ICML 2025 Papers — Page 9
International Conference on Machine Learning · 3257 papers
Do We Need to Verify Step by Step? Rethinking Process Supervision from a Theoretical Perspective
Zeyu Jia (Massachusetts Institute of Technology), Tengyang Xie (University of Wisconsin-Madison)
Reinforcement Learning
🎯 What it does: This paper proposes a method to transform result supervision data that only contains total rewards into process supervision data with stepwise rewards, and provides a theoretical proof showing that under the condition of state-action coverage, the statistical complexity difference between the two supervision methods is only polynomial in nature; it also presents a new 'trajectory measure transformation lemma' and proves that the advantage function can serve as the optimal process reward model, indicating that using the Q function may lead to suboptimal results; an improved sample complexity analysis of preference supervision and the DPO algorithm is also provided.
Do We Really Need Message Passing in Brain Network Modeling?
Liang Yang (Hebei University of Technology), Xiaochun Cao
Graph Neural NetworkTransformerBiomedical DataMagnetic Resonance ImagingAlzheimer's Disease
🎯 What it does: This paper proposes the Brain Quadratic Network (BQN), which replaces traditional message passing with quadratic networks and Hadamard multiplication for brain network modeling.
DocKS-RAG: Optimizing Document-Level Relation Extraction through LLM-Enhanced Hybrid Prompt Tuning
Xiaolong Xu (Nanjing University of Information Science and Technology), Wanchun Dou (Nanjing University)
RetrievalOptimizationGraph Neural NetworkLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes a new framework called DocKS-RAG to enhance document-level relation extraction (RE) performance.
DocVXQA: Context-Aware Visual Explanations for Document Question Answering
Mohamed Ali Souibgui (Computer Vision Center Universitat Aut` onoma de Barcelona), Dimosthenis Karatzas (Computer Vision Center Universitat Aut` onoma de Barcelona)
Explainability and InterpretabilityVision Language ModelText
🎯 What it does: This paper proposes DocVXQA, a model capable of self-explanation when answering visual questions about documents, which can generate contextually relevant visual explanation heatmaps.
Does Data Scaling Lead to Visual Compositional Generalization?
Arnas Uselis (University of Tübingen), Seong Joon Oh (University of Tübingen)
Representation LearningConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper systematically studies the combinatorial generalization ability of visual models under different data scales and diversities, and proposes a controllable (n,k) experimental framework.
Does Generation Require Memorization? Creative Diffusion Models using Ambient Diffusion
Kulin Shah (University of Texas at Austin), Giannis Daras (Massachusetts Institute of Technology)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: A framework is proposed for training diffusion models using Ambient Diffusion during high noise stages, significantly reducing the model's memorization of training samples while maintaining generation quality.
Does Graph Prompt Work? A Data Operation Perspective with Theoretical Analysis
Qunzhong Wang (Chinese University of Hong Kong), Hong Cheng (Chinese University of Hong Kong)
Graph Neural NetworkPrompt EngineeringGraph
🎯 What it does: This paper conducts a systematic theoretical analysis of Graph Prompt from the perspective of data manipulation, proposing the concepts of bridge set and ε-extended bridge set. It provides a guarantee theorem that Graph Prompt can approximate graph data transformation operators and derives the error upper bounds and error distributions for single graphs and batch graphs.
Does learning the right latent variables necessarily improve in-context learning?
Sarthak Mittal (Mila Quebec AI Institute), Dhanya Sridhar (Google DeepMind)
Transformer
🎯 What it does: This study incorporates a bottleneck structure into the Transformer to force the model to explicitly infer the latent variables of the task from the context, in order to examine whether explicit latent variable inference can enhance the generalization ability of ICL (in-context learning).
Does Low Rank Adaptation Lead to Lower Robustness against Training-Time Attacks?
Zi Liang (Hong Kong Polytechnic University), RongHua Li
OptimizationAdversarial AttackTransformerSupervised Fine-TuningText
🎯 What it does: Study the robustness of LoRA under training-time attacks (data poisoning and backdoor attacks);
Does One-shot Give the Best Shot? Mitigating Model Inconsistency in One-shot Federated Learning
Hui Zeng (National University of Defense Technology), Zhiping Cai (National University of Defense Technology)
Federated LearningContrastive LearningImage
🎯 What it does: Proposes the FAFI framework to address the issue of local model inconsistency in one-round federated learning.
DOLPHIN: A Programmable Framework for Scalable Neurosymbolic Learning
Aaditya Naik (University of Pennsylvania), Eric Wong (University of Pennsylvania)
Computational EfficiencyImageVideoText
🎯 What it does: This paper proposes and implements DOLPHIN, a programmable and scalable neuro-symbolic learning framework.
Domain-Adapted Diffusion Model for PROTAC Linker Design Through the Lens of Density Ratio in Chemical Space
Zixing Song (University of Cambridge), José Miguel Hernández-Lobato (University of Cambridge)
Domain AdaptationDrug DiscoveryGraph Neural NetworkDiffusion modelTabular
🎯 What it does: A domain-adaptive diffusion model for PROTAC linker design, DAD-PROTAC, is proposed to address the mismatch between small molecules and the chemical space distribution of PROTACs.
