ICLR 2026 Papers — Page 18
International Conference on Learning Representations · 5356 papers
FIRE: Frobenius-Isometry Reinitialization for Balancing the Stability–Plasticity Tradeoff
Isaac Han, KyungJoong Kim
OptimizationConvolutional Neural NetworkTransformerReinforcement LearningImageText
🎯 What it does: Propose a novel reinitialization method called FIRE (Frobenius-Isometry REinitialization), achieving a balance between stability and plasticity by solving a constrained optimization problem;
First is Not Really Better Than Last: Evaluating Layer Choice and Aggregation Strategies in Language Model Data Influence Estimation
Dmytro Vitel (University of South Florida), Anshuman Chhabra (University of South Florida)
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: Studied the layer selection and aggregation strategies of influence functions in large language models (LLM), and proposed new aggregation methods and evaluation metrics.
Fisher-Rao Sensitivity for Out-of-Distribution Detection in Deep Neural Networks
Anthony Nguyen (Paris-Saclay University), Franck FLORIN
Anomaly DetectionConvolutional Neural NetworkTransformerImage
🎯 What it does: From the perspective of information geometry, we study the local Fisher-Rao sensitivity of deep networks and propose an OoD detection method based on statistical manifolds.
Fixing the Broken Compass: Diagnosing and Improving Inference-Time Reward Modeling
Jiachun Li (University of Chinese Academy of Sciences), Jun Zhao (University of Chinese Academy of Sciences)
Computational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposed a reward model-based reasoning acceleration algorithm called CRISP, which enhances LLM inference effectiveness through answer clustering, reward aggregation, and step-by-step prefix guidance.
FLARE: Fully Integration of Vision-Language Representations for Deep Cross-Modal Understanding
Zheng Liu (Peking University), Wentao Zhang
Representation LearningTransformerLarge Language ModelVision Language ModelDiffusion modelMultimodalityBenchmark
🎯 What it does: Developed FLARE, a Vision-Language Model (VLM) that achieves deep vision-language fusion through text-guided visual encoding and context-aware alignment decoding.
Flash-Mono: Feed-Forward Accelerated Gaussian Splatting Monocular SLAM
Zicheng Zhang (Fudan University), Wenchao Ding (Fudan University)
OptimizationComputational EfficiencyTransformerGaussian SplattingSimultaneous Localization and MappingImage
🎯 What it does: Propose Flash-Mono, a real-time monocular Gaussian Splatting SLAM system that utilizes a recursive feedforward model to directly predict camera poses and pixel-level 2D Gaussian attributes, with the backend performing only lightweight optimization, significantly improving speed;
Flash-Searcher: Fast and Effective Web Agents via DAG-Based Parallel Execution
Tianrui Qin (OPPO AI Center), Wangchunshu Zhou (OPPO Reseach Institute)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelAgentic AITextTabularBenchmark
🎯 What it does: Propose the Flash-Searcher framework, which utilizes a DAG structure to decompose complex tasks into parallelizable subtasks, achieving efficient Web Agent reasoning through multi-path parallel inference and tool calls.
FlashDLM: Accelerating Diffusion Language Model Inference via Efficient KV Caching and Guided Diffusion
Zhanqiu Hu (Cornell University), Udit Gupta (Cornell University)
Computational EfficiencyTransformerLarge Language ModelDiffusion modelText
🎯 What it does: Propose two training-agnostic acceleration methods: FreeCache uses KV caching to reduce redundant computations in DLM inference, while Guided Diffusion employs a lightweight AR model to provide consistency guidance during the diffusion process, thereby reducing denoising steps and improving parallel generation efficiency.
FlashVID: Efficient Video Large Language Models via Training-free Tree-based Spatiotemporal Token Merging
Ziyang Fan (Harbin Institute of Technology), Zhuotao Tian (Harbin Institute of Technology)
CompressionComputational EfficiencyTransformerVision Language ModelVideoMultimodality
🎯 What it does: Proposed FlashVID, a training-agnostic, plug-and-play acceleration framework that leverages attention + diversity screening and tree-structured spatiotemporal merging to significantly compress video visual tokens, achieving efficient inference.
FlashWorld: High-quality 3D Scene Generation within Seconds
Xinyang Li (Xiamen University), Liujuan Cao (Tencent)
GenerationData SynthesisTransformerDiffusion modelGaussian SplattingImageVideoTextMultimodality
🎯 What it does: Propose FlashWorld, a 3D scene generation framework based on dual-mode pre-training and cross-modal fine-tuning;
Flatness Guided Test-Time Adaptation for Vision-Language Models
Aodi Li (University of Science and Technology of China), Shafei Wang (Peng Cheng Laboratory)
Domain AdaptationTransformerPrompt EngineeringVision Language ModelContrastive LearningMultimodality
🎯 What it does: Propose a visual-language model test-time adaptation framework FGA based on the flatness of the loss landscape. First, during the training phase, Sharpness-Aware Prompt Tuning is used to find flat minima. Then, during testing, Sharpness-based Test Sample Selection is employed to select augmented samples aligned with the training flat regions, avoiding parameter updates to improve out-of-distribution (OOD) generalization.
