NeurIPS 2025 Papers — Page 17
Conference on Neural Information Processing Systems · 5275 papers
Fast Data Attribution for Text-to-Image Models
Sheng-Yu Wang (Carnegie Mellon University), Jun-Yan Zhu (Carnegie Mellon University)
GenerationComputational EfficiencyData-Centric LearningDiffusion modelContrastive LearningImageText
🎯 What it does: A scalable and efficient data attribution method is proposed, utilizing unlearning of generated images as a teacher to distill attribution results into a feature embedding space, enabling rapid retrieval of the influence of training samples.
Fast exact recovery of noisy matrix from few entries: the infinity norm approach
BaoLinh Tran (Yale University), Van Vu (Yale University)
RestorationOptimizationImage
🎯 What it does: This paper proposes a fast algorithm that requires only three basic assumptions: low rank, unbiasedness, and sufficient sampling (along with a signal strength condition) to achieve accurate recovery in noise matrix completion;
Fast Inference for Augmented Large Language Models
Rana Shahout, Michael Mitzenmacher
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes the MARS (Memory-and-API-Rooted Scheduler) framework, specifically designed for API composite LLM requests. It can pre-allocate processing strategies by predicting output length and API duration under GPU memory constraints, and based on this, perform queue scheduling based on total memory usage, significantly reducing end-to-end latency and the time to generate the first token.
Fast Last-Iterate Convergence of SGD in the Smooth Interpolation Regime
Amit Attia (Tel Aviv University), Tomer Koren (Tel Aviv University)
Optimization
🎯 What it does: This paper studies the convergence properties of the last iteration of stochastic gradient descent (SGD) in the case of interpolation (low noise) for smooth convex objective functions, and provides a theoretical upper bound on the step size 0 < η < 2/β.
Fast Local Search Algorithms for Clustering with Adaptive Sampling and Bandit Strategies
Junyu Huang (Central South University), Jianxin Wang (Central South University)
OptimizationTabular
🎯 What it does: A local search algorithm combining adaptive sampling and multi-armed bandit strategies is designed to quickly complete k-means and k-median clustering on large-scale datasets with constant approximation guarantees.
Fast Monte Carlo Tree Diffusion: 100× Speedup via Parallel and Sparse Planning
Jaesik Yoon (Korea Advanced Institute of Science and Technology), Sungjin Ahn (Korea Advanced Institute of Science and Technology)
Robotic IntelligenceReinforcement LearningDiffusion modelGraph
🎯 What it does: This paper proposes Fast-MCTD, which combines and optimizes Monte Carlo Tree Diffusion (MCTD) to achieve significantly accelerated trajectory planning.
Fast MRI for All: Bridging Access Gaps by Training without Raw Data
Yasar Utku Alcalar (University of Minnesota), Mehmet Akcakaya (University of Minnesota)
RestorationOptimizationImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Train an unsupervised physics-driven deep learning MRI reconstruction model using only available clinical reconstructed images.
Fast Non-Log-Concave Sampling under Nonconvex Equality and Inequality Constraints with Landing
Kijung Jeon (Georgia Institute of Technology), Molei Tao (Georgia Institute of Technology)
OptimizationComputational EfficiencyData-Centric LearningTabularFinance RelatedStochastic Differential Equation
🎯 What it does: A projection-free over-damped Langevin sampling framework OLLA is proposed, which can simultaneously handle non-convex equality and inequality constraints.
Fast Projection-Free Approach (without Optimization Oracle) for Optimization over Compact Convex Set
Chenghao Liu (City University of Hong Kong), Minghua Chen (Chinese University of Hong Kong)
Optimization
🎯 What it does: A projection-free first-order method named Hom-PGD is proposed, which utilizes gauge mapping to map any compact convex set to the unit ball, thereby completing optimization without the need for expensive projections or linear optimization oracles.
Fast Rate Bounds for Multi-Task and Meta-Learning with Different Sample Sizes
Hossein Zakerinia (Institute of Science and Technology Austria), Christoph H. Lampert (Institute of Science and Technology Austria)
OptimizationMeta LearningImageTabular
🎯 What it does: A fast convergence PAC-Bayes generalization bound for multi-task learning and meta-learning under the condition of imbalanced task sample sizes is proposed.
Fast Solvers for Discrete Diffusion Models: Theory and Applications of High-Order Algorithms
Yinuo Ren (Stanford University), Lexing Ying (Stanford University)
Diffusion modelImageText
🎯 What it does: Developed two high-order inference algorithms for discrete diffusion models (θ-RK-2 and θ-Trapezoidal), achieving more efficient and accurate sampling;
Fast Training of Large Kernel Models with Delayed Projections
Amirhesam Abedsoltan (University of California San Diego), Mikhail Belkin (University of California San Diego)
OptimizationComputational EfficiencyImageAudio
🎯 What it does: A new kernel method training framework, EigenPro 4, is proposed, which utilizes a delayed projection strategy to achieve large-scale model training while maintaining the advantages of kernel methods.
Fast Zeroth-Order Convex Optimization with Quantum Gradient Methods
Junhyung Lyle Kim (JPMorganChase), Shouvanik Chakrabarti (JPMorganChase)
Optimization
🎯 What it does: This paper proposes a zero-order quantum gradient method that only uses noise function value queries, proving that its query complexity in convex optimization matches that of classical first-order gradient methods. The framework is extended to non-Euclidean spaces, solving white-box problems such as semidefinite programming, linear programming, and zero-sum games through quantum mirror descent.