Domain2Vec: Vectorizing Datasets to Find the Optimal Data Mixture without Training
Mozhi Zhang (Fudan University), Xipeng Qiu (Fudan University)
Computational EfficiencyData-Centric LearningLarge Language ModelText
🎯 What it does: The DOMAIN2VEC method is proposed, treating any dataset as a linear combination of several meta-domains. By training an unsupervised meta-domain classifier, the domain vector of the dataset is obtained, and the optimal data mixing ratio is predicted without training using the distribution alignment hypothesis or RegMix.
Don't Restart, Just Reuse: Reoptimizing MILPs with Dynamic Parameters
Sijia Zhang (University of Science and Technology of China), Xiangyang Li
OptimizationGraph Neural NetworkReinforcement LearningTabular
🎯 What it does: This paper proposes a two-stage re-optimization framework VP-OR, which quickly obtains high-quality feasible solutions for dynamic parameter MILP problems.
Double Machine Learning for Causal Inference under Shared-State Interference
Chris Hays (Massachusetts Institute of Technology), Manish Raghavan (Massachusetts Institute of Technology)
🎯 What it does: In the framework of causal inference with shared state interference, a Double Machine Learning (DML) method is proposed for efficient estimation of Average Direct Effects (ADE) and Global Average Treatment Effects (GATE);
Double-Filter: Efficient Fine-tuning of Pre-trained Vision-Language Models via Patch&Layer Filtering
Yaoqin He (University of Glasgow), Xuri Ge (Shandong University)
Object DetectionRetrievalComputational EfficiencyTransformerSupervised Fine-TuningVision Language ModelImageMultimodality
🎯 What it does: Proposes the Double-Filter method, which utilizes image Patch filtering and network-level fine-grained filtering to achieve efficient fine-tuning of Vision-Language Pre-trained (VLP) models.
Doubly Protected Estimation for Survival Outcomes Utilizing External Controls for Randomized Clinical Trials
Chenyin Gao (North Carolina State University), Douglas Faries (Eli Lilly & Company)
Biomedical Data
🎯 What it does: This paper proposes a dual protection estimation framework that uses external control data to supplement the control group of randomized clinical trials to assess differences in survival outcomes (RMST).
Doubly Robust Conformalized Survival Analysis with Right-Censored Data
Matteo Sesia (University of Southern California), Vladimir Svetnik (Merck and Co Inc)
TabularBiomedical Data
🎯 What it does: A dual robust conformal survival analysis method for right-censored data is designed, which can construct lower prediction bounds (LPB) for survival time and provide theoretical guarantees.
Doubly Robust Fusion of Many Treatments for Policy Learning
Ke Zhu (North Carolina State University), Shu Yang (North Carolina State University)
TabularBiomedical DataElectronic Health Records
🎯 What it does: A treatment fusion method based on calibrated weights is proposed, which first balances covariates for multi-level treatments and then uses weighted fusion Lasso to cluster similar treatments. Subsequently, existing individualized treatment rule (ITR) learning methods (such as decision trees) are employed in the aggregated treatment space to obtain optimal decisions.
DPCore: Dynamic Prompt Coreset for Continual Test-Time Adaptation
Yunbei Zhang (Tulane University), Jihun Hamm (Tulane University)
Domain AdaptationPrompt EngineeringImage
🎯 What it does: The DPCore method is proposed to achieve Continuous Testing Time-domain Adaptation (CTTA), efficiently adapting to the continuously changing target domain through visual prompts, a prompt core set, and dynamic updates.
DPO Meets PPO: Reinforced Token Optimization for RLHF
Han Zhong (Peking University), Liwei Wang (Peking University)
OptimizationReinforcement Learning from Human FeedbackReinforcement LearningText
🎯 What it does: Re-modeling Reinforcement Learning with Human Feedback (RLHF) from the traditional sentence-level Bandit framework to a fine-grained token-level Markov Decision Process (MDP), and proposing the Reinforced Token Optimization (RTO) algorithm: first, using DPO to learn token-level rewards from offline preference data, and then applying PPO for reinforcement learning on that reward;
DRAG: Data Reconstruction Attack using Guided Diffusion
Wa-Kin Lei (National Taiwan University), Shang-Tse Chen (National Taiwan University)
Data SynthesisAdversarial AttackTransformerDiffusion modelImage
🎯 What it does: In the Split Inference scenario, data reconstruction attacks are conducted on the intermediate representations of large visual foundation models (such as CLIP-ViT and DINOv2).
DragLoRA: Online Optimization of LoRA Adapters for Drag-based Image Editing in Diffusion Model
Siwei Xia (Shanghai Key Laboratory of Multidimensional Information Processing), Qingli Li (Key Laboratory of Advanced Theory and Application in Statistics and Data Science)
GenerationOptimizationTransformerDiffusion modelImage
🎯 What it does: Implement drag-and-drop image editing on pre-trained diffusion models using the online optimization LoRA adapter framework called DragLoRA.
DragSolver: A Multi-Scale Transformer for Real-World Automotive Drag Coefficient Estimation
Ye Liu (Shanghai Jiao Tong University), Yuntian Chen (Eastern Institute of Technology)
Autonomous DrivingTransformerMesh
🎯 What it does: A multi-scale network based on Transformer is implemented to quickly and accurately estimate the drag coefficient from 3D vehicle models.
