ICML 2025 Papers — Page 11
International Conference on Machine Learning · 3257 papers
Exploiting Curvature in Online Convex Optimization with Delayed Feedback
Hao Qiu (Universita degli Studi di Milano), Mengxiao Zhang (University of Iowa)
OptimizationTabular
🎯 What it does: This study investigates online convex optimization under delayed feedback, utilizing loss curvature to improve the regret upper bound.
Exploiting Presentative Feature Distributions for Parameter-Efficient Continual Learning of Large Language Models
Xin Cheng (Southeast University), Lei Feng (Southeast University)
ClassificationGenerationTransformerLarge Language ModelSupervised Fine-TuningTextSequentialBenchmark
🎯 What it does: This paper proposes a method for achieving parameter-efficient online continual learning by utilizing the distribution of performance features from pre-trained large language models, aiming to avoid information leakage (IL) and enhance CL performance.
Exploiting Similarity for Computation and Communication-Efficient Decentralized Optimization
Yuki Takezawa (Kyoto University), Sebastian U Stich
OptimizationComputational EfficiencyImage
🎯 What it does: Two new decentralized optimization methods are proposed: Stabilized Proximal Decentralized Optimization (SPDO) and its accelerated version Accelerated-SPDO, which improve the existing Proximal Decentralized Optimization (PDO) framework and reduce communication and computational costs.
ExPLoRA: Parameter-Efficient Extended Pre-Training to Adapt Vision Transformers under Domain Shifts
Samar Khanna (Stanford University), Stefano Ermon (Stanford University)
Domain AdaptationTransformerSupervised Fine-TuningImage
🎯 What it does: This paper proposes ExPLoRA, a pre-training method for efficiently extending parameters of pre-trained ViT on unsupervised objectives, helping the model achieve better performance under domain transfer.
Exploring and Mitigating Adversarial Manipulation of Voting-Based Leaderboards
Yangsibo Huang (Google), Chiyuan Zhang (Google)
Anomaly DetectionAdversarial AttackLarge Language ModelPrompt EngineeringText
🎯 What it does: The paper analyzes the feasibility of manipulating voting rankings on voting-based LLM evaluation platforms (such as Chatbot Arena) by adversaries exploiting model identification in the absence of protective measures. It proposes a two-step attack: first, training a model recognizer to accurately identify the source of responses, and then voting against the target model. Subsequently, it proposes and implements various mitigation measures (authentication, rate limiting, malicious user detection, etc.) to significantly increase the cost of attacks.
Exploring Criteria of Loss Reweighting to Enhance LLM Unlearning
Puning Yang (Hong Kong Baptist University), Bo Han (Hong Kong Baptist University)
OptimizationSafty and PrivacyTransformerLarge Language ModelText
🎯 What it does: This paper explores the role of loss reweighting in the context of zero-shot learning in large language models through experiments, proposing two types of reweighting objectives: Saturation and Importance, and based on this, designs a new SatImp scheme.
Exploring Invariance in Images through One-way Wave Equations
Yinpeng Chen (Google DeepMind), Zicheng Liu (AMD)
RestorationObject DetectionRepresentation LearningAuto EncoderImage
🎯 What it does: A model based on first-order norm-plus-linear autoregression (FINOLA) is proposed, which can recursively generate complete image feature maps in the latent space using a single central vector, achieving high-quality reconstruction.
Exploring Large Action Sets with Hyperspherical Embeddings using von Mises-Fisher Sampling
Walid Bendada (Deezer Research), Tristan Cazenave (LAMSADE)
Recommendation SystemReinforcement LearningAudio
🎯 What it does: A scalable exploration method based on the von Mises-Fisher distribution (vMF-exp) is proposed, capable of exploring RL problems where the action set is large and actions are represented as unit sphere embedded vectors.
Exploring Representations and Interventions in Time Series Foundation Models
Michał Wiliński (Carnegie Mellon University), Artur Dubrawski (Carnegie Mellon University)
TransformerTime SeriesElectrocardiogram
🎯 What it does: Conduct a similarity analysis of the internal representations of the Time Series Fundamental Model (TSFM), identify redundant blocks, and propose block-based pruning; simultaneously, use linear probing to locate concepts such as trends and seasonality in the model, and implement 'concept-driven' interventions in model predictions using concept vectors during the inference phase.
Exploring Vision Semantic Prompt for Efficient Point Cloud Understanding
Yixin Zha (University of Science and Technology of China), Feng Wu (University of Science and Technology of China)
ClassificationRecognitionSegmentationTransformerPrompt EngineeringPoint Cloud
🎯 What it does: This study proposes a parameter-efficient transfer learning framework that integrates frozen 2D visual pre-trained model semantic prompts into a 3D point cloud pre-trained model, achieving cross-modal fusion through a Hybrid Attention Adapter, significantly enhancing the performance of downstream tasks on point clouds.
Exponential Family Variational Flow Matching for Tabular Data Generation
Andrés Guzmán-Cordero (Vector Institute), Jan-Willem van de Meent (Bosch Delta Lab)
GenerationData SynthesisTransformerFlow-based ModelTabular
🎯 What it does: A variational flow matching (EF-VFM) framework based on the exponential family is proposed, and the TabbyFlow model is implemented for generating mixed continuous and discrete tabular data.