Fast-in-Slow: A Dual-System VLA Model Unifying Fast Manipulation within Slow Reasoning
Hao Chen (Chinese University of Hong Kong), Pheng-Ann Heng (Chinese University of Hong Kong)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerReinforcement LearningVision Language ModelVision-Language-Action ModelDiffusion modelMultimodalitySequential
🎯 What it does: Proposes the Fast-in-Slow dual-system VLA model, embedding a high-speed execution module within the same pre-trained VLM to achieve real-time robot control.
Fast-Slow Thinking GRPO for Large Vision-Language Model Reasoning
Wenyi Xiao (Zhejiang University), Leilei Gan (Zhejiang University)
OptimizationComputational EfficiencyTransformerReinforcement LearningVision Language ModelTextMultimodality
🎯 What it does: We propose FAST-GRPO, a fast-slow thinking framework for large visual language models (LVLM) that dynamically adjusts the depth of reasoning to avoid overthinking.
FAST: Foreground‑aware Diffusion with Accelerated Sampling Trajectory for Segmentation‑oriented Anomaly Synthesis
Xichen Xu (Global Institute of Future Technology, Shanghai Jiao Tong University), Zhichao Lu (City University of Hong Kong)
SegmentationData SynthesisAnomaly DetectionDiffusion modelImage
🎯 What it does: A foreground-aware diffusion framework FAST is proposed for industrial anomaly segmentation, supporting controllable and rapid anomaly synthesis.
FastDINOv2: Frequency Based Curriculum Learning Improves Robustness and Training Speed
Jiaqi Zhang (Brown University), Randall Balestriero (Brown University)
ClassificationRecognitionComputational EfficiencyKnowledge DistillationTransformerContrastive LearningImage
🎯 What it does: FastDINOv2 is proposed, achieving efficient pre-training of DINOv2 through two-stage frequency curriculum learning and Gaussian noise patching, significantly improving convergence speed while maintaining robustness.
Faster Algorithms for Structured John Ellipsoid Computation
Yang Cao (Wyoming Seminary), Tianyi Zhou (University of Southern California)
OptimizationComputational EfficiencyTabular
🎯 What it does: This paper studies the problem of approximately computing the John ellipsoid within a symmetric polytope and proposes two efficient algorithms.
Faster Fixed-Point Methods for Multichain MDPs
Matthew Zurek (University of Wisconsin Madison), Yudong Chen (University of Wisconsin Madison)
OptimizationReinforcement Learning
🎯 What it does: The value iteration algorithm for multi-chain average reward Markov decision processes has been studied and improved.
Faster Generic Identification in Tree-Shaped Structural Causal Models
Yasmine Briefs (Max Planck Institute for Informatics), Markus Bläser (Saarland University)
OptimizationComputational EfficiencyGraph
🎯 What it does: The general identifiability of parameters in tree-structured linear structural causal models was studied, and a randomized polynomial-time algorithm was proposed.
Faster Video Diffusion with Trainable Sparse Attention
Peiyuan Zhang (University of California San Diego), Hao Zhang (University of California San Diego)
GenerationOptimizationComputational EfficiencyKnowledge DistillationTransformerDiffusion modelVideo
🎯 What it does: Designed and implemented a trainable, hardware-friendly video sparse attention mechanism (VSA) to accelerate the training and inference of video diffusion Transformers.
FastJAM: a Fast Joint Alignment Model for Images
Omri Hirsch (Ben Gurion University of the Negev), Oren Freifeld (Ben Gurion University of the Negev)
Image TranslationOptimizationComputational EfficiencyGraph Neural NetworkImage
🎯 What it does: A FastJAM framework based on graph neural networks is proposed, utilizing sparse keypoint matching for rapid joint alignment;
FastLongSpeech: Enhancing Large Speech-Language Models for Efficient Long-Speech Processing
Shoutao Guo (Chinese Academy of Sciences), Yang Feng (Chinese Academy of Sciences)
CompressionComputational EfficiencyTransformerLarge Language ModelBenchmarkAudio
🎯 What it does: The FastLongSpeech framework is proposed, which enables large-scale speech-language models (LSLM) to efficiently handle long speech through iterative fusion and dynamic compression training.
FastVID: Dynamic Density Pruning for Fast Video Large Language Models
Leqi Shen (Tsinghua University), Guiguang Ding (Tsinghua University)
CompressionComputational EfficiencyTransformerLarge Language ModelVision Language ModelVideoBenchmark
🎯 What it does: A fast video large language model (FastVID) framework based on dynamic density pruning is proposed, focusing on spatiotemporal redundancy compression during the inference phase, significantly improving the inference speed of video LLMs and reducing computational costs.
Feasibility-Aware Decision-Focused Learning for Predicting Parameters in the Constraints
Jayanta Mandi (KU Leuven), Tias Guns (KU Leuven)
OptimizationTabularOrdinary Differential Equation
🎯 What it does: The ODECE framework is proposed, which utilizes decision-focused learning to simultaneously predict constraint parameters in prediction-optimization problems, balancing feasibility and suboptimality through adjustable weights.