DreamDPO: Aligning Text-to-3D Generation with Human Preferences via Direct Preference Optimization
Zhenglin Zhou (Zhejiang University), Tat-Seng Chua (National University of Singapore)
GenerationData SynthesisOptimizationLarge Language ModelReinforcement LearningNeural Radiance FieldTextPoint Cloud
🎯 What it does: Achieving text-to-3D content generation and alignment with human preferences through Direct Preference Optimization (DreamDPO), constructing online comparative samples and evaluating preferences using a reward model or large multimodal models, thereby updating 3D representations;
DriveGPT: Scaling Autoregressive Behavior Models for Driving
Xin Huang (Cruise LLC), Siddhartha Srinivasa
Autonomous DrivingTransformerSupervised Fine-TuningSequential
🎯 What it does: Developed DriveGPT, a Transformer-based autoregressive behavior model for predicting and planning the future trajectories of autonomous vehicles in complex road scenarios.
Drug-TTA: Test-Time Adaptation for Drug Virtual Screening via Multi-task Meta-Auxiliary Learning
Ao Shen (Fudan University), Manning Wang (Fudan University)
Meta LearningDrug DiscoveryContrastive LearningBiomedical Data
🎯 What it does: This paper proposes a drug virtual screening framework called Drug-TTA based on Test-Time Adaptation (TTA), which enhances the recognition ability of unknown protein-ligand pairs by dynamically adjusting model parameters during the inference phase using self-supervised auxiliary tasks.
DS-VLM: Diffusion Supervision Vision Language Model
Zhen Sun (Xiamen University), Rongrong Ji (Xiamen University)
TransformerMixture of ExpertsVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: DS-VLM constructs a shorter knowledge propagation chain and improves the quality of visual features by introducing a diffusion model during the training phase, providing pixel-level reconstruction supervision for the visual encoder and connector.
DSBRouter: End-to-end Global Routing via Diffusion Schr\"{o}dinger Bridge
Liangliang Shi (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
OptimizationDiffusion modelScore-based ModelGraphStochastic Differential Equation
🎯 What it does: An end-to-end global routing model called DSBRouter is proposed, which can directly generate connected and low-overflow routing results.
DSP: Dynamic Sequence Parallelism for Multi-Dimensional Transformers
Xuanlei Zhao (National University of Singapore), Yang You (National University of Singapore)
TransformerSequential
🎯 What it does: This paper proposes Dynamic Sequence Parallelism (DSP), a parallel method that can dynamically switch sequence dimensions during multi-dimensional Transformer computations, significantly reducing communication overhead and improving throughput.
DTZO: Distributed Trilevel Zeroth Order Learning with Provable Non-Asymptotic Convergence
Yang Jiao (Tongji University), Chengtao Jian (Tongji University)
OptimizationFederated LearningHyperparameter SearchLarge Language ModelTabular
🎯 What it does: A distributed three-layer zero-order learning framework (DTZO) is proposed to solve three-layer nested optimization problems without gradient information, along with a non-asymptotic convergence analysis.
Dual Feature Reduction for the Sparse-group Lasso and its Adaptive Variant
Fabio Feser (Imperial College London), Marina Evangelou (Imperial College London)
OptimizationComputational EfficiencyBiomedical Data
🎯 What it does: This paper proposes a Dual Feature Reduction (DFR) method, which first masks ineffective groups and then removes ineffective variables within groups when solving Sparse Group Lasso (SGL) and its adaptive variant aSGL, significantly reducing the optimization space without affecting the optimal solution.
Dueling Convex Optimization with General Preferences
Aadirupa Saha (University of Illinois), Yishay Mansour (Tel Aviv University)
Optimization
🎯 What it does: A theoretical optimization framework and algorithm are proposed under the environment of only two-point comparison feedback (dueling convex optimization under general transfer functions).
DUNIA: Pixel-Sized Embeddings via Cross-Modal Alignment for Earth Observation Applications
Ibrahim Fayad (Laboratoire des Sciences du Climat et de l'Environnement), Alexandre d'Aspremont (Computer Science)
SegmentationGenerationRetrievalTransformerDiffusion modelContrastive LearningImageMultimodalityPoint CloudAgriculture Related
🎯 What it does: Learn pixel-level embeddings and achieve multi-task zero-shot and few-shot earth observation predictions through cross-modal alignment;
DVI:A Derivative-based Vision Network for INR
Runzhao Yang (Tsinghua University), Jinli Suo (Tsinghua University)
Super ResolutionConvolutional Neural NetworkTransformerImageVideoMultimodality
🎯 What it does: This paper proposes the DVI network, which integrates high-order derivative information from Implicit Neural Representation (INR) into existing raster visual networks to achieve unified processing of various visual tasks.
DyCodeEval: Dynamic Benchmarking of Reasoning Capabilities in Code Large Language Models Under Data Contamination
Simin Chen (Columbia University), Baishakhi Ray (Columbia University)
AI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: This paper proposes DyCodeEval, a dynamic evaluation framework based on the idea of meta-testing, which utilizes a four-stage LLM agent to automatically generate programming problems that are semantically equivalent to the original questions but have diverse contexts, thereby assessing the reasoning ability of Code LLMs under data contamination.
Dynamic Mixture of Curriculum LoRA Experts for Continual Multimodal Instruction Tuning
Chendi Ge (Tsinghua University), Wenwu Zhu (Tsinghua University)
OptimizationTransformerLarge Language ModelMixture of ExpertsAuto EncoderImageTextMultimodalityBenchmark
🎯 What it does: This paper proposes a method for dynamically evolving architecture to automatically allocate LoRA experts in Continuous Multi-modal Instruction Tuning (CMIT) of large multi-modal language models (MLLMs) to adapt to the hierarchical and modal requirements of different tasks.
