NeurIPS 2025 Papers — Page 18
Conference on Neural Information Processing Systems · 5275 papers
FlexVAR: Flexible Visual Autoregressive Modeling without Residual Prediction
Siyu Jiao (Beijing Jiaotong University), ZEQUN JIE
GenerationData SynthesisTransformerAuto EncoderImage
🎯 What it does: A scalable visual autoregressive image generation framework called FlexVAR is proposed, which can generate images at any resolution, aspect ratio, and inference steps, and supports zero-shot image-to-image tasks.
FlexWorld: Progressively Expanding 3D Scenes for Flexible-View Exploration
Luxi Chen (Renmin University of China), Chongxuan Li (Renmin University of China)
GenerationData SynthesisDepth EstimationOptimizationDiffusion modelGaussian SplattingImageVideo
🎯 What it does: Achieving the generation of 3D scenes with flexible viewpoints through the gradual expansion and optimization of 3D Gaussian splatting based on a single image, utilizing a video-to-video diffusion model (V2V) to complete new viewpoint synthesis and geometrically integrate the generated content into the scene;
Flick: Empowering Federated Learning with Commonsense Knowledge
Ran Zhu (Delft University of Technology), Qing Wang (Delft University of Technology)
Data SynthesisFederated LearningTransformerLarge Language ModelVision Language ModelDiffusion modelImageText
🎯 What it does: The Flick framework is designed and implemented, which generates synthetic data by combining server-side common knowledge from large language models (LLMs) with low-sensitivity cross-client local summaries to alleviate data heterogeneity (label imbalance and domain shift) in federated learning.
Flow based approach for Dynamic Temporal Causal models with non-Gaussian or Heteroscedastic Noises
Abdellah Rahmani (Ecole Polytechnique Fédérale de Lausanne), Pascal Frossard (Ecole Polytechnique Fédérale de Lausanne)
Graph Neural NetworkFlow-based ModelTime SeriesSequential
🎯 What it does: Proposes the FANTOM framework, which jointly infers multi-interval causal graphs and interval boundaries under multi-period non-Gaussian or heteroscedastic noise.
Flow Density Control: Generative Optimization Beyond Entropy-Regularized Fine-Tuning
Riccardo De Santi (ETH Zurich), Andreas Krause (ETH Zurich)
GenerationOptimizationDrug DiscoverySupervised Fine-TuningDiffusion modelFlow-based ModelImageTabular
🎯 What it does: A generalized generative optimization framework is proposed, and the Flow Density Control (FDC) algorithm is designed for fine-tuning flow models and diffusion models to optimize arbitrary objective functions and any diversity.
Flow Equivariant Recurrent Neural Networks
T. Anderson Keller (Kempner Institute for the Study of Natural and Artificial Intelligence Harvard University)
Recurrent Neural NetworkFlow-based ModelSequential
🎯 What it does: A flow equivariant recurrent neural network (FERNN) is proposed, enabling sequence models to maintain equivariance to time-parameterized symmetric transformations.
Flow Field Reconstruction with Sensor Placement Policy Learning
Ruoyan Li (University of California), Yizhou Sun (University of California)
Graph Neural NetworkReinforcement LearningMeshTime Series
🎯 What it does: In a three-dimensional turbulent flow field, the entire flow field is reconstructed using sparse boundary sensor data, and a direction transmission perception graph neural network and a two-stage constrained PPO sensor layout strategy are proposed.
Flow Matching Neural Processes
Hussen Abu Hamad (University of Haifa), Dan Rosenbaum (University of Haifa)
GenerationData SynthesisTransformerFlow-based ModelImageTime SeriesOrdinary Differential Equation
🎯 What it does: A neural process model based on flow matching (FlowNP) is proposed, which can directly generate or estimate the conditional distribution of any target point given context points;
Flow Matching-Based Autonomous Driving Planning with Advanced Interactive Behavior Modeling
Tianyi Tan (Institute for AI Industry Research), Jingjing Liu (Institute for AI Industry Research)
Autonomous DrivingTransformerFlow-based ModelMultimodalityOrdinary Differential Equation
🎯 What it does: A flow matching-based autonomous driving planning framework called Flow Planner is proposed, focusing on interactive behavior modeling.
Flow-Based Policy for Online Reinforcement Learning
Lei Lv (Shanghai Research Institute for Intelligent Autonomous Systems), Xiao Ma (ByteDance)
Reinforcement LearningFlow-based Model
🎯 What it does: In this paper, the authors design an online reinforcement learning framework called FlowRL, which utilizes Continuous Normalizing Flows (CNF) to represent policies as state-dependent velocity fields, and achieves value-driven policy optimization through a constrained policy search with Wasserstein-2 regularization.
Flow-GRPO: Training Flow Matching Models via Online RL
Jie Liu (Chinese University of Hong Kong), Wanli Ouyang (Chinese University of Hong Kong)
GenerationReinforcement Learning from Human FeedbackReinforcement LearningFlow-based ModelImageTextStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: The Flow-GRPO method is proposed, integrating online policy gradient reinforcement learning into the flow matching model to enhance performance in text-to-image tasks.
FlowCut: Rethinking Redundancy via Information Flow for Efficient Vision-Language Models
Jintao Tong (Huazhong University of Science and Technology), Ruixuan Li (Huazhong University of Science and Technology)
OptimizationComputational EfficiencyTransformerVision Language ModelImageVideoMultimodality
🎯 What it does: Proposes the FlowCut framework, which cuts visual tokens of LVLM from the perspective of information flow, significantly reducing computational and memory burdens.