ExpProof : Operationalizing Explanations for Confidential Models with ZKPs
Chhavi Yadav (University of California San Diego), Kamalika Chaudhuri (University of California San Diego)
Explainability and InterpretabilityTabular
🎯 What it does: This paper proposes ExpProof, a protocol that combines cryptographic commitments and zero-knowledge proofs to ensure that predictions and explanations can be generated using the same model while maintaining model confidentiality, and to verify the correctness of the explanation calculations.
Expressive Power of Graph Neural Networks for (Mixed-Integer) Quadratic Programs
Ziang Chen (Massachusetts Institute of Technology), Wotao Yin (Alibaba US)
OptimizationGraph Neural NetworkGraph
🎯 What it does: This paper studies the expressiveness of Graph Neural Networks (GNN) in solving Linear Constrained Quadratic Programming (LCQP) and its mixed-integer variant (MI-LCQP) regarding feasibility, optimal objective values, and optimal solutions.
Expressive Score-Based Priors for Distribution Matching with Geometry-Preserving Regularization
Ziyu Gong (Purdue University), David I. Inouye (Purdue University)
Data SynthesisDomain AdaptationScore-based ModelAuto EncoderImageTabular
🎯 What it does: This paper proposes using a score function instead of an explicit prior density for distribution matching, combined with geometric preservation regularization to achieve more stable non-adversarial distribution matching.
ExtPose: Robust and Coherent Pose Estimation by Extending ViTs
Rongyu Chen (National University of Singapore), Angela Yao (National University of Singapore)
Pose EstimationTransformerImageVideo
🎯 What it does: By extending the Vision Transformer framework, 2D pose information and temporal attention are introduced to achieve robust and pixel-level aligned 3D pose estimation for the human body and hands.
Extracting Rare Dependence Patterns via Adaptive Sample Reweighting
Yiqing Li (Mohamed Bin Zayed University of Artificial Intelligence), Kun Zhang (Carnegie Mellon University)
Anomaly DetectionOptimizationTabularTime SeriesFinance Related
🎯 What it does: A method that combines kernel-based independence testing with adaptive sample importance reweighting is proposed to detect rare dependence patterns that are significant only in small regions but exist in the overall sample; this test is embedded in the PC algorithm to achieve causal structure learning under rare dependence.
Extractive Structures Learned in Pretraining Enable Generalization on Finetuned Facts
Jiahai Feng (University of California Berkeley), Jacob Steinhardt (University of California Berkeley)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposes the 'extractive structures' framework to explain how pre-trained language models learn facts through fine-tuning and perform context-free inference during reasoning.
Extreme Value Policy Optimization for Safe Reinforcement Learning
Shiqing Gao (Shanghai Jiao Tong University), Xinbing Wang (Shanghai Jiao Tong University)
OptimizationSafty and PrivacyReinforcement Learning
🎯 What it does: This study investigates extreme events in safety reinforcement learning and proposes the EVO algorithm, which utilizes extreme value theory to reduce constraint violations.
FAB-PPI: Frequentist, Assisted by Bayes, Prediction-Powered Inference
Stefano Cortinovis (University of Oxford), Francois Caron (University of Oxford)
OptimizationData-Centric LearningDrug DiscoveryTabular
🎯 What it does: A framework that embeds Bayesian prior information into Predictive-Driven Inference (PPI), namely FAB-PPI, is proposed to improve the accuracy of parameter estimation and confidence intervals in a semi-supervised setting with a large number of high-quality machine learning predictions.
FACTER: Fairness-Aware Conformal Thresholding and Prompt Engineering for Enabling Fair LLM-Based Recommender Systems
Arya Fayyazi (University of Southern California), Massoud Pedram (University of Southern California)
Recommendation SystemLarge Language ModelPrompt EngineeringText
🎯 What it does: The FACTER framework is proposed, which achieves fairness monitoring and adaptive correction in LLM recommendation systems through conformal prediction and dynamic prompt engineering.
FactTest: Factuality Testing in Large Language Models with Finite-Sample and Distribution-Free Guarantees
Fan Nie (Stanford University), Linjun Zhang (Rutgers University)
ClassificationTransformerLarge Language ModelText
🎯 What it does: The FACTTEST framework is proposed, modeling the factuality detection of LLMs as a Neyman-Pearson hypothesis test, allowing the model to provide answers under the condition of meeting a user-specified Type I error threshold, otherwise rejecting to answer.
Fair Clustering via Alignment
Kunwoong Kim (Seoul National University), Yongdai Kim (Seoul National University)
OptimizationImageTabular
🎯 What it does: A new fair clustering algorithm, FCA (Fair Clustering via Alignment), is proposed, which aligns data from different protected groups under a joint distribution and then performs standard K-means in the aligned space, maximizing clustering utility while satisfying proportional fairness.
FairICP: Encouraging Equalized Odds via Inverse Conditional Permutation
Yuheng Lai (University of Wisconsin-Madison), Leying Guan (Yale University)
OptimizationFlow-based ModelGenerative Adversarial NetworkTabular
🎯 What it does: A learning framework based on Inverse Conditional Permutation (ICP), called FairICP, is proposed to achieve equalized odds under multidimensional sensitive attributes during training.
Fairness on Principal Stratum: A New Perspective on Counterfactual Fairness
Haoxuan Li (Peking University), Kun Zhang (Carnegie Mellon University)
OptimizationTabular
🎯 What it does: A 'principal counterfactual fairness' framework based on principal stratification is proposed to impose fairness constraints only on the subgroup when sensitive attributes have no individual causal impact on the target outcome.