Dynamic Similarity Graph Construction with Kernel Density Estimation
Steinar Laenen (University of Edinburgh), He Sun (University of Edinburgh)
Computational EfficiencyImageTabular
🎯 What it does: An algorithm for dynamically maintaining kernel density estimation (KDE) and similarity graphs is proposed, and based on this, dynamic spectral clustering is implemented.
Dynamic Sparse Training of Diagonally Sparse Networks
Abhishek Tyagi (University of Rochester), Yuhao Zhu (University of Rochester)
OptimizationComputational EfficiencyTransformerImageText
🎯 What it does: A dynamic sparse training method DynaDiag based on diagonal sparse patterns is proposed, which can maintain the sparse structure during training and achieve efficient inference and training.
Dynamical Modeling of Behaviorally Relevant Spatiotemporal Patterns in Neural Imaging Data
Sayed Mohammad Hosseini (University of Southern California), Maryam Shanechi
Convolutional Neural NetworkRecurrent Neural NetworkTransformerImageBiomedical DataUltrasound
🎯 What it does: A deep learning framework called SBIND is proposed and implemented, which performs spatiotemporal neural dynamic modeling directly on raw panoramic/functional ultrasound imaging data, separating behavior-related dynamics from other neural dynamics.
Dynamical phases of short-term memory mechanisms in RNNs
Bariscan Kurtkaya (Koc University), Nina Miolane (Fudan University)
Recurrent Neural NetworkSequential
🎯 What it does: This study explores two main sequence generation mechanisms in short-term memory tasks—slow-point manifold and limit cycle—through large-scale training of RNNs, revealing their relationships with task design, learning rate, and delay duration.
DynaMind: Reasoning over Abstract Video Dynamics for Embodied Decision-Making
Ziru Wang (State Grid Corporation of China), Jingdong Wang (Baidu)
Robotic IntelligenceConvolutional Neural NetworkTransformerVision-Language-Action ModelVideoTextMultimodality
🎯 What it does: The DynaMind framework is proposed, which abstracts videos into high-level dynamic representations, utilizing dynamic reasoning and a dynamic guidance decision module to achieve the integration of natural language instructions and visual content.
DyPolySeg: Taylor Series-Inspired Dynamic Polynomial Fitting Network for Few-shot Point Cloud Semantic Segmentation
Changshuo Wang (Nanyang Technological University), Prayag Tiwari (Halmstad University)
SegmentationConvolutional Neural NetworkPoint Cloud
🎯 What it does: This paper proposes DyPolySeg, a pre-trained, dependency-free dynamic convolutional network based on multi-order polynomial fitting for few-shot point cloud semantic segmentation.
e-GAI: e-value-based Generalized $\alpha$-Investing for Online False Discovery Rate Control
Yifan Zhang (Shanghai Jiao Tong University), Changliang Zou (Nankai University)
Anomaly DetectionTime SeriesFinance Related
🎯 What it does: This paper proposes an online hypothesis testing framework based on e-values, called e-GAI, and designs two algorithms, e-LORD and e-SAFFRON, which can achieve online FDR control through e-values under any dependency structure. It also provides long-term improvement plans for the α-death problem with mem-e-LORD and mem-e-SAFFRON; additionally, the framework is extended to p-values that satisfy conditional super-uniformity. Its effectiveness is validated through theoretical proofs and simulations/real data experiments.
E-LDA: Toward Interpretable LDA Topic Models with Strong Guarantees in Logarithmic Parallel Time
Adam Breuer (Dartmouth College)
OptimizationExplainability and InterpretabilityComputational EfficiencyText
🎯 What it does: This paper proposes a combination optimization-based LDA topic allocation algorithm E-LDA, which can approximately optimally solve the topic-word allocation problem under the constraint of document sparsity.
EAGLES: Towards Effective, Efficient, and Economical Federated Graph Learning via Unified Sparsification
Zitong Shi (Wuhan University), Mang Ye (Wuhan University)
Federated LearningComputational EfficiencyGraph Neural NetworkMixture of ExpertsGraph
🎯 What it does: Proposes the EAGLES unified sparse framework, which balances graph structure and model parameter sparsity, achieving efficient, economical, and effective federated graph learning.
EARL-BO: Reinforcement Learning for Multi-Step Lookahead, High-Dimensional Bayesian Optimization
Mujin Cheon (Korea Advanced Institute of Science and Technology), Calvin Tsay (Imperial College London)
OptimizationHyperparameter SearchReinforcement LearningTabular
🎯 What it does: The EARL-BO method is proposed, which combines the Attention-DeepSets encoder with PPO-based reinforcement learning to achieve multi-step foresight Bayesian Optimization, addressing the challenges of high-dimensional black-box optimization.
Earley-Driven Dynamic Pruning for Efficient Structured Decoding
Xintong Sun (Rice University), Shiwen Ni (Shenzhen Institutes of Advanced Technology)
GenerationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: A constraint decoding engine named Formatron has been constructed, utilizing CFG and the Earley parser to achieve structured generation in large language models, while dynamically pruning invalid states during the inference process.
EARTH: Epidemiology-Aware Neural ODE with Continuous Disease Transmission Graph
Guancheng Wan (Emory University), Wei Jin (Emory University)
Graph Neural NetworkTime SeriesOrdinary Differential Equation
🎯 What it does: A full-process model called EARTH based on Neural Ordinary Differential Equations (Neural ODE) is proposed for predicting infectious disease transmission in continuous time.