FEAT: Free energy Estimators with Adaptive Transport
Yuanqi Du (Cornell University), Eric Vanden-Eijnden (Courant Institute of Mathematical Sciences, New York University)
OptimizationComputational EfficiencyReinforcement LearningTabularPhysics Related
🎯 What it does: This paper proposes the FEAT framework to estimate free energy differences through learning adaptive transport, utilizing the non-equilibrium Jarzynski equality and Crooks theorem.
Feature Distillation is the Better Choice for Model-Heterogeneous Federated Learning
Yichen Li (Huazhong University of Science and Technology), Ruixuan Li (Huazhong University of Science and Technology)
Federated LearningKnowledge DistillationImage
🎯 What it does: Proposes FedFD, a heterogeneous federated learning framework based on feature distillation and orthogonal projection;
Feature Unlearning: Theoretical Foundations and Practical Applications with Shuffling
Yue Yang (Maincode), Hao Wang (Monash University)
Supervised Fine-TuningImageTabular
🎯 What it does: Achieve a feature-level 'forgetting' function by randomly shuffling specified features in a trained model and fine-tuning on this dataset.
Feature-aware Modulation for Learning from Temporal Tabular Data
Haorun Cai, Han-Jia Ye (Nanjing University)
ClassificationDomain AdaptationTabularTime SeriesFinance Related
🎯 What it does: A feature-aware modulation mechanism based on temporal context is proposed, which achieves alignment of feature semantics under temporal drift through a learnable Yeo-Johnson transformation of feature distributions (mean, standard deviation, skewness), thereby enhancing the temporal generalization ability of tabular data.
Feature-Based Instance Neighbor Discovery: Advanced Stable Test-Time Adaptation in Dynamic World
Qinting Jiang (Tsinghua University), Zhi Wang (Tsinghua University)
Domain AdaptationTransformerImage
🎯 What it does: The FIND framework is proposed to achieve efficient adaptation during testing in dynamic multi-distribution scenarios.
FedEL: Federated Elastic Learning for Heterogeneous Devices
Letian Zhang (Middle Tennessee State University), Jie Xu (University of Florida)
Federated LearningImageTextAudio
🎯 What it does: The FedEL framework is proposed, which achieves a simultaneous improvement in training efficiency and accuracy in federated learning on heterogeneous devices through sliding window training and tensor importance adjustment.
Federated Continual Learning via Orchestrating Multi-Scale Expertise
Xiaoyang Yi (Nankai University), Jian Zhang (Nankai University)
Federated LearningKnowledge DistillationTransformerMixture of ExpertsImage
🎯 What it does: The MultiFCL framework is proposed, which achieves a dual enhancement of stability and plasticity for pre-trained models in federated continual learning through lightweight adapters, cross-modal prototype initialization, multi-scale expert training, and multi-teacher self-distillation.
Federated Dialogue-Semantic Diffusion for Emotion Recognition under Incomplete Modalities
Xihang Qiu (Shenzhen MSU-BIT University), Chun Li (Shenzhen MSU-BIT University)
RecognitionFederated LearningGraph Neural NetworkDiffusion modelMultimodality
🎯 What it does: The FedDISC framework is proposed, which combines federated learning with diffusion model-based modality recovery techniques to address the issue of missing modalities in multimodal emotion recognition.
Federated Multi-armed Bandits with Efficient Bit-Level Communications
Haoran Zhang (Nanjing University), Yang Gao (Nanjing University)
OptimizationFederated LearningTabular
🎯 What it does: This paper proposes a communication-efficient algorithm EpoInc-SE for the federated multi-armed bandit problem with heterogeneous rewards in a fully distributed communication network.
FedFACT: A Provable Framework for Controllable Group-Fairness Calibration in Federated Learning
Li Zhang (Zhejiang University), Chaochao Chen (Hangzhou Dianzi University)
OptimizationFederated LearningTabular
🎯 What it does: The FedFACT framework is proposed to achieve a controllable balance between global and local group fairness in federated learning.
FedFree: Breaking Knowledge-sharing Barriers through Layer-wise Alignment in Heterogeneous Federated Learning
Haizhou Du (Shanghai University of Electric Power), Guodong Long (University of Technology Sydney)
Federated LearningKnowledge DistillationImage
🎯 What it does: Proposes the FedFree framework to address the knowledge sharing barriers caused by model and data heterogeneity in heterogeneous federated learning, completely without the need for proxy data and model matching;
FedGPS: Statistical Rectification Against Data Heterogeneity in Federated Learning
Zhiqin Yang (Chinese University of Hong Kong), Yixuan Yuan (Chinese University of Hong Kong)
Federated LearningConvolutional Neural NetworkGaussian SplattingImage
🎯 What it does: This paper conducts large-scale experiments on various heterogeneous scenarios of federated learning, finding that existing methods lack robustness, and proposes the FedGPS framework, which integrates collaborative corrections of statistical distribution layers and gradient layers to enhance model performance.
FedIGL: Federated Invariant Graph Learning for Non-IID Graphs
Lingren Wang (Hainan University), Jingxin Liu (Hainan University)
Federated LearningGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: Proposes the FedIGL framework, which identifies and separates irrelevant subgraphs shared by clients and client-specific subgraphs through invariant learning methods in federated graph learning, enhancing graph classification and clustering performance.