FlowDAS: A Stochastic Interpolant-based Framework for Data Assimilation
Siyi Chen (University of Michigan), Jeffrey A Fessler
GenerationData SynthesisOptimizationTime SeriesPhysics RelatedStochastic Differential Equation
🎯 What it does: A generative data assimilation framework based on stochastic interpolation, FlowDAS, is proposed for estimating the state of PDE control systems under noise and sparse observations.
FlowFeat: Pixel-Dense Embedding of Motion Profiles
Nikita Araslanov (Technical University of Munich), Daniel Cremers (Technical University of Munich)
SegmentationDepth EstimationOptical FlowVideo
🎯 What it does: A pixel-level high-resolution feature representation called FlowFeat is trained using optical flow and video data through a self-supervised approach.
FLOWING: Implicit Neural Flows for Structure-Preserving Morphing
Arthur Bizzi (University of Buenos Aires), Tiago Novello (Universidade Federal Rural de Pernambuco)
Image TranslationGenerationFlow-based ModelGaussian SplattingImagePoint CloudOrdinary Differential Equation
🎯 What it does: We propose FLOWING—a morphology transformation framework based on flow-based implicit neural representations (INR), which utilizes reversible flow models (NODE and NCF) to achieve structure-preserving, continuous, and reversible deformations between 2D images and 3D Gaussian splatting.
FlowMixer: A Depth-Agnostic Neural Architecture for Interpretable Spatiotemporal Forecasting
Fares B. Mehouachi (New York University), Saif Eddin Jabari (New York University)
Explainability and InterpretabilityComputational EfficiencyTime SeriesSequential
🎯 What it does: This paper presents FlowMixer, a single-layer neural network architecture that utilizes constrained non-negative matrix mixing and reversible mappings to capture spatio-temporal patterns while maintaining dimensionality, and supports the direct extraction of Koopman-style Kronecker-Koopman spectra.
FlowMo: Variance-Based Flow Guidance for Coherent Motion in Video Generation
Ariel Shaulov (Tel Aviv University), Hila Chefer (Tel Aviv University)
GenerationTransformerDiffusion modelVideoText
🎯 What it does: A guiding method called FlowMo is proposed for inference, which enhances the temporal consistency of videos by utilizing the temporal statistical features (patch-wise variance) of the latent space of a pre-trained text-to-video diffusion model, without additional training or introducing external motion signals.
FlowMoE: A Scalable Pipeline Scheduling Framework for Distributed Mixture-of-Experts Training
Yunqi Gao (Zhejiang University), Merouane Abdelkader DEBBAH
OptimizationComputational EfficiencyTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: FlowMoE is proposed, a scalable pipeline scheduling framework for distributed Mixture-of-Experts (MoE) training; it unifies the scheduling of multiple types of tasks (MHA computation, gating, expert computation, A2A communication), and overlaps tensor block priority scheduling with all-reduce communication, while using Bayesian optimization to automatically tune block sizes.
FlowNet: Modeling Dynamic Spatio-Temporal Systems via Flow Propagation
Yutong Feng (Hong Kong University of Science and Technology), Yuxuan Liang (Hong Kong University of Science and Technology)
Graph Neural NetworkFlow-based ModelTime Series
🎯 What it does: A physics-inspired flow network (FlowNet) is proposed, which achieves the conservation transfer of information between dynamic nodes through flow tokens, thereby predicting complex spatiotemporal systems.
FlowPrune: Accelerating Attention Flow Calculation by Pruning Flow Network
Shuo Xu (Southeast University), Xu Yang (Southeast University)
Computational EfficiencyTransformerImageText
🎯 What it does: This paper proposes the FlowPrune framework to accelerate global attention flow computation in Transformer models.
FlowRefiner: A Robust Traffic Classification Framework against Label Noise
Mingwei Zhan (Shanghai Jiao Tong University), Ke Xu (Tsinghua University)
ClassificationAnomaly DetectionAuto EncoderTime Series
🎯 What it does: Proposes the FLOWREFINER framework for classifying network traffic in label noise environments.
FLUX: Efficient Descriptor-Driven Clustered Federated Learning under Arbitrary Distribution Shifts
Dario Fenoglio (Università della Svizzera italiana), Martin Gjoreski (Università della Svizzera italiana)
Federated LearningImage
🎯 What it does: FLUX is proposed—a clustering-based federated learning framework that utilizes implicit feature descriptors on the client side to handle four types of distribution shifts simultaneously and supports automatic allocation of unlabelled new clients during testing.
Flux4D: Flow-based Unsupervised 4D Reconstruction
Jingkang Wang (University of Toronto), Raquel Urtasun (University of Toronto)
Autonomous DrivingOptimizationFlow-based ModelOptical FlowVideoPoint Cloud
🎯 What it does: Designed and implemented Flux4D, an unsupervised 4D dynamic scene reconstruction framework that completes the reconstruction of full HD multi-view driving scenes in seconds using photometric loss and a 'stay as still as possible' regularization.
FlyLoRA: Boosting Task Decoupling and Parameter Efficiency via Implicit Rank-Wise Mixture-of-Experts
Heming Zou (Tsinghua University), Xiangyang Ji (Tsinghua University)
Large Language ModelMixture of ExpertsText
🎯 What it does: This paper proposes FlyLoRA, an implicit MoE LoRA variant inspired by the olfactory circuits of moths, which uses frozen sparse random projections as built-in routers to achieve more efficient parameter tuning and multi-task integration.