EasyInv: Toward Fast and Better DDIM Inversion
Ziyue Zhang (Xiamen University), Rongrong Ji (Xiamen University)
RestorationGenerationComputational EfficiencyDiffusion modelImage
🎯 What it does: A DDIM inversion method named EasyInv is proposed, which utilizes a weighted fusion of the initial latent states to eliminate noise errors in traditional iterative optimization, achieving faster and more stable image inversion.
EasyRef: Omni-Generalized Group Image Reference for Diffusion Models via Multimodal LLM
Zhuofan Zong (Chinese University of Hong Kong), Hongsheng Li (Chinese University of Hong Kong)
GenerationData SynthesisTransformerLarge Language ModelDiffusion modelImageMultimodalityBenchmark
🎯 What it does: EasyRef is proposed, a plugin-based adaptation method using a multimodal large language model for personalized generation of diffusion models controlled by instructions under multiple reference images.
EcoMapper: Generative Modeling for Climate-Aware Satellite Imagery
Muhammed Goktepe (Technical University of Munich), Malte von Bloh (Technical University of Munich)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: Proposed the EcoMapper framework, which combines climate data with Sentinel-2 satellite images to train diffusion models for generating synthetic remote sensing images under various climate conditions.
Edge-Colored Clustering in Hypergraphs: Beyond Minimizing Unsatisfied Edges
Alex Crane (University of Utah), Nate Veldt (Texas A&M University)
OptimizationGraph
🎯 What it does: A clustering framework for edge-colored hypergraphs is proposed, and research is conducted from various objectives such as minimizing unsatisfied edges and maximizing satisfied edges.
Editable Concept Bottleneck Models
Lijie Hu (King Abdullah University of Science and Technology), Di Wang (King Abdullah University of Science and Technology)
Computational EfficiencyBiomedical Data
🎯 What it does: This paper proposes Editable Concept Bottleneck Models (ECBMs), which allow for rapid modifications of training data, concept labels, and concept sets without the need for retraining from scratch.
Editable Noise Map Inversion: Encoding Target-image into Noise For High-Fidelity Image Manipulation
Mingyu Kang (Hanyang University), Yong Suk Choi (Hanyang University)
Image TranslationGenerationDiffusion modelImageVideo
🎯 What it does: This paper proposes an Editable Noise Map Inversion (ENM Inversion) method that encodes target image information into a noise vector to achieve high-fidelity text-guided image and video editing.
EditLord: Learning Code Transformation Rules for Code Editing
Weichen Li (University of Chicago), Kexin Pei (University of Chicago)
OptimizationAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: A code editing framework named EDITLORD is proposed, which utilizes language models to extract explicit editing rules and functional specifications from training pairs to enhance code editing effectiveness.
EduLLM: Leveraging Large Language Models and Framelet-Based Signed Hypergraph Neural Networks for Student Performance Prediction
Ming Li (Zhejiang Normal University), Pietro Lio (University of Cambridge)
Graph Neural NetworkLarge Language ModelTabular
🎯 What it does: The EduLLM framework is proposed, which combines large language models with signature hypergraph learning to predict student performance.
EEG-Language Pretraining for Highly Label-Efficient Clinical Phenotyping
Sam Gijsen (Charite University Medicine Berlin), Kerstin Ritter (Hertie Institute for AI in Brain Health)
ClassificationRetrievalConvolutional Neural NetworkContrastive LearningMultimodalityTime SeriesBiomedical DataElectronic Health Records
🎯 What it does: This study proposes a multimodal pre-training model (ELM) based on clinical reports and EEG signals, achieving sub-unit level cross-modal alignment through the trimming of EEG time series, segmentation of reports, and the integration of multi-instance learning (MIL).
EFDTR: Learnable Elliptical Fourier Descriptor Transformer for Instance Segmentation
Jiawei Cao (Shanghai Jiao Tong University), Changsheng Lu (Australian National University)
Object DetectionSegmentationTransformerImage
🎯 What it does: A learnable Elliptical Fourier Descriptor Transformer (EFDTR) framework is proposed, which directly predicts the polygon boundaries of instances through a two-stage decoding process.
Effective and Efficient Masked Image Generation Models
Zebin You (Renmin University of China), Chongxuan Li (Renmin University of China)
GenerationData SynthesisComputational EfficiencyTransformerDiffusion modelImage
🎯 What it does: A unified framework is proposed that integrates mask image generation models with mask diffusion models, resulting in the design of an efficient eMIGM model, which enhances sampling efficiency and generation quality through various techniques.
Efficient and Privacy-Preserving Soft Prompt Transfer for LLMs
Xun Wang (CISPA Helmholtz Center for Information Security), Adam Dziedzic (CISPA Helmholtz Center for Information Security)
Safty and PrivacyComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes the POST framework, which can transfer soft prompts obtained from fine-tuning on small models to large LLMs while ensuring user data privacy.
Efficient and Scalable Density Functional Theory Hamiltonian Prediction through Adaptive Sparsity
Erpai Luo (Tsinghua University), Jia Zhang (Microsoft Research)
Computational EfficiencyDrug DiscoveryGraph Neural NetworkGraphPhysics Related
🎯 What it does: This paper presents SPHNet, an adaptive sparse Hamiltonian prediction model introduced in SE(3) equivariant networks.
