI)
Large Language ModelSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
🎯 What it does: Combining large language models, reinforcement learning, and Monte Carlo tree search, we utilize feedback from the Lean verifier to enhance the completeness and efficiency of formal theorem proving.
DeepTAGE: Deep Temporal-Aligned Gradient Enhancement for Optimizing Spiking Neural Networks
Wei Liu (Beijing Key Laboratory of Super-Intelligent Security of Multi-Modal Information), Weiming Hu (School of Artificial Intelligence)
OptimizationSpiking Neural NetworkImage
🎯 What it does: The DeepTAGE method is proposed, which significantly improves the training effectiveness of SNNs by dynamically adjusting the surrogate gradient at each time step and introducing deep supervision at multiple stages and time steps in the network.
DeFT: Decoding with Flash Tree-attention for Efficient Tree-structured LLM Inference
Jinwei Yao (Westlake University), Tao Lin (Westlake University)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper proposes DEFT-Flatten, an efficient prefix sharing and load balancing attention algorithm for tree-structured LLM inference.
DELIFT: Data Efficient Language model Instruction Fine-Tuning
Ishika Agarwal (University of Illinois Urbana-Champaign), Marina Danilevsky (IBM Research)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper presents DELIFT, a unified data subset selection framework that efficiently selects the most valuable training samples during three stages: instruction tuning, task-specific fine-tuning, and continuous fine-tuning, significantly reducing the amount of data required.
DeLLMa: Decision Making Under Uncertainty with Large Language Models
Ollie Liu (University of Southern California), Willie Neiswanger (University of Southern California)
TransformerLarge Language ModelTabularTime SeriesAgriculture RelatedFinance RelatedChain-of-Thought
🎯 What it does: This paper presents DeLLMa, a decision support framework based on large language models that can make auditable and interpretable decisions in uncertain environments.
DelTA: An Online Document-Level Translation Agent Based on Multi-Level Memory
Yutong Wang (Harbin Institute of Technology), Min Zhang (Harbin Institute of Technology)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: An online document-level translation agent named DELTA has been developed, utilizing a multi-layer memory structure (proper noun records, bilingual summaries, long-term and short-term memory) combined with LLM for sentence-level translation, aiming to enhance the consistency and quality of document translation.
DELTA: DENSE EFFICIENT LONG-RANGE 3D TRACKING FOR ANY VIDEO
Tuan Duc Ngo (University of Massachusetts Amherst), Chaoyang Wang (Snap Inc.)
Object TrackingDepth EstimationComputational EfficiencyTransformerVideo
🎯 What it does: A dense 3D tracking method named DELTA is proposed, capable of real-time tracking of the 3D trajectory of each pixel in monocular video and supporting long sequences.
Democratic Training Against Universal Adversarial Perturbations
Bing Sun (Singapore Management University), Wei Zhao (Singapore Management University)
OptimizationAdversarial AttackConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: A defense framework called Democratic Training is proposed to resist Universal Adversarial Perturbation (UAP) attacks on deep networks, primarily by using low-entropy samples during the model fine-tuning process to 'weaken' the dominant features of UAP.
Demystifying Online Clustering of Bandits: Enhanced Exploration Under Stochastic and Smoothed Adversarial Contexts
Zhuohua Li (Xidian University), John C.S. Lui (Chinese University of Hong Kong)
Recommendation SystemReinforcement LearningTabular
🎯 What it does: Proposes two online clustering bandit algorithms, UniCLUB and UniSCLUB, incorporating a pure exploration phase to enhance clustering identification.
Demystifying the Token Dynamics of Deep Selective State Space Models
Thieu Vo, Tan Minh Nguyen
ImageTextStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Analyzed the dynamic behavior of tokens in the pre-trained Mamba (Selective State Space Model), proving the existence of both convergence and divergence in the one-dimensional case, and exploring its impact on model performance.
Demystifying Topological Message-Passing with Relational Structures: A Case Study on Oversquashing in Simplicial Message-Passing
Diaaeldin Taha (Max Planck Institute for Mathematics in the Sciences), Guido Montufar (University of California Los Angeles)
ClassificationOptimizationGraph Neural NetworkGraph
🎯 What it does: This study constructs an axiomatic framework that unifies degenerate complexes and relational structures, and analyzes the phenomenon of excessive compression in topological message passing based on this framework, while proposing a scalable relational reconnection strategy.