FNOPE: Simulation-based inference on function spaces with Fourier Neural Operators
Guy Moss (University of Tübingen), Cornelius Schröder (University of Tübingen)
Flow-based ModelTime Series
🎯 What it does: A simulation inference method based on the Fourier Neural Operator (FNO) called FNOPE is proposed for inferring the posterior distribution of function value parameters.
FocalCodec: Low-Bitrate Speech Coding via Focal Modulation Networks
Luca Della Libera (Concordia University), Mirco Ravanelli (Concordia University)
GenerationCompressionAudio
🎯 What it does: FocalCodec is proposed, a low-bitrate single-codebook speech encoder that achieves high-quality speech reconstruction, speech conversion, and downstream tasks at 0.16–0.65 kbps.
Focus-Then-Reuse: Fast Adaptation in Visual Perturbation Environments
Jiahui Wang (Nanjing University), Yang Yu (Nanjing University)
Domain AdaptationRobotic IntelligenceReinforcement LearningVision Language ModelVideo
🎯 What it does: The Focus-Then-Reuse (FTR) method is proposed, which utilizes object-level filters to quickly deploy existing visual reinforcement learning strategies in visually disturbed environments.
FOCUS: Internal MLLM Representations for Efficient Fine-Grained Visual Question Answering
Liangyu Zhong (Technical University of Berlin), Leo Schwinn (Technical University of Munich)
RecognitionObject DetectionComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringImage
🎯 What it does: FOCUS is proposed, a training-independent visual cropping method that utilizes the similarity of the internal KV cache of MLLM to locate detail areas in images related to the question, thereby achieving efficient fine-grained visual question answering.
FOCUS: Unified Vision-Language Modeling for Interactive Editing Driven by Referential Segmentation
Fan Yang (Chinese Academy of Sciences), Jinqiao Wang (Chinese Academy of Sciences)
SegmentationGenerationTransformerLarge Language ModelVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: A unified large-scale visual language model FOCUS is proposed, integrating pixel-level segmentation perception and region-controlled generation, supporting interactive editing.
FoGE: Fock Space inspired encoding for graph prompting
Sotirios Panagiotis Chytas (University of Wisconsin-Madison), Vikas Singh (University of Wisconsin-Madison)
Graph Neural NetworkLarge Language ModelPrompt EngineeringGraph
🎯 What it does: This paper proposes a parameter-free graph encoding method based on Fock space, called FoGE, which is accessed as a prefix through a linear adapter to a frozen LLM, enabling question-answer reasoning for various graph structures (ordinary graphs, hypergraphs, protein graphs, etc.);
Follow the Energy, Find the Path: Riemannian Metrics from Energy-Based Models
Louis Béthune, Victor Boutin (Centre National de la Recherche Scientifique)
GenerationData SynthesisAuto EncoderImage
🎯 What it does: A method is proposed to directly derive Riemannian metrics from pre-trained energy-based models (EBMs) to calculate the shortest paths between data points in high-dimensional space.
Follow-the-Perturbed-Leader Nearly Achieves Best-of-Both-Worlds for the m-Set Semi-Bandit Problems
Jingxin Zhan (Peking University), Zhihua Zhang (Peking University)
OptimizationReinforcement LearningTabular
🎯 What it does: This paper studies the Follow-the-Perturbed-Leader (FTPL) strategy in the m-set semi-bandit problem, proving that it can achieve an approximate optimal regret bound of O(√nm(√d log(d) + m^(5/6))) in adversarial environments, and a regret bound of O(∑i, ∆i > 0 log(n) ∆i) in stochastic environments.
For Better or for Worse, Transformers Seek Patterns for Memorization
Madhur Panwar (École Polytechnique Fédérale de Lausanne), Antoine Bosselut (École Polytechnique Fédérale de Lausanne)
TransformerLarge Language ModelText
🎯 What it does: This study investigates how memorization occurs during the training process of Transformer language models by tracking the dynamics of memorization across different datasets and model sizes.
Force Prompting: Video Generation Models Can Learn And Generalize Physics-based Control Signals
Nate Gillman (Brown University), Chen Sun (Brown University)
GenerationData SynthesisPrompt EngineeringVideoPhysics Related
🎯 What it does: Proposes the Force Prompting method, allowing video generation models to create physics-compliant videos through physical force control (local point forces and global wind forces).
ForceFM: Enhancing Protein-Ligand Predictions through Force-Guided Flow Matching
Huanlei Guo (Southern University of Science and Technology), Bingyi Jing (Southern University of Science and Technology)
Drug DiscoveryFlow-based ModelBiomedical Data
🎯 What it does: A molecular docking model based on force-guided flow matching, ForceFM, is proposed, which can use physical energy functions as a force field guide during the generation process to produce low-energy, physically reasonable ligand conformations.
ForceVLA: Enhancing VLA Models with a Force-aware MoE for Contact-rich Manipulation
Jiawen Yu (Fudan University), Cewu Lu (Shanghai Jiao Tong University)
Robotic IntelligenceTransformerMixture of ExpertsVision-Language-Action ModelMultimodalityTime Series
🎯 What it does: The ForceVLA framework is proposed, integrating 6-axis force/torque perception as a first-class modality into the Vision-Language-Action system, and achieving multimodal fusion through FVLMoE dynamic routing to enhance the completion rate of contact-rich tasks.