Efficient and Separate Authentication Image Steganography Network
Junchao Zhou (Harbin Institute of Technology), Guangming Lu (Harbin Institute of Technology)
Image TranslationData SynthesisSafty and PrivacyComputational EfficiencyFlow-based ModelImage
🎯 What it does: An efficient and independently authenticable image steganography network (AIS) is proposed, which achieves multi-image steganography and decryption through a two-stage reversible network and supports independent authentication.
Efficient ANN-SNN Conversion with Error Compensation Learning
Chang Liu (Dalian University of Technology), Gang Pan (Zhejiang University)
ClassificationComputational EfficiencySpiking Neural NetworkSupervised Fine-TuningImage
🎯 What it does: This paper proposes an ANN-to-SNN conversion framework based on error compensation learning, addressing issues of pruning, quantization, and non-uniform errors.
Efficient Bisection Projection to Ensure Neural-Network Solution Feasibility for Optimization over General Set
Enming Liang (City University of Hong Kong), Minghua Chen (Chinese University of Hong Kong)
OptimizationTabular
🎯 What it does: A Bisection Projection framework is proposed, which uses interior points and bisection methods to project infeasible solutions predicted by neural networks into a general compact constraint set, ensuring feasibility while maintaining a small optimality loss.
Efficient Core-set Selection for Deep Learning Through Squared Loss Minimization
Jianting Chen (Tongji University)
OptimizationData-Centric LearningConvolutional Neural NetworkImage
🎯 What it does: Selects the core set by minimizing the squared loss of all samples and based on loss reduction attribution;
Efficient Curvature-Aware Hypergradient Approximation for Bilevel Optimization
Youran Dong (Nanjing University), Jin Zhang (Southern University of Science and Technology)
OptimizationHyperparameter SearchMeta LearningImageTabular
🎯 What it does: An efficient curvature-aware supergradient approximation method is proposed for bilevel optimization problems, particularly in hyperparameter optimization and meta-learning.
Efficient Diffusion Models for Symmetric Manifolds
Oren Mangoubi (Worcester Polytechnic Institute), Nisheeth K. Vishnoi (Yale University)
GenerationData SynthesisComputational EfficiencyDiffusion modelMultimodalityPhysics RelatedStochastic Differential Equation
🎯 What it does: An efficient diffusion model for symmetric space Riemannian manifolds (such as tori, spheres, special orthogonal groups SO(n), and unitary groups U(n)) is proposed, utilizing variable covariance forward diffusion and achieving training and sampling through the projection of Euclidean Brownian motion.
Efficient Distributed Optimization under Heavy-Tailed Noise
Su Hyeong Lee (University of Chicago), Tian Li (University of Chicago)
OptimizationText
🎯 What it does: This paper proposes the TailOPT framework and the BiClip optimizer to address heavy-tailed noise in distributed training.
Efficient Federated Incomplete Multi-View Clustering
Suyuan Liu (National University of Defence Technology), Xinwang Liu (National University of Defence Technology)
Federated LearningSafty and PrivacyComputational EfficiencyMultimodality
🎯 What it does: A federated incomplete multi-view clustering framework EFIMVC is proposed, which can handle missing views, has high communication efficiency, and maintains privacy.
Efficient Fine-Grained Guidance for Diffusion Model Based Symbolic Music Generation
Tingyu Zhu (University of California), Zeyu Zheng (University of California)
GenerationDiffusion modelAudio
🎯 What it does: This paper proposes the Fine-Grained Guidance (FGG) method, which integrates fine-grained chord and rhythm constraints into the training and sampling of diffusion models, achieving high-precision generation of symbolic music.
Efficient First-Order Optimization on the Pareto Set for Multi-Objective Learning under Preference Guidance
Lisha Chen (Rensselaer Polytechnic Institute), Tianyi Chen (Peking University)
OptimizationImageAudio
🎯 What it does: A first-order penalty function method based on a smooth superiority function is proposed to solve the multi-objective learning problem (OPS) under user-specified preferences, with theoretical convergence guarantees provided.
Efficient Generative Modeling with Residual Vector Quantization-Based Tokens
Jaehyeon Kim (NVIDIA), Jaewoong Cho (KRAFTON)
GenerationData SynthesisDiffusion modelImageAudio
🎯 What it does: The ResGen model is designed to achieve high-fidelity generation using a discrete diffusion framework with Residual Vector Quantization (RVQ), while maintaining fast sampling.
Efficient Graph Continual Learning via Lightweight Graph Neural Tangent Kernels-based Dataset Distillation
Rihong Qiu (Peking University), Yasha Wang (Peking University)
Computational EfficiencyKnowledge DistillationGraph Neural NetworkGraph
🎯 What it does: A lightweight Graph Neural Tangent Kernel (LIGHTGNTK) framework is proposed for efficient distillation of graph datasets, enabling continual learning on graphs.
Efficient Heterogeneity-Aware Federated Active Data Selection
Ying-Peng Tang (Nanyang Technological University), Han Yu (Nanyang Technological University)
Federated LearningSafty and PrivacyComputational EfficiencyScore-based ModelImageTabular
🎯 What it does: A federated active learning method named FALE is designed, which securely obtains global feature information using FedSVD, combined with score sampling to achieve one-time global data selection and global regression model training;
Efficient Length-Generalizable Attention via Causal Retrieval for Long-Context Language Modeling
Xiang Hu (Ant Group), Kewei Tu (ShanghaiTech University)
GenerationRetrievalOptimizationComputational EfficiencyTransformerTextRetrieval-Augmented Generation
🎯 What it does: A long-context attention mechanism called GCA is proposed, which supports pre-training and is used to construct a retrievable Transformer (DRT), achieving a combination of causal retrieval and long text generation.