Fairness Overfitting in Machine Learning: An Information-Theoretic Perspective
Firas Laakom (Center of Excellence for Generative AI, KAUST), Yuheng Bu (University of Florida)
TabularBenchmark
🎯 What it does: This study investigates the generalization error of fairness metrics during the training phase of machine learning models and constructs a theoretical framework based on information theory to quantify this phenomenon of 'fairness overfitting'.
FairPFN: A Tabular Foundation Model for Causal Fairness
Jake Robertson (University of Freiburg), Frank Hutter
TransformerTabular
🎯 What it does: This paper proposes FairPFN, a Transformer-based foundational model for tabular data, aimed at eliminating the causal influence of protected attributes under the condition of having only observational data, thereby enhancing algorithmic fairness.
Falcon: Fast Visuomotor Policies via Partial Denoising
Haojun Chen (Peking University), Yaodong Yang (Peking University)
Computational EfficiencyRobotic IntelligenceReinforcement LearningDiffusion modelSequential
🎯 What it does: A training-free, partially denoised acceleration framework called Falcon is proposed, which can leverage the sequential dependencies of historical actions in visual motion tasks, significantly reducing the number of denoising steps and achieving real-time control.
False Coverage Proportion Control for Conformal Prediction
Alexandre Blain (INRIA), Pierre Neuvial (Institut de Mathématiques de Toulouse)
Tabular
🎯 What it does: The CoJER method is proposed and implemented, utilizing Joint Error Rate (JER) control to achieve high probability control of the error coverage proportion (FCP) for multi-point predictions, providing non-parametric confidence interval generation.
Falsification of Unconfoundedness by Testing Independence of Causal Mechanisms
Rickard Karlsson, JH Krijthe
Tabular
🎯 What it does: In multi-environment observational data, a two-stage algorithm MINT is proposed to falsify the no unmeasured confounding hypothesis using independence tests of causal mechanism parameters.
Fast and Low-Cost Genomic Foundation Models via Outlier Removal
Haozheng Luo (Northwestern University), Han Liu (Northwestern University)
ClassificationOptimizationTransformerSupervised Fine-TuningBiomedical Data
🎯 What it does: A genome-based model named GERM has been constructed, utilizing an outlier-free attention layer to eliminate outliers in attention, thereby achieving fast low-rank adaptation and robust post-training quantization.
Fast and Provable Algorithms for Sparse PCA with Improved Sample Complexity
Jian-Feng Cai (Hong Kong University of Science and Technology), Jiaxi Ying (Hong Kong University of Science and Technology)
OptimizationComputational Efficiency
🎯 What it does: This paper explores the unimodal covariance model in sparse principal component analysis (PCA) and proposes a novel threshold-based algorithm that requires only Ω(k log p) samples to recover sparse unit vectors, significantly improving sample efficiency.
Fast and Robust: Task Sampling with Posterior and Diversity Synergies for Adaptive Decision-Makers in Randomized Environments
Yun Qu (Tsinghua University), Xiangyang Ji (Tsinghua University)
Robotic IntelligenceReinforcement Learning
🎯 What it does: A task sampling method based on posterior sampling and diversity regularization (PDTS) is proposed to achieve more robust active task sampling in domain randomization and meta reinforcement learning.
Fast Estimation of Partial Dependence Functions using Trees
Jinyang Liu (University of Copenhagen), Munir Hiabu (University of Copenhagen)
Explainability and InterpretabilityComputational EfficiencyTabularBenchmark
🎯 What it does: The FastPD algorithm is proposed for the rapid and consistent estimation of the bias function of any subset in tree models, thereby obtaining complete functional decomposition and SHAP values.
Fast Exact Unlearning for In-Context Learning Data for LLMs
Andrei Ioan Muresanu (University of Waterloo), Nicolas Papernot (University of Toronto)
Computational EfficiencyData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes a context learning (in-context learning) method based on large language models (LLM) for task adaptation, and further designs the ERASE algorithm, which enables efficient and precise unlearning when it is necessary to 'forget' certain refined training data, achieving performance comparable to traditional parameter fine-tuning.
Fast Incomplete Multi-view Clustering by Flexible Anchor Learning
Yalan Qin (Shanghai University), Xinpeng Zhang (Shanghai University)
OptimizationGraph Neural NetworkContrastive LearningImage
🎯 What it does: A unified framework for large-scale incomplete multi-view clustering, FIML, is proposed, which integrates graph construction, anchor point learning, and graph partitioning to achieve clustering results in one go.
Fast Inference with Kronecker-Sparse Matrices
Antoine Gonon (Ecole Polytechnique Federale de Lausanne), TUNG QUOC LE
OptimizationComputational EfficiencyTransformerTabularBenchmark
🎯 What it does: A fused output-resident GPU kernel for Kronecker-sparse matrix multiplication is proposed, significantly reducing memory transfer, improving speed and energy consumption;
Fast Large Language Model Collaborative Decoding via Speculation
Jiale Fu (Southeast University), Xu Yang (Southeast University)
GenerationComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: A framework named CoS (Collaborative decoding via Speculation) is proposed to accelerate multi-model collaborative decoding while maintaining the quality of generation.