Forecasting in Offline Reinforcement Learning for Non-stationary Environments
Suzan Ece Ada (Bogazici University), Erhan Oztop (Ozyegin University)
Reinforcement LearningDiffusion modelTime Series
🎯 What it does: A framework named FORL is proposed, which utilizes zero-shot forecasting models and conditional diffusion models to predict and correct unobservable temporal shifts in offline reinforcement learning, thereby maintaining efficient decision-making in non-stationary environments.
Foresight: Adaptive Layer Reuse for Accelerated and High-Quality Text-to-Video Generation
Muhammad Adnan (University of British Columbia), Prashant J. Nair (University of British Columbia)
GenerationComputational EfficiencyTransformerDiffusion modelVideoText
🎯 What it does: A training-independent adaptive layer reuse method called Foresight is proposed for the Diffusion Transformer (DiT) model in text-to-video generation, which dynamically decides whether to reuse the outputs of each layer during inference to reduce redundant computations and accelerate generation.
ForgerySleuth: Empowering Multimodal Large Language Models for Image Manipulation Detection
Zhihao Sun (Fudan University), Yu-Gang Jiang (Fudan University)
SegmentationAnomaly DetectionTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageMultimodality
🎯 What it does: Proposes the ForgerySleuth framework, utilizing multimodal large language models and low-level trace encoders to achieve image tampering detection and interpretation.
Forging Time Series with Language: A Large Language Model Approach to Synthetic Data Generation
Cécile Rousseau (IBM Research Europe), Juan Bernabe Moreno (IBM Research Europe)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningTabularTime SeriesFinance Related
🎯 What it does: The SDForger framework is proposed, utilizing large language models to generate high-quality multivariate time series in low-sample environments.
FORLA: Federated Object-centric Representation Learning with Slot Attention
Guiqiu Liao (University of Pennsylvania), Daniel A Hashimoto
Federated LearningRepresentation LearningContrastive LearningImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: This paper proposes the FORLA framework, which utilizes lightweight feature adapters and Slot Attention in federated learning to achieve unsupervised object-level representation learning, avoiding the drawbacks of data sharing and model sharing.
Formal Models of Active Learning from Contrastive Examples
Farnam Mansouri (University of Waterloo), Sandra Zilles (University of Regina)
OptimizationData-Centric LearningContrastive Learning
🎯 What it does: This paper proposes a general theoretical framework for studying the sample complexity when using contrastive examples in active learning, and provides upper and lower bounds for several classes of geometric and Boolean concept classes based on different contrastive set definitions (minimum distance model, neighborhood model).
Fortifying Time Series: DTW-Certified Robust Anomaly Detection
Shijie Liu (University of Melbourne), Sarah Monazam Erfani (University of Melbourne)
Anomaly DetectionTime Series
🎯 What it does: Designed a DTW certified robust defense for time series anomaly detection, utilizing random smoothing and the Keogh lower bound to obtain the DTW certified radius.
Foundation Cures Personalization: Improving Personalized Models’ Prompt Consistency via Hidden Foundation Knowledge
Yiyang Cai (Hong Kong University of Science and Technology), Wenhan Luo (Hong Kong University of Science and Technology)
SegmentationGenerationDiffusion modelImage
🎯 What it does: Proposes the FreeCure framework, which maintains the identity embedding while leveraging the prompt consistency knowledge of the base model to freely enhance the attribute consistency of facial personalization models.
Foundations of Top-$k$ Decoding for Language Models
Georgy Noarov (University of Pennsylvania), Edgar Dobriban (University of Pennsylvania)
TransformerLarge Language ModelText
🎯 What it does: This paper constructs an interpretable and scalable theoretical framework by viewing Top-k decoding as a Bregman divergence minimization problem for sparsifying the original LLM word distribution, and further proposes new sparse decoding methods such as α-Bregman decoding.
Fourier Analysis Network
Yihong Dong (Peking University), Jingjing Xu (ByteDance)
TextTime Series
🎯 What it does: A new neural network called FAN (Fourier Analysis Networks) is proposed to address the shortcomings of existing neural networks in modeling and reasoning about periodic phenomena.
Fourier Clouds: Fast Bias Correction for Imbalanced Semi-Supervised Learning
Jiawei Gu (Great Bay University), Ziyue Qiao (University of California)
ClassificationImage
🎯 What it does: This paper proposes a method for category bias correction in long-tail semi-supervised learning using random phase images. By subtracting the model output from reference images that contain no semantic information but maintain the true amplitude spectrum, the method estimates and corrects the model's inherent bias.
Fourier Token Merging: Understanding and Capitalizing Frequency Domain for Efficient Image Generation
Jiesong Liu (North Carolina State University), Xipeng Shen (North Carolina State University)
GenerationComputational EfficiencyDiffusion modelImage
🎯 What it does: A Token aggregation method based on Fourier transform is proposed to accelerate the generation process of diffusion models.
FP4 All the Way: Fully Quantized Training of Large Language Models
Brian Chmiel (Nvidia), Daniel Soudry (Intel)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Achieved full precision FP4 training of large language models for the first time, covering weights, activations, and gradients;
FP64 is All You Need: Rethinking Failure Modes in Physics-Informed Neural Networks
Chenhui Xu (University at Buffalo), Jinjun Xiong (University at Buffalo)
OptimizationComputational EfficiencyPhysics Related
🎯 What it does: This study investigates the 'failure modes' of Physics-Informed Neural Networks (PINN) under low precision (FP32) training, addressing the issue by increasing the arithmetic precision to double precision (FP64) and proposing the Same-Basin hypothesis, which contradicts the traditional 'loss barrier' assumption.