FedLPA: Local Prior Alignment for Heterogeneous Federated Generalized Category Discovery
Geeho Kim (Seoul National University), Bohyung Han (Seoul National University)
Federated LearningGraph Neural NetworkContrastive LearningImage
🎯 What it does: The FedLPA framework is proposed, achieving universal category discovery of new classes without prior knowledge by constructing a local similarity graph for clients, Infomap clustering, and local prior alignment in a federated learning environment.
FedMGP: Personalized Federated Learning with Multi-Group Text-Visual Prompts
Weihao Bo (Nanjing University of Science and Technology), Zechao Li (Baidu VIS)
Federated LearningSafty and PrivacyPrompt EngineeringVision Language ModelImageMultimodality
🎯 What it does: In the context of federated learning, a personalized adaptation method for visual-text models called FedMGP is proposed, which achieves the dual goals of local personalization and global generalization while maintaining privacy.
FedQS: Optimizing Gradient and Model Aggregation for Semi-Asynchronous Federated Learning
Yunbo Li (Shanghai Jiao Tong University), Yue Wu (Shanghai Jiao Tong University)
OptimizationFederated LearningTabular
🎯 What it does: Proposes the FedQS framework, which simultaneously optimizes gradient aggregation and model aggregation in semi-asynchronous federated learning (SAFL).
FedRACE: A Hierarchical and Statistical Framework for Robust Federated Learning
Gang Yan (University of California), Wan Du (University of California)
Federated LearningImage
🎯 What it does: A joint framework named FEDRACE is proposed for robust training in a federated learning environment with frozen large model backbones, which includes a hierarchical statistical network HStat-Net and a bias-based client evaluation mechanism DevGuard.
FedRAM: Federated Reweighting and Aggregation for Multi-Task Learning
Fan Wu (Nanyang Technological University), Wei Yang Bryan Lim (Nanyang Technological University)
Federated LearningComputational EfficiencyTransformerSupervised Fine-TuningText
🎯 What it does: The FedRAM framework is proposed in federated multi-task learning, achieving task and client weighted adaptation through a three-step training process (reference model, proxy model, proxy model), enhancing both local and global performance.
FedRTS: Federated Robust Pruning via Combinatorial Thompson Sampling
Hong Huang (City University of Hong Kong), Dapeng Wu (City University of Hong Kong)
Federated LearningImageText
🎯 What it does: A robust pruning framework FedRTS based on combinatorial Thompson sampling is proposed in federated learning, which achieves adaptive adjustment of sparse network topology through the TSAdj module;
FedRW: Efficient Privacy-Preserving Data Reweighting for Enhancing Federated Learning of Language Models
Pukang Ye (East China Normal University), Yunbo Yang (East China Normal University)
Federated LearningSafty and PrivacyTransformerLarge Language ModelText
🎯 What it does: This study investigates soft deduplication (sample reweighting) for language models in federated learning to enhance performance and protect privacy.
FedSVD: Adaptive Orthogonalization for Private Federated Learning with LoRA
Seanie Lee (KAIST), Sung Ju Hwang (KAIST)
OptimizationFederated LearningSafty and PrivacyTransformerLarge Language ModelText
🎯 What it does: LoRA is introduced for pre-trained language models in federated learning, and the FedSVD method is designed to reparameterize B·A through singular value decomposition (SVD) after each round of aggregation, thereby avoiding the amplification of DP-SGD noise in matrix multiplication while allowing the A matrix to adapt over time.
FedWMSAM: Fast and Flat Federated Learning via Weighted Momentum and Sharpness-Aware Minimization
Tianle Li (Shenzhen University), Kaishun Wu (Hong Kong University of Science and Technology)
OptimizationFederated LearningConvolutional Neural NetworkImage
🎯 What it does: This paper proposes FedWMSAM, a federated learning framework that combines weighted momentum with Sharpness-Aware Minimization, aimed at addressing two major issues: local-global curvature mismatch and momentum echo oscillation in non-IID environments.
Feed-Forward Bullet-Time Reconstruction of Dynamic Scenes from Monocular Videos
Hanxue Liang (NVIDIA), Jiahui Huang (NVIDIA)
Data SynthesisOptimizationTransformerGaussian SplattingVideo
🎯 What it does: This paper presents BTimer, a real-time monocular dynamic scene modeling and novel view synthesis model based on 3D Gaussian rendering.
FEEDBACK FRICTION: LLMs Struggle to Fully Incorporate External Feedback
Dongwei Jiang (Johns Hopkins University), Daniel Khashabi (Johns Hopkins University)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelText
🎯 What it does: This paper systematically evaluates the limits of self-improvement of large models after receiving high-quality external feedback within a controlled experimental framework, and finds that even with nearly perfect feedback, the model still experiences 'feedback friction', making it difficult to fully absorb the feedback.
Feedback Guidance of Diffusion Models
Felix Koulischer (Ghent University), Luca Ambrogioni (Radboud University)
GenerationData SynthesisDiffusion modelImageText
🎯 What it does: A dynamic guidance mechanism based on state feedback (FeedBack Guidance, FBG) is proposed, which adaptively adjusts the guidance scale during the inference process of the diffusion model based on the posterior estimation of the current trajectory, thereby improving sample quality while maintaining diversity.
Feedback-Aware MCTS for Goal-Oriented Information Seeking
Harshita Chopra (University of Washington), Chirag Shah (University of Washington)
RetrievalOptimizationTransformerLarge Language ModelReinforcement LearningTextBiomedical Data
🎯 What it does: Designed and implemented the MISQ-HF framework, which combines LLM-generated information retrieval questions, uses MCTS for decision-making on the question-answer tree, and enhances dialogue efficiency and success rate through a hierarchical feedback mechanism.