Fast Min-$\epsilon$ Segmented Regression using Constant-Time Segment Merging
Ansgar Lößer (TU Darmstadt), Björn Scheuermann (TU Darmstadt)
OptimizationComputational EfficiencyTime Series
🎯 What it does: A new greedy algorithm is proposed that utilizes constant time segment merging and maintains error to solve the minimum MSE piecewise regression problem for a given k segments.
Fast Tensor Completion via Approximate Richardson Iteration
Mehrdad Ghadiri (Massachusetts Institute of Technology), Ali Jadbabaie (Massachusetts Institute of Technology)
OptimizationComputational EfficiencyImageMagnetic Resonance Imaging
🎯 What it does: This paper proposes an algorithm that transforms the tensor completion regression problem into a structured tensor decomposition subproblem through lifting, and implements sublinear time tensor completion using approximate Richardson iteration.
Fast Video Generation with Sliding Tile Attention
Peiyuan Zhang (University of California), Hao Zhang (University of California)
GenerationData SynthesisComputational EfficiencyTransformerDiffusion modelVideo
🎯 What it does: Proposes Sliding Tile Attention (STA), which replaces full 3D attention in video diffusion Transformers, significantly reducing inference time.
Fast, Accurate Manifold Denoising by Tunneling Riemannian Optimization
Shiyu Wang (Columbia University), John Wright (Columbia University)
OptimizationTime Series
🎯 What it does: This paper proposes an online learning hybrid-order Riemannian optimization framework to efficiently denoise noisy samples on unknown low-dimensional manifolds during testing.
FastCAV: Efficient Computation of Concept Activation Vectors for Explaining Deep Neural Networks
Laines Schmalwasser (German Aerospace Center), Julia Niebling (German Aerospace Center)
Explainability and InterpretabilityComputational EfficiencyConvolutional Neural NetworkTransformerImageBiomedical Data
🎯 What it does: This paper proposes FastCAV, an algorithm that quickly computes the Concept Activation Vector (CAV) directly through the difference vector between the mean of concept samples and the global mean, without training a linear SVM.
Faster and Stronger: When ANN-SNN Conversion Meets Parallel Spiking Calculation
Zecheng Hao (Peking University), Tiejun Huang (Peking University)
ClassificationComputational EfficiencySpiking Neural NetworkImage
🎯 What it does: A parallel ANN-SNN conversion framework is proposed, achieving high-precision SNN inference with low time steps by constructing a mathematical equivalent mapping of parallel synaptic computations and ANN activation functions.
Faster Approximation Algorithms for k-Center via Data Reduction
Arnold Filtser (Bar Ilan University), Qin Zhang (Indiana University)
OptimizationComputational EfficiencyTabular
🎯 What it does: This paper studies a fast approximation algorithm for the Euclidean k-center problem in the case of large k, proposing a data reduction method based on α-coresets and constructing two types of k·o(n) sized coreset, thereby reducing the time complexity of the approximation algorithm from O(nk) to nearly O(n).
Faster Global Minimum Cut with Predictions
Helia Niaparast (Carnegie Mellon University), Karan Singh (Carnegie Mellon University)
OptimizationComputational EfficiencyGraph
🎯 What it does: This paper proposes an accelerated version of the Karger and Karger-Stein minimum cut algorithms using machine learning predictions, which can significantly reduce the time complexity of solving the global minimum cut under the premise of known prediction error rates.
Faster Rates for Private Adversarial Bandits
Hilal Asi (Apple), Kunal Talwar (University of Michigan)
OptimizationSafty and PrivacyAdversarial AttackReinforcement Learning
🎯 What it does: A new differentially private algorithm is designed for adversarial Bandits and Bandits with expert advice.
Faster Stochastic Optimization with Arbitrary Delays via Adaptive Asynchronous Mini-Batching
Amit Attia (Tel Aviv University), Tomer Koren (Google Research)
OptimizationImageStochastic Differential Equation
🎯 What it does: This paper proposes a method to adaptively convert standard stochastic first-order optimization methods into asynchronous versions, utilizing batching and quantization delays to improve convergence speed under arbitrary delay models.
FDGen: A Fairness-Aware Graph Generation Model
Zichong Wang (Florida International University), Wenbin Zhang (Florida International University)
GenerationData SynthesisGraph Neural NetworkDiffusion modelGraph
🎯 What it does: This paper studies a fair graph generation model FDGen, aimed at simultaneously eliminating structural bias and feature bias.
Feasible Action Search for Bandit Linear Programs via Thompson Sampling
Aditya Gangrade (Boston University), Venkatesh Saligrama (Boston University)
OptimizationReinforcement Learning from Human FeedbackReinforcement Learning
🎯 What it does: An efficient algorithm named FAST is proposed for online searching for safe actions that satisfy unknown linear constraints, or for determining the infeasibility of constraints, while providing approximate optimal safety margins within a finite sample.
FEAT-KD: Learning Concise Representations for Single and Multi-Target Regression via TabNet Knowledge Distillation
Kei Sen Fong (National University of Singapore), Mehul Motani (National University of Singapore)
OptimizationKnowledge DistillationRepresentation LearningTransformerTabular
🎯 What it does: Proposes FEAT-KD, which achieves interpretable symbolic feature combinations for regression through block knowledge distillation of a trained TabNet model.