FPSAttention: Training-Aware FP8 and Sparsity Co-Design for Fast Video Diffusion
Akide Liu (Monash University), Bohan Zhuang (ZIP Lab)
GenerationComputational EfficiencyDiffusion modelVideo
🎯 What it does: This paper proposes FPSAttention, a training-aware module that combines FP8 quantization and sparse attention, specifically designed for the 3D bidirectional attention mechanism of video diffusion generative models, aimed at significantly improving inference speed without compromising generation quality.
FracFace: Breaking The Visual Clues—Fractal-Based Privacy-Preserving Face Recognition
Wanying Dai (Sichuan University), Jin Song Dong (National University of Singapore)
RecognitionSafty and PrivacyImage
🎯 What it does: A fractal-based frequency domain privacy-preserving face recognition framework, FracFace, is proposed, which can suppress reconstructable information while ensuring recognition accuracy.
Fractional Diffusion Bridge Models
Gabriel Nobis (Fraunhofer Heinrich Hertz Institute), Wojciech Samek (Fraunhofer Heinrich Hertz Institute)
Image TranslationProtein Structure PredictionTransformerDiffusion modelImageBiomedical DataStochastic Differential Equation
🎯 What it does: Proposes Fractional Diffusion Bridge Models (FDBM), which approximates fractional Brownian motion (fBM) as a Markov process MA-fBM to generate noise for bridge creation, enabling diffusion bridge learning for paired and unpaired data.
Fractional Langevin Dynamics for Combinatorial Optimization via Polynomial-Time Escape
Shiyue Wang (East China Normal University), Junchi Yan (Shanghai Jiao Tong University)
OptimizationReinforcement LearningGraphStochastic Differential Equation
🎯 What it does: A combination optimization method based on Fractional Lagrangian Dynamics (FLD) is proposed and implemented within a sampling and data-driven framework to address the issue of traditional LD struggling to escape narrow local optima.
FRAM: Frobenius-Regularized Assignment Matching with Mixed-Precision Computing
Binrui Shen (Beijing Normal University), Shengxin Zhu (Beijing Normal University)
OptimizationImageGraph
🎯 What it does: This paper proposes a graph matching framework called FRAM based on Frobenius regularization linear assignment (FRA), which solves the quadratic assignment problem through continuous relaxation and controls the relaxation error with adjustable parameters.
Frame Context Packing and Drift Prevention in Next-Frame-Prediction Video Diffusion Models
Lvmin Zhang (Stanford University), Maneesh Agrawala (Stanford University)
GenerationData SynthesisCompressionTransformerDiffusion modelVideo
🎯 What it does: Proposes the FramePack structure, which gradually compresses video context based on the importance of frames, thereby addressing the issues of forgetting and drift in next-frame prediction models.
Frame In-N-Out: Unbounded Controllable Image-to-Video Generation
Boyang Wang (University of Virginia), Zezhou Cheng (University of Virginia)
GenerationData SynthesisTransformerDiffusion modelImageVideoText
🎯 What it does: Proposes the Frame In-N-Out task, utilizing a boundaryless canvas to allow objects in images to exit or enter the frame, achieving more freedom in motion control.
FrameShield: Adversarially Robust Video Anomaly Detection
Mojtaba Nafez (Sharif University of Technology), Mohammad Hossein Rohban (Okinawa Institute of Science and Technology)
Anomaly DetectionGenerative Adversarial NetworkVideo
🎯 What it does: This paper proposes FrameShield, which addresses the adversarial robustness of weakly supervised video anomaly detection (WSVAD) by employing frame-by-frame adversarial training with pseudo-labels and pseudo-anomaly generation, significantly enhancing the model's detection performance under attacks.
FraPPE: Fast and Efficient Preference-Based Pure Exploration
Udvas Das (University of Lille), Debabrota Basu (University of Lille)
OptimizationTabular
🎯 What it does: This paper proposes a pure exploration algorithm for multi-objective bandits called FraPPE, which can efficiently identify all Pareto optimal arms under a given confidence level and achieve optimal sample complexity under any preference cone.
FRBNet: Revisiting Low-Light Vision through Frequency-Domain Radial Basis Network
Fangtong Sun (National University of Defense Technology), Yiying Li (National University of Defense Technology)
Object DetectionSegmentationImage
🎯 What it does: A frequency domain-based plugin module FRBNet is proposed for feature enhancement in low-light vision tasks.
Fréchet Geodesic Boosting
Yidong Zhou (University of California), Hans-Georg Müller (University of California)
Tabular
🎯 What it does: This paper proposes Fréchet geodesic boosting (FGBoost), a regression framework for gradient boosting that can operate in any uniquely defined geodesic space.
Free-Lunch Color-Texture Disentanglement for Stylized Image Generation
Jiang Qin (Harbin Institute of Technology), Joost van de Weijer (Computer Vision Center)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: This paper proposes a training-free color and texture separation image stylization method called SADis, achieving independent control over color and texture when generating images from text.
FreeControl: Efficient, Training-Free Structural Control via One-Step Attention Extraction
Jiang Lin (Nanjing University), Zili Yi (Nanjing University)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: A training-free, test-time augmentation structural control method is proposed, which injects spatial and semantic structure into the diffusion generation process by utilizing the original reference image through a single attention extraction at key time steps.
FreeInv: Free Lunch for Improving DDIM Inversion
Yuxiang Bao (Beihang University), Guoliang Kang (Beihang University)
RestorationGenerationDiffusion modelImageVideoBenchmark
🎯 What it does: This paper proposes a cost-free method for enhancing DDIM inversion called FreeInv, which achieves multi-trajectory integration by applying the same transformation to latent representations at each step, significantly reducing the trajectory deviation between inversion and reconstruction.