FerretNet: Efficient Synthetic Image Detection via Local Pixel Dependencies
Shuqiao Liang (Jinan University), Quanlong Guan (Jinan University)
ClassificationAnomaly DetectionConvolutional Neural NetworkImage
🎯 What it does: This paper presents FerretNet, a lightweight synthetic image detection model that utilizes local pixel dependency (LPD) features to identify texture and edge anomalies in generative models.
Few-Shot Knowledge Distillation of LLMs With Counterfactual Explanations
Faisal Hamman (University of Maryland), Sanghamitra Dutta (University of Maryland)
Explainability and InterpretabilityKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: A few-shot task-aware knowledge distillation framework COD is proposed, which expands the training set by generating and utilizing minimally perturbed counterfactual explanations (CFE) for the teacher model, thereby helping the student model to more accurately approach the decision boundary of the teacher.
Few-Shot Learning from Gigapixel Images via Hierarchical Vision-Language Alignment and Modeling
Bryan Wong (KAIST), Mun Yong Yi (KAIST)
ClassificationGraph Neural NetworkPrompt EngineeringVision Language ModelImageMultimodalityBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A hierarchical audiovisual multimodal multiple instance learning framework HiVE-MIL is proposed for few-shot whole slide image classification.
FFN Fusion: Rethinking Sequential Computation in Large Language Models
Akhiad Bercovich (NVIDIA), Ran El-Yaniv (NVIDIA)
Computational EfficiencyKnowledge DistillationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Proposes the FFN Fusion technology, which merges consecutive Feed-Forward network layers into wider parallel layers, significantly reducing the sequential computation load of LLMs.
FHGS: Feature-Homogenized Gaussian Splatting
Q.G. Duan, Ben M. Chen (Chinese University of Hong Kong)
SegmentationOptimizationGaussian SplattingPoint Cloud
🎯 What it does: A Feature-Homogenized Gaussian Splatting (FHGS) framework is proposed, which embeds frozen 2D pre-trained semantic features into a 3D Gaussian splatting model to achieve real-time semantic mapping and 3D reconstruction.
FIGRDock: Fast Interaction-Guided Regression for Flexible Docking
Shikun Feng (Tsinghua University), Yanyan Lan (Tsinghua University)
Drug DiscoveryProtein Structure PredictionTabular
🎯 What it does: This paper presents FIGRDock, an interactive-guided regression-based flexible docking framework;
Filter Like You Test: Data-Driven Data Filtering for CLIP Pretraining
Mikey Shechter (Tel Aviv University), Yair Carmon (Tel Aviv University)
Data-Centric LearningContrastive LearningImageText
🎯 What it does: The FLYT algorithm is proposed, which evaluates the value of CLIP pre-training data by training a scoring model and utilizing downstream task gradients; M-FLYT is implemented to mix various scoring methods into a learnable linear model; a Soft Cap Sampling (SCS) strategy is designed to reduce the repetition of high-scoring samples and generate a high-quality filtered dataset.
Fin3R: Fine-tuning Feed-forward 3D Reconstruction Models via Monocular Knowledge Distillation
Weining Ren (University of Hong Kong), Kai Han (University of Hong Kong)
Depth EstimationKnowledge DistillationSupervised Fine-TuningPoint Cloud
🎯 What it does: Fin3R, a lightweight fine-tuning framework, has been developed that significantly enhances the detail reconstruction and robustness of feed-forward 3D reconstruction models through monocular knowledge distillation and re-normalization LoRA, while maintaining multi-view consistency by only updating the encoder.
Final-Model-Only Data Attribution with a Unifying View of Gradient-Based Methods
Dennis Wei (IBM Research), Maria Chang (IBM Research)
Explainability and InterpretabilityData-Centric LearningImageTextTabular
🎯 What it does: For scenarios where only the final model is available, a FiMO setting for training data attribution is proposed, which is reconstructed as a measurement of the model's sensitivity to training instances.
Find your Needle: Small Object Image Retrieval via Multi-Object Attention Optimization
Michael Green (Hebrew University of Jerusalem), Rami Ben-Ari (OriginAI)
Object DetectionRetrievalTransformerContrastive LearningImage
🎯 What it does: A small target image retrieval method is proposed, utilizing multi-target attention to optimize the construction of a single global feature;
Finding and Reactivating Post-Trained LLMs' Hidden Safety Mechanisms
Mingjie Li (CISPA Helmholtz Center for Information Security), Yisen Wang (Peking University)
OptimizationSafty and PrivacyRepresentation LearningAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningTextFinance Related
🎯 What it does: This paper investigates the reasons for the decline in safety performance of post-trained LLMs (especially large inference models) and proposes a lightweight method called SafeReAct to reactivate hidden safety mechanisms, thereby enhancing the model's safety under harmful prompts without significantly reducing its performance on specialized tasks.