DenoiseVAE: Learning Molecule-Adaptive Noise Distributions for Denoising-based 3D Molecular Pre-training
Yurou Liu (Renmin University of China), Bing Su (Renmin University of China)
Drug DiscoveryGraph Neural NetworkAuto EncoderGraph
🎯 What it does: A 3D molecular pre-training framework DenoiseVAE is designed based on a variational autoencoder, which learns molecule-specific noise distributions and performs denoising learning.
Denoising as Adaptation: Noise-Space Domain Adaptation for Image Restoration
Kang Liao (Nanyang Technological University), Chen Change Loy (Nanyang Technological University)
RestorationDomain AdaptationDiffusion modelContrastive LearningImage
🎯 What it does: Proposes using diffusion models in the noise space for domain adaptation of image restoration models to narrow the distribution gap between synthetic and real images.
Denoising Autoregressive Transformers for Scalable Text-to-Image Generation
Jiatao Gu (Apple), Shuangfei Zhai (Apple)
GenerationData SynthesisTransformerDiffusion modelImageText
🎯 What it does: This paper proposes and implements the Denoising Autoregressive Transformer (DART), which combines non-Markov diffusion with autoregressive Transformers for scalable text-to-image generation.
Denoising Task Difficulty-based Curriculum for Training Diffusion Models
Jin-Young Kim (Ajou University), Hyun-Gyoon Kim (Ajou University)
GenerationDiffusion modelImage
🎯 What it does: Proposes a curriculum learning method based on denoising difficulty to improve diffusion model training.
Denoising with a Joint-Embedding Predictive Architecture
Dengsheng Chen (Meituan), Enhua Wu (Key Laboratory of System Software, Chinese Academy of Sciences and State Key Laboratory of Computer Science, Institute of Software, Chinese Academy of Sciences)
GenerationData SynthesisTransformerDiffusion modelFlow-based ModelImageVideoAudio
🎯 What it does: By combining the Joint Embedding Prediction Architecture (JEPA) with diffusion/flow matching loss, a generative model D-JEPA is proposed that can perform autoregressive sampling in continuous space, achieving high-quality conditional image generation.
Dense Video Object Captioning from Disjoint Supervision
Xingyi Zhou (Google DeepMind), Cordelia Schmid (Google DeepMind)
Object DetectionObject TrackingGenerationTransformerVideoText
🎯 What it does: This paper proposes the Dense Video Object Captioning task, which unifies detection, tracking, and generating coherent captions for all objects in a video.
DenseGrounding: Improving Dense Language-Vision Semantics for Ego-centric 3D Visual Grounding
Henry Zheng (Tsinghua University), Gao Huang (Tsinghua University)
Object DetectionSegmentationTransformerLarge Language ModelMultimodalityPoint Cloud
🎯 What it does: This paper proposes DenseGrounding, which combines LLM-assisted text enhancement and Hierarchical Scene Semantic Enhancement (HSSE) to achieve refined processing of ego-centric 3D visual localization.
DenseMatcher: Learning 3D Semantic Correspondence for Category-Level Manipulation from a Single Demo
Junzhe Zhu (Tepan Inc.), Huazhe Xu (Shanghai AI Lab)
Robotic IntelligenceDiffusion modelPoint CloudMesh
🎯 What it does: The DenseMatcher framework is proposed to achieve cross-category and cross-instance 3D dense semantic correspondence for single demonstration robotic operations and color transfer.
Density estimation with LLMs: a geometric investigation of in-context learning trajectories
Toni J.B. Liu (Cornell University), Christopher Earls
TransformerLarge Language ModelPrompt Engineering
🎯 What it does: Exploring the ability of large language models to estimate probability density functions in context learning and using InPCA to visualize their learning trajectories; interpreted as adaptive kernel density estimation.
DEPfold: RNA Secondary Structure Prediction as Dependency Parsing.
KE WANG, Shay B Cohen
Protein Structure PredictionTransformerLarge Language ModelBiomedical Data
🎯 What it does: A framework called DEPfold is proposed to transform RNA secondary structure prediction into a dependency parsing-based approach.
DEPT: Decoupled Embeddings for Pre-training Language Models
Alex Iacob (University of Cambridge), Nicholas Donald Lane
OptimizationFederated LearningComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: A communication-efficient pre-training framework called DEPT is proposed, which decouples word embeddings from the Transformer architecture and supports training on multi-source heterogeneous corpora.