FreqExit: Enabling Early-Exit Inference for Visual Autoregressive Models via Frequency-Aware Guidance
Ying Li (Westlake University), Huan Wang (Westlake University)
GenerationComputational EfficiencyKnowledge DistillationImage
🎯 What it does: The FreqExit framework is proposed to achieve early exit inference for visual autoregressive models.
FreqPolicy: Efficient Flow-based Visuomotor Policy via Frequency Consistency
Yifei Su (Chinese Academy of Sciences), Jian Tang (Chinese Academy of Sciences)
Computational EfficiencyRobotic IntelligenceReinforcement LearningFlow-based ModelImagePoint Cloud
🎯 What it does: A flow matching visual motion strategy named FreqPolicy is proposed, which achieves one-shot action generation through frequency domain consistency constraints, significantly improving inference speed and success rate.
FreqPolicy: Frequency Autoregressive Visuomotor Policy with Continuous Tokens
Yiming Zhong (ShanghaiTech University), Yuexin Ma (ShanghaiTech University)
Robotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningDiffusion modelMultimodality
🎯 What it does: The FreqPolicy proposes a frequency-domain visual-motor strategy that utilizes hierarchical frequency domain modeling and continuous tokens to achieve coarse-to-fine action generation.
Frequency-Aware Token Reduction for Efficient Vision Transformer
Dong-Jae Lee (Korea Advanced Institute of Science and Technology), Junmo Kim (Korea Advanced Institute of Science and Technology)
Computational EfficiencyRepresentation LearningTransformerImage
🎯 What it does: A frequency-aware Token reduction method is proposed to reduce the computational complexity of self-attention in Vision Transformers while avoiding rank collapse issues caused by excessive smoothing.
FRN: Fractal-Based Recursive Spectral Reconstruction Network
Ge Meng (Xiamen University), Xinghao Ding (Xiamen University)
RestorationGenerationSuper ResolutionConvolutional Neural NetworkImage
🎯 What it does: A spectral reconstruction network based on fractal recursion (FRN) is proposed, which can recursively generate hyperspectral images from broadband to narrowband step by step.
From Average-Iterate to Last-Iterate Convergence in Games: A Reduction and Its Applications
Yang Cai (Yale University), Weiqiang Zheng (Yale University)
OptimizationReinforcement Learning
🎯 What it does: A black-box simplification technique A2L is proposed, which transforms the average iteration of any decoupled learning algorithm into its latest iteration, achieving optimal final iteration convergence in self-play games;
From Black-box to Causal-box: Towards Building More Interpretable Models
Inwoo Hwang (Causal Artificial Intelligence Lab Columbia University), Elias Bareinboim (Causal Artificial Intelligence Lab Columbia University)
Explainability and InterpretabilityImage
🎯 What it does: The concept of causal interpretability is proposed, and a graph-based criterion is constructed to determine the interpretability of predictive models when satisfying given counterfactual queries, along with the maximum feature set for implementing interpretable models.
From Bytes to Ideas: Language Modeling with Autoregressive U-Nets
Mathurin VIDEAU, David Lopez-Paz (Meta AI)
TransformerLarge Language ModelText
🎯 What it does: Proposes a self-regressive U-Net architecture that directly processes raw bytes and dynamically aggregates them at multiple levels to form word-level and multi-word-level embeddings, eliminating traditional fixed tokenizers and large-scale embedding tables;
From Condensation to Rank Collapse: A Two-Stage Analysis of Transformer Training Dynamics
Zheng-An Chen (Shanghai Jiao Tong University), Tao Luo (Shanghai Jiao Tong University)
TransformerText
🎯 What it does: This study investigates the training dynamics of Transformers under small initialization and proposes a two-phase analytical framework from condensation to rank collapse.
From Contextual Combinatorial Semi-Bandits to Bandit List Classification: Improved Sample Complexity with Sparse Rewards
Liad Erez (Tel Aviv University), Tomer Koren (Google Research)
Reinforcement Learning
🎯 What it does: This paper studies the Contextual Combinatorial Semi-Bayesian (CCSB) problem with sparse rewards, providing an analysis of sample complexity and scheduling complexity for PAC learning and online loss minimization.
From Counterfactuals to Trees: Competitive Analysis of Model Extraction Attacks
Awa Khouna (Polytechnique Montreal), Thibaut Vidal (Polytechnique Montreal)
Explainability and InterpretabilityAdversarial AttackTabular
🎯 What it does: This paper proposes a functional equivalence extraction attack (TRA) for decision trees and tree ensemble models using local optimal counterfactual explanations, and provides a theoretical analysis of its query complexity and competitive ratio.
From Cradle to Cane: A Two-Pass Framework for High-Fidelity Lifespan Face Aging
Tao Liu (Nankai University), Yaxing Wang (Nankai University)
Image TranslationGenerationDiffusion modelGenerative Adversarial NetworkImage
🎯 What it does: A two-stage facial age transformation framework called Cradle2Cane is proposed, utilizing a few-step text-image diffusion model to achieve high-fidelity facial aging across the entire life cycle.
From Dormant to Deleted: Tamper-Resistant Unlearning Through Weight-Space Regularization
Shoaib Ahmed Siddiqui (University of Cambridge), Eleni Triantafillou (Google DeepMind)
ClassificationData-Centric LearningConvolutional Neural NetworkImage
🎯 What it does: This study investigates the anti-tampering capability of machine unlearning in visual classification models, finding that existing methods can restore the performance of forgotten samples by fine-tuning only on the retained set after refinement, and proposes a new unlearning method based on weight space regularization.