FeatSharp: Your Vision Model Features, Sharper
Mike Ranzinger (NVIDIA), Andrew Tao (NVIDIA)
ClassificationObject DetectionSegmentationRetrievalKnowledge DistillationTransformerImage
🎯 What it does: A feature upsampling framework named FeatSharp is proposed, which can elevate the feature maps of low-resolution visual encoders to higher resolutions at a lower cost while maintaining the original model representation, and utilizes tile information for fine-grained refinement.
Feature Importance Metrics in the Presence of Missing Data
Henrik von Kleist (Helmholtz Munich - German Research Center for Environmental Health), Carsten Marr (Helmholtz Munich - German Research Center for Environmental Health)
Time SeriesElectronic Health Records
🎯 What it does: This paper studies how to define, identify, and estimate feature importance in the presence of missing values. It proposes two evaluation frameworks: one for complete data and one for observed data, and introduces the Feature Measurement Importance Gradient (FMIG) to assess the impact of incremental increases in measurement probability on predictive performance.
Feature Learning beyond the Lazy-Rich Dichotomy: Insights from Representational Geometry
Chi-Ning Chou (Flatiron Institute), SueYeon Chung (New York University)
Representation LearningConvolutional Neural NetworkRecurrent Neural NetworkImage
🎯 What it does: This paper proposes a framework based on representational geometry and set capacity to quantify the depth, strategy, and stages of neural networks in the feature learning process, and validates its effectiveness through experiments.
Feature learning from non-Gaussian inputs: the case of Independent Component Analysis in high dimensions
Fabiola Ricci (International School of Advanced Studies), Sebastian Goldt (International School of Advanced Studies)
OptimizationRepresentation LearningImage
🎯 What it does: This paper studies the sample complexity of the popular ICA algorithm FastICA and stochastic gradient descent (SGD) for feature learning from non-Gaussian inputs.
Feature out! Let Raw Image as Your Condition for Blind Face Restoration
Xinmin Qiu (University of Chinese Academy of Sciences), Zicheng Zhang (University of Chinese Academy of Sciences)
RestorationDiffusion modelImageStochastic Differential Equation
🎯 What it does: A pseudo-hash based Schrödinger bridge model P-I2SB is proposed, using original low-quality images directly as conditions to achieve blind face restoration.
Feature Shift Localization Network
Míriam Barrabés (Stanford University), Alexander G. Ioannidis (Stanford University)
Domain AdaptationAnomaly DetectionConvolutional Neural NetworkTabular
🎯 What it does: This paper proposes a neural network called FSL-Net that can quickly locate feature distribution shifts between data sources, helping to identify the specific features that cause the shifts.
Feature-Mapping Topology Optimization with Neural Heaviside Signed Distance Functions
Aleksandr Kolomeitsev (Skolkovo Institute of Science and Technology), ANH-HUY PHAN
OptimizationAuto EncoderMesh
🎯 What it does: A feature mapping topology optimization method based on a neural Heaviside symbol distance function is proposed, which can directly generate manufacturable voids during the optimization process.
Features are fate: a theory of transfer learning in high-dimensional regression
Javan Tahir (Stanford University), Grant M. Rotskoff
Tabular
🎯 What it does: This paper constructs a theoretical framework for transfer learning in high-dimensional regression tasks by analyzing deep linear networks and shallow ReLU networks. It derives a phase diagram of transfer performance and verifies that the overlap of feature spaces is a key indicator determining the success of transfer.
FedBEns: One-Shot Federated Learning based on Bayesian Ensemble
Jacopo Talpini (University of Milano-Bicocca), Giovanni Neglia (Inria)
Federated LearningImage
🎯 What it does: This paper proposes FedBEns, a single-round federated learning algorithm based on Bayesian inference, which utilizes multimodal Laplace approximation to aggregate client posteriors, construct a global posterior, and make predictions through model averaging.
FedClean: A General Robust Label Noise Correction for Federated Learning
Xiaoqian Jiang (Southeast University), Jing Zhang (Southeast University)
Federated LearningMixture of ExpertsImage
🎯 What it does: We propose FedClean, a general federated learning label noise correction framework that can automatically correct noisy labels and improve global model performance in the absence of clean clients.
FedECADO: A Dynamical System Model of Federated Learning
Aayushya Agarwal (Carnegie Mellon University), Lawrence Pileggi
Federated LearningConvolutional Neural NetworkImageTextOrdinary Differential Equation
🎯 What it does: The FedECADO algorithm is proposed, utilizing dynamic systems and equivalent circuit models to address the model inconsistency issues caused by non-IID data distribution and heterogeneous computation in federated learning.
Federated Causal Structure Learning with Non-identical Variable Sets
Yunxia Wang (Shanxi University), Jiye Liang (Shanxi University)
Federated LearningSafty and PrivacyTabular
🎯 What it does: The FedCDnv algorithm is proposed to handle the situation where different client observation variables are not completely the same in a federated learning environment, detecting and eliminating pseudo-correlation caused by non-overlapping variable pairs, aggregating 'correct and important' causal relationships to obtain a global causal graph and its determined subgraphs.
Federated Disentangled Tuning with Textual Prior Decoupling and Visual Dynamic Adaptation
Yihao Yang (Wuhan University), Mang Ye (Wuhan University)
Domain AdaptationFederated LearningPrompt EngineeringMixture of ExpertsVision Language ModelImageMultimodality
🎯 What it does: This paper proposes the FedDDA method for parameter-efficient fine-tuning of visual-language models in a federated learning environment to address the performance degradation caused by data heterogeneity.