Finding Low-Rank Matrix Weights in DNNs via Riemannian Optimization: RAdaGrad and RAdamW
Fengmiao Bian (Hong Kong University of Science and Technology), Jian-Feng Cai (Hong Kong University of Science and Technology)
CompressionOptimizationTransformerSupervised Fine-TuningImageText
🎯 What it does: A low-rank matrix weight optimization framework based on Riemannian geometry is proposed, introducing two adaptive optimizers, RAdaGrad and RAdamW, which can directly update weights on fixed-rank matrix manifolds, eliminating redundancy and condition number issues caused by factorization.
Finding separatrices of dynamical flows with Deep Koopman Eigenfunctions
Kabir Vinay Dabholkar, Omri Barak (Technion Israel Institute of Technology)
OptimizationRecurrent Neural NetworkTime SeriesSequentialPhysics RelatedOrdinary Differential Equation
🎯 What it does: A numerical method is proposed and implemented to approximate the separating surface using deep Koopman feature functions (positive eigenvalues), which can identify the boundary of the attraction domain in high-dimensional black-box dynamical systems and design minimal perturbations.
Fine-grained List-wise Alignment for Generative Medication Recommendation
Chenxiao Fan (University of Science and Technology of China), Fuli Feng (University of Science and Technology of China)
Recommendation SystemDrug DiscoveryTransformerLarge Language ModelReinforcement LearningBiomedical DataElectronic Health Records
🎯 What it does: A fine-grained list-based drug recommendation framework FLAME is constructed based on large language models, achieving safer and more accurate prescription generation through drug-by-drug decision-making and the integration of multi-source medical knowledge.
Fine-Grained Preference Optimization Improves Spatial Reasoning in VLMs
Yifan Shen (University of Illinois Urbana-Champaign), Ismini Lourentzou (University of Illinois Urbana-Champaign)
OptimizationTransformerLarge Language ModelVision Language ModelImageMultimodalityBenchmarkChain-of-Thought
🎯 What it does: We propose SpatialReasoner-R1, a visual-language model capable of generating long chain reasoning (LongCoT) from two-dimensional images, significantly enhancing fine-grained spatial reasoning abilities.
Fine-Tuning Discrete Diffusion Models with Policy Gradient Methods
Oussama Zekri (Institut Polytechnique de Paris), Nicolas Boulle
GenerationOptimizationReinforcement Learning from Human FeedbackReinforcement LearningDiffusion modelTextSequentialBiomedical Data
🎯 What it does: This paper studies a fine-tuning method for discrete diffusion models, utilizing policy gradients from reinforcement learning to directly optimize non-differentiable rewards.
FineRS: Fine-grained Reasoning and Segmentation of Small Objects with Reinforcement Learning
Lu Zhang (Dalian University of Technology), You He
Object DetectionSegmentationTransformerLarge Language ModelReinforcement LearningVision Language ModelImageMultimodality
🎯 What it does: The FINERS framework is proposed, achieving instruction-driven fine-grained reasoning and segmentation of ultra-small targets in high-resolution images using a multimodal large language model.
Finite Sample Analyses for Continuous-time Linear Systems: System Identification and Online Control
Hongyi Zhou (Tsinghua University), Jingzhao Zhang (Tsinghua University)
OptimizationReinforcement LearningTime SeriesStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: A finite sample system identification algorithm for continuous-time linear systems is proposed, and it is applied to online LQR control, achieving an expected cumulative regret of O(√T);
Finite Sample Analysis of Linear Temporal Difference Learning with Arbitrary Features
Zixuan Xie (University of Virginia), Shangtong Zhang (University of Virginia)
Reinforcement Learning
🎯 What it does: A finite sample convergence rate analysis of linear TD(λ) under arbitrary features is proposed, providing the convergence speed in L2 norm;
Finite-Sample Analysis of Policy Evaluation for Robust Average Reward Reinforcement Learning
Yang Xu (Purdue University), Vaneet Aggarwal (Purdue University)
Reinforcement Learning
🎯 What it does: A finite sample analysis of policy evaluation for robust average reward Markov decision processes is proposed.
Finite-Time Analysis of Stochastic Nonconvex Nonsmooth Optimization on the Riemannian Manifolds
Emre Sahinoglu (Northeastern University), Shahin Shahrampour (Northeastern University)
Optimization
🎯 What it does: This paper proposes two algorithms, RO2NC and ZO-RO2NC, to address the finite-time convergence problem of non-smooth non-convex stochastic optimization on Riemannian manifolds, providing a theoretical guarantee of sample complexity O(δ^{-1}ε^{-3}), and conducts numerical validation on the sparse principal component problem.
Finite-Time Bounds for Average-Reward Fitted Q-Iteration
Jongmin Lee (Seoul National University), Ernest K. Ryu (UCLA)
Reinforcement Learning
🎯 What it does: This paper proposes the Anchored Fitted Q-Iteration (Anc-F-QI) algorithm, which combines an anchoring mechanism and fitted Q-iteration to address the sample complexity issue in average reward offline reinforcement learning.
FIPER: Factorized Features for Robust Image Super-Resolution and Compression
Yang-Che Sun (National Yang Ming Chiao Tung University), Yu-Lun Liu (National Yang Ming Chiao Tung University)
Super ResolutionCompressionTransformerImage
🎯 What it does: A unified image representation called Factorized Features is proposed, which combines the decomposition of coefficients and bases, applied to single image super-resolution and image compression.
Fira: Can We Achieve Full-rank Training of LLMs Under Low-rank Constraint?
Xi Chen (Beijing Institute of Technology), Guoren Wang (Beijing Institute of Technology)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: The Fira framework is proposed, achieving full-rank training while maintaining low-rank constraints.