From Euler to AI: Unifying Formulas for Mathematical Constants
Tomer Raz (Technion Israel Institute of Technology), Ido Kaminer (Technion Israel Institute of Technology)
OptimizationTransformerLarge Language ModelPrompt EngineeringTextPhysics Related
🎯 What it does: A complete automation framework has been developed to extract, verify, and unify infinite series, continued fractions, and other formulas of π and other constants using large language models, constructing a unified Conservative Matrix Field (CMF) structure, and proving the equivalence of formulas through the UMAPS algorithm.
From Experts to a Generalist: Toward General Whole-Body Control for Humanoid Robots
Yuxuan Wang (Peking University), Zongqing Lu (Peking University)
Robotic IntelligenceTransformerReinforcement LearningAuto EncoderText
🎯 What it does: The BumbleBee framework is designed to achieve full-body control from expert to general. It first uses an autoencoder and text to cluster movements, trains experts for each cluster, and then completes sim-to-real through iterative delta action compensation. Finally, knowledge distillation is performed using a Transformer to obtain a unified general controller.
From Faults to Features: Pretraining to Learn Robust Representations against Sensor Failures
Jens U. Brandt (TH Köln), Thomas Bartz-Beielstein (TH Köln)
Anomaly DetectionAutonomous DrivingRepresentation LearningTransformerSupervised Fine-TuningMultimodalityTime Series
🎯 What it does: A self-supervised mask pre-training scheme is proposed, specifically simulating sensor faults (such as bias, drift, hard faults, noise, etc.) and training the model to reconstruct the original signal, thereby learning representations that are robust to sensor failures.
From Flat to Hierarchical: Extracting Sparse Representations with Matching Pursuit
Valérie Costa (École Polytechnique Fédérale de Lausanne), Demba E. Ba (Harvard University)
Representation LearningAuto EncoderMultimodality
🎯 What it does: This study investigates the limitations of Sparse Autoencoders (SAE) in interpreting neural network representations and proposes and validates the Matching Pursuit Autoencoder (MP-SAE) to capture hierarchical, nonlinear, and multimodal features.
From Forecasting to Planning: Policy World Model for Collaborative State-Action Prediction
Zhida Zhao (Dalian University of Technology), Huchuan Lu (Dalian University of Technology)
Autonomous DrivingOptimizationTransformerGenerative Adversarial NetworkWorld ModelVideoMultimodality
🎯 What it does: This paper proposes the Policy World Model (PWM), a unified driving world model that can perform action-independent video prediction of future states and directly plan trajectories based on the predicted future states.
From Human Attention to Diagnosis: Semantic Patch-Level Integration of Vision-Language Models in Medical Imaging
Dmitry Lvov (Artificial Intelligence Institute Innopolis University), Ilya Pershin (Artificial Intelligence Institute Innopolis University)
ClassificationRecognitionTransformerVision Language ModelMultimodalityBiomedical DataComputed Tomography
🎯 What it does: The LogitGaze-Med framework is proposed for semantic-based gaze prediction in medical imaging.
From Indicators to Insights: Diversity-Optimized for Medical Series-Text Decoding via LLMs
Xiyuan Jin (Beijing Jiaotong University), Youfang Lin (Beijing Jiaotong University)
ClassificationRecognitionOptimizationTransformerLarge Language ModelPrompt EngineeringTextTime SeriesBiomedical DataElectrocardiogram
🎯 What it does: A multimodal text and time series joint decoding framework based on clinical decision indicators, InDiGO, is proposed, utilizing LLM for efficient analysis of medical time series.
From Information to Generative Exponent: Learning Rate Induces Phase Transitions in SGD
Konstantinos Christopher Tsiolis (University of Toronto), Murat A Erdogdu
OptimizationTabular
🎯 What it does: This paper studies the impact of the learning rate in gradient descent on sample complexity in single-index high-dimensional model learning, and proposes a unified framework to analyze phase transitions under different gradient update methods. It also introduces an 'Alternating SGD' algorithm that achieves non-correlation enhancement through two-step updates with different learning rates. Theoretical proofs demonstrate that under different learning rates, one can transition from information index to generative index, achieving near-linear sample complexity.
From Judgment to Interference: Early Stopping LLM Harmful Outputs via Streaming Content Monitoring
Yang Li (Institute of Computing Technology, Chinese Academy of Sciences), Juan Cao (Institute of Computing Technology, Chinese Academy of Sciences)
GenerationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes a real-time streaming content monitoring framework that can terminate harmful outputs in advance during the generation of large language models.
From Kolmogorov to Cauchy: Shallow XNet Surpasses KANs
Xin Li (Dongguan University of Technology), Zhihong Jeff Xia
OptimizationReinforcement LearningPhysics Related
🎯 What it does: This study investigates a shallow XNet neural network architecture based on the Cauchy integral formula, demonstrating its effectiveness in function approximation, solving partial differential equations, and reinforcement learning.
From Likelihood to Fitness: Improving Variant Effect Prediction in Protein and Genome Language Models
Charles W. J. Pugh (Centre for Genomic Regulation), Jonathan Frazer (Centre for Genomic Regulation)
Drug DiscoveryTransformerLarge Language ModelBiomedical Data
🎯 What it does: Proposed and implemented the Likelihood-Fitness Bridging (LFB) method to improve the prediction of mutation effects in protein and genomic language models.