Federated Generalised Variational Inference: A Robust Probabilistic Federated Learning Framework
Terje Mildner (University of Warwick), Theodoros Damoulas (University of Warwick)
Federated LearningTabularSequential
🎯 What it does: A Federated Generalised Variational Inference (FEDGVI) framework is proposed to address the mis-specification of priors and likelihoods in federated learning, providing robust and calibrated uncertainty estimates.
Federated In-Context Learning: Iterative Refinement for Improved Answer Quality
Ruhan Wang (Indiana University), Dongruo Zhou (Indiana University)
Federated LearningTransformerLarge Language ModelText
🎯 What it does: Proposes the Fed-ICL framework, which combines federated learning with context learning to iteratively improve question-answering responses through multiple rounds of client-server interactions.
Federated Incomplete Multi-view Clustering with Globally Fused Graph Guidance
Guoqing Chao (Harbin Institute of Technology), Dianhui Chu (Harbin Institute of Technology)
Federated LearningGraph Neural NetworkGraph
🎯 What it does: This paper proposes a Federated Incomplete Multi-View Clustering method (FIMCFG), which achieves end-to-end feature extraction and clustering through a globally fused graph-guided dual-head graph convolutional encoder and fusion module.
Federated Learning for Feature Generalization with Convex Constraints
Dongwon Kim (Sungkyunkwan University), Kwangsu Kim (Korea Advanced Institute of Science and Technology)
OptimizationFederated LearningConvolutional Neural NetworkImage
🎯 What it does: Proposes FedCONST, which incorporates convex constraints in client training to achieve feature generalization and robust aggregation of the global model;
Federated Node-Level Clustering Network with Cross-Subgraph Link Mending
Jingxin Liu (Hainan University), Jieren Cheng (Hainan University)
Federated LearningSafty and PrivacyGraph Neural NetworkGraph
🎯 What it does: This paper proposes the Federated Node-level Clustering Network (FedNCN), a federated learning framework for unsupervised node clustering on distributed subgraphs across multiple clients. The core idea is to utilize clustering prior information to repair the missing connections caused by graph partitioning, and to achieve information sharing and model collaboration through an MLP projector and a GNN generator.
Federated Oriented Learning: A Practical One-Shot Personalized Federated Learning Framework
Guan Huang (Auburn University), Tao Shu (Auburn University)
Federated LearningKnowledge DistillationSupervised Fine-TuningImage
🎯 What it does: Proposes Federated Oriented Learning (FOL), a federated learning framework that achieves client personalized model enhancement under single-round communication.
FedOne: Query-Efficient Federated Learning for Black-box Discrete Prompt Learning
Ganyu Wang (Western University), Charles Ling (Western University)
Federated LearningComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Proposes the FedOne framework for black-box discrete prompt learning of cloud-based LLMs, achieving prompt optimization through federated learning and activating only one client per round to minimize query costs.
FedPHA: Federated Prompt Learning for Heterogeneous Client Adaptation
Chengying Fang (Wuhan University), Mang Ye (Wuhan University)
Domain AdaptationFederated LearningPrompt EngineeringVision Language ModelContrastive LearningImageMultimodality
🎯 What it does: A FedPHA framework has been developed to implement prompt learning for visual language models in federated learning to adapt to heterogeneous clients.
FedSMU: Communication-Efficient and Generalization-Enhanced Federated Learning through Symbolic Model Updates
Xinyi Lu (Shanghai Jiao Tong University), Hongkai Xiong (Shanghai Jiao Tong University)
OptimizationFederated LearningConvolutional Neural NetworkTransformerImageText
🎯 What it does: This paper proposes FedSMU, an algorithm that symbolically updates client models in federated learning (using sign operations) and employs the Lion optimizer for split execution, which reduces communication costs and enhances the generalization performance of the global model.
FedSSI: Rehearsal-Free Continual Federated Learning with Synergistic Synaptic Intelligence
Yichen Li (Huazhong University of Science and Technology), Ruixuan Li (Huazhong University of Science and Technology)
Federated LearningImage
🎯 What it does: This paper proposes a regularization-based continual federated learning method called FedSSI, aimed at alleviating catastrophic forgetting without using sample replay while considering data heterogeneity.
Feedforward Few-shot Species Range Estimation
Christian Lange (University of Edinburgh), Oisin Mac Aodha (University of Edinburgh)
TransformerImageText
🎯 What it does: This paper proposes FS-SINR, a Transformer-based few-shot species distribution range estimation model that can infer the spatial distribution of unknown species in a single forward pass with only a few observation locations or metadata such as text/images provided.
Ferret: Federated Full-Parameter Tuning at Scale for Large Language Models
Yao Shu (Hong Kong University of Science and Technology), Fei Yu (Guangdong Laboratory of Artificial Intelligence and Digital Economy)
OptimizationFederated LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Large-scale full-parameter fine-tuning of large language models in a federated environment (Ferret)
Few-Shot Learner Generalizes Across AI-Generated Image Detection
Shiyu Wu (Institute of Automation, Chinese Academy of Sciences), Yequan Wang (Beijing Academy of Artificial Intelligence)
ClassificationObject DetectionConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: An AI-generated image detector based on few-shot learning (Few-Shot Detector, FSD) is proposed, which can identify fake images generated by a given small number of samples from unknown generative models.