First Attentions Last: Better Exploiting First Attentions for Efficient Parallel Training
Gyudong Kim (Korea University), Young Geun Kim (Korea University)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Redesign the Transformer structure by using the output of the first layer's attention to replace the MHA-MLP connection in each layer, proposing two efficient models: FAL and FAL+.
First SFT, Second RL, Third UPT: Continual Improving Multi-Modal LLM Reasoning via Unsupervised Post-Training
Lai Wei (Shanghai Jiao Tong University), Lichao Sun (Lehigh University)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextMultimodality
🎯 What it does: This paper proposes an MM-UPT framework that achieves unsupervised post-training of multimodal large language models, enhancing reasoning capabilities through a self-reward mechanism based on majority voting within the GRPO algorithm.
Fisher meets Feynman: score-based variational inference with a product of experts
Diana Cai (Flatiron Institute), Lawrence K. Saul (Columbia University)
Mixture of ExpertsScore-based Model
🎯 What it does: Construct and use a product model (PoE) based on multivariate t-distribution experts for black-box variational inference, utilizing its sample-ability to achieve score-based BBVI.
Fit the Distribution: Cross-Image/Prompt Adversarial Attacks on Multimodal Large Language Models
Hai Yan (Huazhong University of Science and Technology), Pan Zhou (Huazhong University of Science and Technology)
Adversarial AttackTransformerLarge Language ModelVision Language ModelImageMultimodality
🎯 What it does: A cross-image/prompt adversarial attack method for multimodal large language models (MLLMs) is proposed, capable of generating harmful outputs on unknown combinations of images and prompts.
Fix False Transparency by Noise Guided Splatting
Aly El Hakie (OpsiClear), Yehe Liu (Case Western Reserve University)
Gaussian SplattingPoint Cloud
🎯 What it does: This paper proposes a noise-guided Gaussian polishing (NGS) method, which utilizes internal noise Gaussian as an occlusion barrier and α-consistency constraints to significantly reduce the pseudo-transparency problem in 3D Gaussian polishing.
Fixed-Point RNNs: Interpolating from Diagonal to Dense
Sajad Movahedi (ELLIS Institute), Antonio Orvieto (Max Planck Institute for Intelligent Systems)
Recurrent Neural NetworkSequential
🎯 What it does: Proposes a method to implicitly represent dense linear RNNs as fixed points of diagonal RNNs, allowing for adjustable parallelism and expressive power while keeping the number of parameters unchanged;
FLAME: Fast Long-context Adaptive Memory for Event-based Vision
Biswadeep Chakraborty (Georgia Institute of Technology), Saibal Mukhopadhyay (Georgia Institute of Technology)
Spiking Neural NetworkImage
🎯 What it does: Proposes the FLAME architecture, which utilizes event attention layers and event-aware HiPPO for efficient event-level processing of asynchronous event camera data.
Flash Invariant Point Attention
Andrew Liu (Flagship Pioneering), Olivia Viessmann (Flagship Pioneering)
Protein Structure PredictionTransformerReinforcement LearningGraphBiomedical Data
🎯 What it does: Improved IPA by rewriting it into a linear attention form compatible with FlashAttention, achieving linear memory and time complexity on GPUs, supporting structural modeling of thousands of residues.
FlashBias: Fast Computation of Attention with Bias
Haixu Wu (Tsinghua University), Mingsheng Long (Tsinghua University)
Computational EfficiencyProtein Structure PredictionTransformerLarge Language ModelImageText
🎯 What it does: This paper proposes FlashBias, an efficient computation method for the attention plus bias module using low-rank decomposition.
FlashMD: long-stride, universal prediction of molecular dynamics
Filippo Bigi (Ecole Polytechnique Federale de Lausanne), Michele Ceriotti (Ecole Polytechnique Federale de Lausanne)
Drug DiscoveryGraph Neural NetworkTransformerGraphTime SeriesPhysics RelatedStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: The FlashMD method is proposed, which directly predicts the displacement and momentum of molecular dynamics (MD) trajectories using long strides, avoiding traditional small-step integration, and supporting various thermodynamic ensembles such as isothermal and isobaric.
FlashMo: Geometric Interpolants and Frequency-Aware Sparsity for Scalable Efficient Motion Generation
Zeyu Zhang (Australian National University), Richard Hartley (Australian National University)
GenerationData SynthesisPose EstimationComputational EfficiencyTransformerDiffusion modelVideoMultimodality
🎯 What it does: This paper proposes FlashMo, a diffusion model that generates 3D human motions through frequency-aware sparsification and the MotionSiT Transformer.
FlashMoE: Fast Distributed MoE in a Single Kernel
Osayamen Jonathan Aimuyo (Cornell University), Rachee Singh (Cornell University)
OptimizationComputational EfficiencyTransformerMixture of Experts
🎯 What it does: This paper proposes FlashMoE, a method that fully integrates distributed Mixture-of-Experts (MoE) operations into a single persistent GPU kernel, addressing the low utilization and high latency issues caused by CPU scheduling, synchronous AlltoAll communication, and frequent kernel launches in traditional implementations.