Efficient LiDAR Reflectance Compression via Scanning Serialization
Jiahao Zhu (Hangzhou Normal University), Zhan Ma (Nanjing University)
CompressionAutonomous DrivingPoint Cloud
🎯 What it does: A lossless compression framework for LiDAR reflectance based on scan serialization, called SerLiC, is proposed, which converts 3D point clouds into 1D sequences and uses the Mamba model for autoregressive encoding.
Efficient Logit-based Knowledge Distillation of Deep Spiking Neural Networks for Full-Range Timestep Deployment
Chengting Yu (Zhejiang University), Aili Wang (Zhejiang University)
Computational EfficiencyKnowledge DistillationSpiking Neural NetworkSupervised Fine-TuningImage
🎯 What it does: A time-decomposed logits knowledge distillation framework is proposed to train deep SNNs, enabling them to maintain high accuracy at any time step without the need for retraining.
Efficient Long Context Fine-tuning with Chunk Flow
Xiulong Yuan (Alibaba Group), Wei Lin (State Key Lab of Processors, Institute of Computing Technology, Chinese Academy of Sciences)
Computational EfficiencyTransformerSupervised Fine-TuningText
🎯 What it does: A chunk-based long context fine-tuning method called ChunkFlow is proposed, which can merge short sequences and split long sequences into uniformly sized chunks for training.
Efficient Molecular Conformer Generation with SO(3)-Averaged Flow Matching and Reflow
Zhonglin Cao (NVIDIA), Emine Kucukbenli (NVIDIA)
GenerationComputational EfficiencyKnowledge DistillationDrug DiscoveryGraph Neural NetworkTransformerFlow-based ModelRectified FlowGraphOrdinary Differential Equation
🎯 What it does: Proposes an SO(3) average flow matching and reflow + distillation method, significantly accelerating molecular conformation generation.
Efficient Motion Prompt Learning for Robust Visual Tracking
Jie Zhao (Dalian University of Technology), Huchuan Lu (Dalian University of Technology)
Object TrackingTransformerPrompt EngineeringVideo
🎯 What it does: A lightweight pluggable Motion Prompt Tracking module (MPT) is proposed, which can encode historical trajectories into visual embeddings and dynamically fuse with a baseline visual tracker to enhance robustness in complex scenes.
Efficient Multi-modal Long Context Learning for Training-free Adaptation
Zehong Ma (Peking University), Qi Tian (Huawei Inc)
CompressionComputational EfficiencyTransformerLarge Language ModelMultimodality
🎯 What it does: A training-independent multimodal long-context learning framework EMLoC is proposed, which compresses long contexts into compact memory using chunk compression and hierarchical adaptive pruning.
Efficient Multivariate Robust Mean Estimation Under Mean-Shift Contamination
Ilias Diakonikolas (University of Wisconsin-Madison), Thanasis Pittas (University of Wisconsin-Madison)
Anomaly DetectionComputational EfficiencyTabular
🎯 What it does: This study investigates the problem of high-dimensional robust mean estimation under mean shift contamination and proposes a computationally efficient algorithm that can tolerate a certain proportion of outliers.
Efficient Network Automatic Relevance Determination
Hongwei Zhang (Fudan University), Yuan Qi (Fudan University)
Explainability and InterpretabilityComputational EfficiencyTabularBiomedical DataFinance Related
🎯 What it does: A sparse regression framework for multi-output linear probability models, called NARD (Network Automatic Relevance Determination), is proposed. Based on this framework, three efficient variants (Sequential NARD, Surrogate NARD, Hybrid NARD) are designed to significantly reduce computational complexity while maintaining sparsity and modeling output correlation.
Efficient Noise Calculation in Deep Learning-based MRI Reconstructions
Onat Dalmaz (Stanford University), Brian Hargreaves (Stanford University)
RestorationComputational EfficiencyBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A method for unbiased estimation based on the Jacobian matrix sketch is proposed for efficiently calculating voxel-level noise variance in deep learning MRI reconstruction.
Efficient Online Reinforcement Learning for Diffusion Policy
Haitong Ma (Harvard University), Bo Dai (Georgia Institute of Technology)
Reinforcement LearningDiffusion modelScore-based ModelSequential
🎯 What it does: A Reweighted Score Matching (RSM) framework is proposed for training diffusion policies in online reinforcement learning, along with the implementation of two efficient algorithms—Diffusion Policy Mirror Descent (DPMD) and Soft Diffusion Actor-Critic (SDAC);
Efficient Optimization with Orthogonality Constraint: a Randomized Riemannian Submanifold Method
Andi Han (University of Sydney), Akiko Takeda (University of Tokyo)
OptimizationTransformerImage
🎯 What it does: A Random Submanifold Descent (RSDM) algorithm is proposed, which significantly reduces the computational complexity of Retraction in orthogonal constrained optimization by updating variables on a low-dimensional random submanifold.