Feynman-Kac Correctors in Diffusion: Annealing, Guidance, and Product of Experts
Marta Skreta (University of Toronto), Kirill Neklyudov (Université de Montréal)
GenerationDrug DiscoveryDiffusion modelImagePhysics RelatedStochastic Differential Equation
🎯 What it does: A weighted stochastic differential equation (FKC) based on the Feynman-Kac equation is proposed, and precise sampling of target distributions such as temperature, guidance, and expert combinations during the inference phase of diffusion models is achieved through SMC resampling.
FG-CLIP: Fine-Grained Visual and Textual Alignment
Chunyu Xie (Beihang University), Yuhui Yin (360 AI Research)
ClassificationRecognitionRetrievalTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes Fine-Granularity CLIP (FG-CLIP), which achieves fine-grained alignment between images and text through a three-stage training process (global contrast, regional contrast, and hard negative sample learning);
FIC-TSC: Learning Time Series Classification with Fisher Information Constraint
Xiwen Chen (Clemson University), Abolfazl Razi (Clemson University)
ClassificationDomain AdaptationTime SeriesBiomedical Data
🎯 What it does: This paper proposes a time series classification learning framework based on Fisher Information Constraint (FIC), called FIC-TSC, to alleviate performance degradation caused by training-testing domain transfer, and enhances generalization ability by guiding the network to converge to flat minima.
FicGCN: Unveiling the Homomorphic Encryption Efficiency from Irregular Graph Convolutional Networks
Zhaoxuan Kan (Chinese Academy of Sciences), Xing Hu (Chinese Academy of Sciences)
Graph Neural NetworkGraph
🎯 What it does: Designed and implemented a CKKS-based FicGCN framework for efficient inference of graph convolutional networks in a homomorphic encryption environment.
Field Matching: an Electrostatic Paradigm to Generate and Transfer Data
Alexander Kolesov (Skolkovo Institute of Science and Technology), Alexander Korotin (Artificial Intelligence Research Institute)
Image TranslationGenerationData SynthesisImageOrdinary Differential Equation
🎯 What it does: Proposes the Electrostatic Field Matching (EFM) method, which utilizes the distribution of positive and negative charges to generate electric field lines and migrate samples along them, achieving a distribution transformation from noise to data and from data to data.
Finding Wasserstein Ball Center: Efficient Algorithm and The Applications in Fairness
Yuntao Wang (University of Science and Technology of China), Hu Ding (University of Science and Technology of China)
OptimizationComputational EfficiencyImagePoint Cloud
🎯 What it does: This paper proposes a new concept of Wasserstein Ball Center (WBC), aimed at finding the minimum Wasserstein ball that covers all input distributions while ensuring fairness, replacing the traditional Wasserstein barycenter (WB); it also provides a linear programming formulation for discrete distributions and designs a significantly accelerated solving algorithm based on the Interior Point Method (IPM);
Fine-Grained Captioning of Long Videos through Scene Graph Consolidation
Sanghyeok Chu (Seoul National University), Bohyung Han (Seoul National University)
GenerationRetrievalOptimizationComputational EfficiencyGraph Neural NetworkTransformerVision Language ModelVideoTextGraph
🎯 What it does: By dividing long videos into short segments, existing visual-language models are first used to generate segment subtitles, then each subtitle is parsed into a scene graph, followed by aggregating all scene graphs into a unified graph, and finally, a lightweight graph-to-text decoder is used to generate fine-grained subtitles for the entire video.
Finite-Sample Convergence Bounds for Trust Region Policy Optimization in Mean Field Games
Antonio Ocello (Ecole Polytechnique), Eric Moulines (Mohamed bin Zayed University of Artificial Intelligence)
OptimizationReinforcement Learning
🎯 What it does: Two mean field game (MFG) algorithms based on trust region policy optimization (TRPO) are proposed—Exact MF-TRPO and Sample-Based MF-TRPO—to solve the steady-state mean field Nash equilibrium (MFNE) with average cost.
Finite-Time Analysis of Discrete-Time Stochastic Interpolants
Yuhao Liu (Tsinghua University), Longbo Huang (Tsinghua University)
OptimizationTabularStochastic Differential Equation
🎯 What it does: This paper proposes a discrete-time stochastic interpolation framework sampler and provides an upper bound on the KL error for finite time, offering theoretical guidance for discrete step size scheduling.
Finite-Time Convergence Rates in Stochastic Stackelberg Games with Smooth Algorithmic Agents
Eric Frankel (University of Washington), Lillian J. Ratliff (University of Washington)
Recommendation SystemOptimizationReinforcement LearningAgentic AI
🎯 What it does: The paper studies how decision-makers can achieve finite-time convergence by controlling the learning behavior of agents in a stochastic Stackelberg game with smooth algorithmic agents.
Finite-Time Global Optimality Convergence in Deep Neural Actor-Critic Methods for Decentralized Multi-Agent Reinforcement Learning
Zhiyao Zhang (Ohio State University), Jia Liu (Ohio State University)
OptimizationReinforcement Learning
🎯 What it does: This paper proposes a novel deep neural network (DNN) benchmark actor-critic framework for decentralized multi-agent reinforcement learning (MARL) and provides theoretical guarantees for its convergence to global optimality within a finite time.