Flat Channels to Infinity in Neural Loss Landscapes
Flavio Martinelli (École Polytechnique Fédérale de Lausanne), Johanni Brea (École Polytechnique Fédérale de Lausanne)
OptimizationTabularOrdinary Differential Equation
🎯 What it does: This paper systematically studies the loss landscape of neural networks and identifies and describes a special structure called 'channels to infinity', which is formed by at least two neurons whose output weights tend to infinity while their input weights approach equality.
Flatness is Necessary, Neural Collapse is Not: Rethinking Generalization via Grokking
Ting Han (Lamarr Institute), Michael Kamp (Lamarr Institute)
ClassificationTransformerSupervised Fine-TuningImageText
🎯 What it does: This paper utilizes the grooksing training framework to separate the memory phase and the generalization phase of the network, systematically evaluating the occurrence timing and impact of neural collapse (NC) and the relative flatness (RF) of the loss landscape across different models.
Flatten Graphs as Sequences: Transformers are Scalable Graph Generators
Dexiong Chen (Max Planck Institute of Biochemistry), Karsten Borgwardt (Max Planck Institute of Biochemistry)
GenerationData SynthesisDrug DiscoveryTransformerLarge Language ModelPoint CloudGraphSequential
🎯 What it does: This paper presents AUTOGRAPH, a Transformer-based autoregressive graph generation framework that transforms graphs into sequences through a reversible Segmented Eulerian Neighborhood Trail (SENT), allowing for direct graph generation using large language models.
Flattening Hierarchies with Policy Bootstrapping
John Luoyu Zhou, Jonathan Kao
Robotic IntelligenceReinforcement Learning
🎯 What it does: A method called SAW (Subgoal Advantage-Weighted Policy Bootstrapping) is proposed for offline goal-conditioned reinforcement learning, which utilizes importance sampling based on subgoal advantages to guide the learning of a single flat policy, eliminating the need for a subgoal generation model required in traditional hierarchical RL.
Flex-Judge: Text-Only Reasoning Unleashes Zero-Shot Multimodal Evaluators
Jongwoo Ko (Microsoft), Se-Young Yun (KAIST)
TransformerLarge Language ModelSupervised Fine-TuningImageVideoTextMultimodalityAudio
🎯 What it does: This paper proposes FLEX-Judge, a cross-modal evaluation model that can perform zero-shot evaluation in multimodal tasks using only a small amount of textual reasoning data.
FlexAC: Towards Flexible Control of Associative Reasoning in Multimodal Large Language Models
Shengming Yuan (University of Electronic Science and Technology of China), Lianli Gao (University of Electronic Science and Technology of China)
Large Language ModelPrompt EngineeringMultimodality
🎯 What it does: This paper proposes FlexAC, a method for controlling the intensity of multimodal large language model associative reasoning that is training-free and lightweight.
FlexEvent: Towards Flexible Event-Frame Object Detection at Varying Operational Frequencies
Dongyue Lu (National University of Singapore), Wei Tsang Ooi (National University of Singapore)
Object DetectionAutonomous DrivingConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a framework for event camera object detection at different frequencies—FlexEvent;
Flexible inference for animal learning rules using neural networks
Yuhan Helena Liu (Princeton University), Jonathan W. Pillow
Recurrent Neural NetworkReinforcement LearningTabularTime Series
🎯 What it does: This paper proposes a framework that uses neural networks to directly infer learning rules from behavioral data of animals learning tasks from scratch, modeling the learning rules as non-parametric mappings of experimental covariates, with decision policies using interpretable GLM.
Flexible Language Modeling in Continuous Space with Transformer-based Autoregressive Flows
Ruixiang ZHANG, Navdeep Jaitly (Apple)
GenerationTransformerFlow-based ModelAuto EncoderText
🎯 What it does: Proposes to transfer language models to continuous latent space, using Transformer autoregressive regularization flow for modeling.
Flexible MOF Generation with Torsion-Aware Flow Matching
Nayoung Kim (Korea Advanced Institute of Science and Technology), Sungsoo Ahn (Korea Advanced Institute of Science and Technology)
GenerationData SynthesisOptimizationTransformerFlow-based ModelGraph
🎯 What it does: A two-stage MOF generation framework called MOFFLOW-2 is proposed, which first generates metal clusters and organic ligands using SMILES, and then predicts their 3D structures (translation, rotation, torsion angles, and lattice parameters) through a flow matching model.
Flexible Realignment of Language Models
Wenhong Zhu (Shanghai Jiao Tong University), Rui Wang (Shanghai Jiao Tong University)
Computational EfficiencyKnowledge DistillationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: A flexible alignment framework is proposed, which includes training-time alignment (TrRa) and inference-time alignment (InRa), achieving efficient and controllable alignment for large language models.
FlexOLMo: Open Language Models for Flexible Data Use
Weijia Shi (Allen Institute for Artificial Intelligence), Sewon Min (Allen Institute for Artificial Intelligence)
Federated LearningSafty and PrivacyTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: Developed FLEXOLMO, a mixed expert language model capable of distributed training without data sharing, and flexible selection of expert modules during inference.
FlexSelect: Flexible Token Selection for Efficient Long Video Understanding
Yunzhuzhang, Linchao Zhu (Zhejiang University)
RecognitionComputational EfficiencyTransformerVision Language ModelVideoMultimodality
🎯 What it does: The FlexSelect framework is proposed, achieving flexible visual token selection based on VideoLLM cross-modal attention, significantly extending the processable temporal length of long videos.