From Linear to Nonlinear: Provable Weak-to-Strong Generalization through Feature Learning
Junsoo Oh (KAIST), Chulhee Yun (KAIST)
OptimizationRepresentation LearningConvolutional Neural NetworkImage
🎯 What it does: This study investigates the weak-to-strong generalization phenomenon when training a strong model (two-layer ReLU CNN) using pseudo-labels from a weak model (linear CNN) after pre-training, providing theoretical analysis and experimental validation.
From Noise to Narrative: Tracing the Origins of Hallucinations in Transformers
Praneet Suresh (Mila Quebec AI Institute), Danilo Bzdok (Mila Quebec AI Institute)
TransformerAuto EncoderText
🎯 What it does: Detecting and quantifying the risk of hallucinations produced by the model under different input noise by training a Sparse Autoencoder (SAE) on the internal activations of the Transformer.
From Pixels to Views: Learning Angular-Aware and Physics-Consistent Representations for Light Field Microscopy
Feng He (University of Science and Technology of China), Quan Wen (University of Science and Technology of China)
RestorationRepresentation LearningTransformerImageBiomedical DataBenchmark
🎯 What it does: Proposes XLFM-Former, a 3D reconstruction framework for light field microscopy based on Swin Transformer;
From Pose to Muscle: Multimodal Learning for Piano Hand Muscle Electromyography
RUOFAN LIU, Hideki Koike (Tokyo Institute of Technology)
RecognitionPose EstimationTransformerContrastive LearningMultimodalityBiomedical Data
🎯 What it does: This paper presents PianoKPM Net—a multimodal network that infers hand electromyography (EMG) from hand posture and piano key strike actions, and releases the largest professional pianist EMG dataset, the PianoKPM Dataset.
From Pretraining to Pathology: How Noise Leads to Catastrophic Inheritance in Medical Models
Hao Sun (Shandong University), Yilong Yin (King Abdullah University of Science and Technology)
ClassificationRepresentation LearningConvolutional Neural NetworkTransformerImageBiomedical Data
🎯 What it does: This paper systematically studies the catastrophic inheritance effect of label noise during the pre-training phase on the performance of medical foundation models in out-of-distribution tasks, and proposes to restore the structure of the representation space by regularizing the skewness and kurtosis of the feature layers.
From Programs to Poses: Factored Real-World Scene Generation via Learned Program Libraries
Joy Hsu (Stanford University), Niloy Mitra
GenerationData SynthesisPose EstimationTransformerLarge Language ModelPoint Cloud
🎯 What it does: The FactoredScenes framework is proposed, which decomposes indoor scenes into two parts: a program library and hierarchical pose prediction. It utilizes LLM to learn reusable spatial structure functions and generates layouts based on this, then assigns real poses to objects through a program-conditioned pose prediction model, ultimately retrieving 3D objects to construct realistic scenes.
From Replication to Redesign: Exploring Pairwise Comparisons for LLM-Based Peer Review
Yaohui Zhang (Stanford University), Weixin Liang (Stanford University)
Large Language ModelText
🎯 What it does: This paper designs and evaluates a peer review mechanism for paper comparison based on a large language model (LLM) agent, replacing traditional absolute scoring;
From Self-Check to Consensus: Bayesian Strategic Decoding in Large Language Models
Weitong Zhang (Imperial College London), Bernhard Kainz (Friedrich-Alexander University Erlangen-Nürnberg)
GenerationOptimizationTransformerLarge Language ModelReinforcement LearningAgentic AITextBiomedical Data
🎯 What it does: A Bayesian Decoding Game (BDG) based on game theory is proposed, allowing LLMs to achieve self-checking and consensus through the interaction between the generator and the verifier during the generation process.
From Sequence to Structure: Uncovering Substructure Reasoning in Transformers
Xinnan Dai (Michigan State University), Jiliang Tang (Michigan State University)
TransformerLarge Language ModelPrompt EngineeringTextGraph
🎯 What it does: This paper studies how to enable a decoder-only Transformer to extract substructures in a graph structure through text sequences, and proposes the Induced Substructure Filtration (ISF) to explain the mechanism of hierarchical identification of substructures.
From Shortcut to Induction Head: How Data Diversity Shapes Algorithm Selection in Transformers
Ryotaro Kawata (University of Tokyo), Denny Wu (Flatiron Institute)
TransformerText
🎯 What it does: This study investigates the learning mechanisms of Transformer under the influence of pre-training data diversity, focusing on the trigger-output copying task. Theoretical and experimental results demonstrate the transfer of induced heads and positional shortcuts.
From Softmax to Score: Transformers Can Effectively Implement In-Context Denoising Steps
Paul Rosu (Duke University), Xiang Cheng (Duke University)
RestorationOptimizationComputational EfficiencyTransformerDiffusion modelScore-based ModelImage
🎯 What it does: This paper proposes that the Transformer can implement various denoising algorithms during forward inference, including manifold-based Laplacian denoising, precise score-based diffusion denoising, and learnable anisotropic diffusion;
From Specificity to Generality: Revisiting Generalizable Artifacts in Detecting Face Deepfakes
Long Ma (University of Science and Technology of China), Hui Lin (China Academy of Electronics and Information Technology)
ClassificationRecognitionAnomaly DetectionConvolutional Neural NetworkAuto EncoderContrastive LearningImageVideo
🎯 What it does: A unified detection framework based on the universal features of deep forgery is proposed, focusing on two types of common forgery traces: Facial Inconsistency (FIA) and Upsampling Artifacts (USA);