FireFlow: Fast Inversion of Rectified Flow for Image Semantic Editing
Yingying Deng (University of Science and Technology Beijing), Fan Tang (Institute of Computing Technology Chinese Academy of Sciences)
Image TranslationGenerationComputational EfficiencyFlow-based ModelRectified FlowImageOrdinary Differential Equation
🎯 What it does: This paper presents FireFlow, a fast inversion and semantic editing method for the ReFlow model, achieving an efficient 8-step reverse and editing process.
Fishers for Free? Approximating the Fisher Information Matrix by Recycling the Squared Gradient Accumulator
Yu Xin Li, Colin Raffel (University of Toronto)
OptimizationTransformerLarge Language ModelText
🎯 What it does: Utilize the existing squared gradient accumulator (Squisher) during the training process to approximate the diagonal of the Fisher information matrix, thereby enabling various Fisher-based importance estimation tasks such as model merging, sparse training, pruning, task embedding, and EWC.
FisherSFT: Data-Efficient Supervised Fine-Tuning of Language Models Using Information Gain
Rohan Deb (University of Illinois), Branislav Kveton (Adobe Research)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes FisherSFT, a data-efficient supervised fine-tuning (SFT) method for LLMs that selects the most representative training sentences using information gain.
Fixed-Confidence Multiple Change Point Identification under Bandit Feedback
Joseph Lazzaro (Imperial College London), Ciara Pike-Burke (Imperial College London)
Reinforcement LearningTime Series
🎯 What it does: This paper proposes a fixed confidence change point detection problem and presents a computationally efficient change point identification algorithm, MCPI, which can adaptively locate all change points under known distribution noise.
Fixing the Double Penalty in Data-Driven Weather Forecasting Through a Modified Spherical Harmonic Loss Function
Christopher Subich (Environment and Climate Change Canada), Jing Yang (Environment and Climate Change Canada)
Graph Neural NetworkTime Series
🎯 What it does: By rewriting the mean squared error loss function and utilizing spherical harmonic decomposition to separate amplitude errors from correlation errors, the double-penalty problem in data-driven weather forecasting is eliminated.
Fixing the Loose Brake: Exponential-Tailed Stopping Time in Best Arm Identification
Kapilan Balagopalan (University of Arizona), Kwang-Sung Jun (University of Arizona)
Optimization
🎯 What it does: Two optimal arm identification algorithms under fixed confidence (FC-DSH and BrakeBooster) are proposed, achieving exponential stopping time tails.
FLAM: Frame-Wise Language-Audio Modeling
Yusong Wu (Mila - Quebec AI Institute, University of Montreal), Justin Salamon (Adobe Research)
ClassificationRetrievalTransformerContrastive LearningTextMultimodalityAudio
🎯 What it does: This paper proposes FLAM, a frame-level audio-language model that achieves fine-grained temporal alignment between audio and text at the open vocabulary level.
FlashTP: Fused, Sparsity-Aware Tensor Product for Machine Learning Interatomic Potentials
Seung Yul Lee (Seoul National University), Jae W. Lee (Seoul National University)
Computational EfficiencyTabular
🎯 What it does: A GPU-accelerated library named FlashTP has been developed to efficiently implement the tensor product layer in machine learning interatomic potentials (MLIP), significantly improving inference and training speed while reducing memory usage.
Flat-LoRA: Low-Rank Adaptation over a Flat Loss Landscape
Tao Li (Shanghai Jiao Tong University), Xiaolin Huang (Shanghai Jiao Tong University)
OptimizationTransformerSupervised Fine-TuningImageText
🎯 What it does: A low-rank adaptation method called Flat-LoRA is proposed to find flat minima in the full parameter space, improving the generalization performance of LoRA.
FlatQuant: Flatness Matters for LLM Quantization
Yuxuan Sun (Huawei Noah's Ark Lab), Jun Yao
TransformerLarge Language ModelText
🎯 What it does: This paper studies a novel post-training quantization method called FLATQUANT, which enhances the flatness of LLM weights and activations through learnable fast affine transformations, significantly reducing quantization error.
Fleet of Agents: Coordinated Problem Solving with Large Language Models
Lars Henning Klein (École Polytechnique Fédérale de Lausanne), Akhil Arora (Aarhus University)
TransformerLarge Language ModelAgentic AIText
🎯 What it does: A multi-agent framework named FOA is proposed, utilizing LLMs as agents to perform autonomous exploration and resampling based on genetic particle filtering in dynamic tree search, addressing complex reasoning and decision-making tasks.
Flex3D: Feed-Forward 3D Generation with Flexible Reconstruction Model and Input View Curation
Junlin Han (Meta), Filippos Kokkinos
GenerationData SynthesisTransformerDiffusion modelNeural Radiance FieldGaussian SplattingImageVideoPoint CloudMesh
🎯 What it does: A two-stage pipeline called Flex3D is proposed, which first generates and selects high-quality views, and then uses FlexRM, capable of handling any number of views, to achieve fast 3D Gaussian reconstruction.
FlexControl: Computation-Aware Conditional Control with Differentiable Router for Text-to-Image Generation
Zheng Fang (University of Warwick), Hongkai Wen (University of Warwick)
GenerationData SynthesisDiffusion modelImageText
🎯 What it does: Proposes the FlexControl framework, which achieves conditional control of dynamic activation by adding learnable routers before each diffusion network block;