Efficient Parallel Training Methods for Spiking Neural Networks with Constant Time Complexity
Wanjin Feng (Institute of Microelectronics Chinese Academy of Sciences), Chunyan Miao (Nanyang Technological University)
Spiking Neural NetworkImageSequential
🎯 What it does: Proposes a Fixed Point Parallel Training (FPT) method, which reduces the time complexity of SNN from O(T) to O(K) through fixed point iteration, achieving parallel training;
Efficient Personalized Adaptation for Physiological Signal Foundation Model
Chenrui Wu (Zhejiang University), Jiajun Bu (Zhejiang University)
ClassificationRecognitionData SynthesisComputational EfficiencyTransformerDiffusion modelTime SeriesBiomedical DataElectrocardiogram
🎯 What it does: This paper proposes the PhysioPFM framework, which achieves efficient personalized adaptation of a time series foundation model (TSFM) based on pre-training with public physiological signals on local medical data.
Efficient Quantification of Multimodal Interaction at Sample Level
Zequn Yang (Renmin University of China), Di Hu (Renmin University of China)
Computational EfficiencyKnowledge DistillationVideoMultimodalityAudio
🎯 What it does: A lightweight sample-level multimodal interaction estimator (LSMI) is proposed, capable of quantifying redundancy, uniqueness, and synergistic information for each sample.
Efficient Robotic Policy Learning via Latent Space Backward Planning
Dongxiu Liu (Tsinghua University), Xianyuan Zhan (Tsinghua University)
Robotic IntelligenceConvolutional Neural NetworkReinforcement LearningSequential
🎯 What it does: A latent space-based inverse planning framework (LBP) is proposed, which first predicts the final latent goal and then recursively generates intermediate sub-goals that are closer to the current state, using this context to train low-level control policies.
Efficient Robust Conformal Prediction via Lipschitz-Bounded Networks
Thomas Massena (IRIT), Sébastien Gerchinovitz (IRT Saint Exupery)
ClassificationAnomaly DetectionComputational EfficiencyConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: The lip-rcp method is proposed, which efficiently computes robust conformal prediction sets using Lipschitz-constrained networks and audits the coverage of Vanilla CP under any attack budget.
Efficient Skill Discovery via Regret-Aware Optimization
He Zhang (Hong Kong University of Science and Technology), Hui Xiong (Hong Kong University of Science and Technology)
OptimizationRobotic IntelligenceReinforcement Learning
🎯 What it does: A framework called 'RSD' based on minimizing and maximizing adversarial processes is proposed to efficiently discover diverse skills under unsupervised conditions.
Efficient Source-free Unlearning via Energy-Guided Data Synthesis and Discrimination-Aware Multitask Optimization
Xiuyuan Wang (Zhejiang University), Xiaolin Zheng (Zhejiang University)
Data SynthesisOptimizationSafty and PrivacyKnowledge DistillationConvolutional Neural NetworkTransformerGenerative Adversarial NetworkImageStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes a two-stage source-free machine forgetting framework called DSDA, which replaces the original training set with synthetic data and achieves efficient forgetting through multi-task optimization.
Efficient Time Series Processing for Transformers and State-Space Models through Token Merging
Leon Götz (Volkswagen AG), Leo Schwinn
Computational EfficiencyTransformerTime SeriesSequential
🎯 What it does: This paper presents the application of token merging in time series models, particularly local merging and causal merging, which can significantly reduce computational load while maintaining sequence order and causality.
Efficiently Access Diffusion Fisher: Within the Outer Product Span Space
Fangyikang Wang (Zhejiang University), Chen Li (Tencent)
OptimizationComputational EfficiencyDiffusion modelImage
🎯 What it does: A diffusion Fisher information structure based on outer product expansion is proposed, and two gradient-free approximation algorithms (DF-TM and DF-EA) are designed for efficiently obtaining the trace and matrix-vector product of the diffusion Fisher;
Efficiently Serving Large Multimodal Models Using EPD Disaggregation
Gursimran Singh (Huawei Technologies), Zhenan Fan (Huawei Technologies)
OptimizationComputational EfficiencyTransformerLarge Language ModelVision Language ModelVideoTextMultimodality
🎯 What it does: An EPD Disaggregation framework is designed, deploying the encoding, pre-filling, and decoding stages of large-scale multimodal models (LMM) on separate resources to reduce latency and memory usage.
Efficiently Vectorized MCMC on Modern Accelerators
Hugh Dance (University College London), Ryan P Adams
OptimizationComputational EfficiencyTabularFinance Related
🎯 What it does: This paper studies an MCMC implementation method based on finite state machines (FSM), which can eliminate loop synchronization bottlenecks when using JAX's vmap vectorization, significantly improving the efficiency of parallel sampling with multiple chains.
EffiCoder: Enhancing Code Generation in Large Language Models through Efficiency-Aware Fine-tuning
Dong HUANG, Jie Zhang
OptimizationComputational EfficiencyAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: By constructing a high-quality and efficient code dataset EFFIINSTRUCT and fine-tuning LLMs with instructions, the correctness and execution efficiency of generated code are improved.
EgoPrivacy: What Your First-Person Camera Says About You?
Yijiang Li (University of California), Nuno Vasconcelos (University of California)
ClassificationRetrievalSafty and PrivacyContrastive LearningVideoMultimodalityBenchmark
🎯 What it does: This paper systematically evaluates the extent to which first-person camera videos leak user privacy and proposes a large-scale benchmark called EgoPrivacy.