Flatter Tokens are More Valuable for Speculative Draft Model Training
Jiaming Fan (Southeast University), Xu Yang (Southeast University)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelText
🎯 What it does: Studied speculative decoding for accelerating large language model inference, proposing a data selection method based on 'flatness' to enhance the training efficiency of the draft model.
Flattery, Fluff, and Fog: Diagnosing and Mitigating Idiosyncratic Biases in Preference Models
Anirudh Bharadwaj (University of Pennsylvania), Mark Yatskar (New York University)
Data-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Systematically evaluated the preference imbalance of language model preference models under five types of biases (length, structure, terminology, flattery, and vagueness), and reduced these biases through post-training using adversarial data augmentation (CDA).
FlexHiNM-GP: Flexible Hierarchical Pruning via Region Allocation and Channel Permutation
YI XIAODIE (Sungkyunkwan University), Dongkun Shin (Sungkyunkwan University)
Computational EfficiencyTransformerImageTextBenchmark
🎯 What it does: Propose the FlexHiNM-GP framework, combining three-layer sparsification (vector layer, N:M layer, all-zero layer) with Gyro-Permutation channel rotation to achieve adaptive sparsity for different network regions and hardware-friendly execution.
FlexiCodec: A Dynamic Neural Audio Codec for Low Frame Rates
Jiaqi Li (Chinese University of Hong Kong), Zhizheng Wu (Chinese University of Hong Kong)
CompressionConvolutional Neural NetworkTransformerAuto EncoderGenerative Adversarial NetworkAudio
🎯 What it does: Developed a dynamic neural audio codec called FlexiCodec that can achieve high-quality audio encoding at extremely low frame rates (3~12.5 Hz).
FlexiVoice: Enabling Flexible Style Control in Zero-Shot TTS with Natural Language Instructions
Dekun Chen (Chinese University Of Hong Kong Shenzhen), Zhizheng Wu
GenerationLarge Language ModelReinforcement LearningTextMultimodalityAudio
🎯 What it does: A TTS system named FlexiVoice was constructed, which can flexibly control speaking style through natural language instructions without relying on zero-copy speech, and adjust voice characteristics via optional reference speech; by leveraging a large-scale instruction speech dataset (FlexiVoice-Instruct) and the Progressive Post-Training (PPT) strategy, the model achieves high-quality, interpretable audio synthesis under multimodal inputs (text, instructions, reference speech);
FlexLinearAttention: Compiling a Unified Abstraction into Scalable Kernels for Linear Attention
Haojie Duanmu (ByteDance Seed), Dahua Lin (Shanghai AI Laboratory)
Computational EfficiencyTransformerBenchmark
🎯 What it does: Propose FlexLA, a domain-specific compiler for linear attention, which can automatically compile high-level PyTorch code into high-performance, scalable distributed kernels.
FlexLoRA: Entropy-Guided Flexible Low-Rank Adaptation
Muqing Liu (Southeast University), Yuheng Jia (Southeast University)
Computational EfficiencyImageTextBenchmark
🎯 What it does: Proposed a flexible low-rank adaptation framework called FlexLoRA, which dynamically increases or decreases the rank of LoRA modules while keeping most of the model's weights frozen, thereby improving parameter efficiency;
FlexRibbon: Joint Sequence and Structure Pretraining for Protein Modeling
Jianwei Zhu (Zhongguancun Academy), Tao Qin (Zhongguancun Academy)
Protein Structure PredictionTransformerDiffusion modelBiomedical Data
🎯 What it does: Propose FlexRibbon, a protein foundation model that integrates sequence and three-dimensional structure, capable of simultaneously generating protein sequences and full-atom structures under single-sequence conditions;
Flipping the Dialogue: Training and Evaluating User Language Models
Tarek Naous (Microsoft Research), Jennifer Neville (Microsoft Research)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Train and evaluate specialized user language models (User LMs) by flipping dialogues (swapping user and assistant utterances) and simulating multi-turn human user behavior through high-level intent conditioning.
FLoC: Facility Location-Based Efficient Visual Token Compression for Long Video Understanding
Janghoon Cho (Qualcomm AI Research), Sungha Choi (Qualcomm AI Research)
CompressionVision Language ModelVideo
🎯 What it does: Proposes a visual token compression framework FLoC based on the facility location function, which selects representative and diverse visual tokens within a given token budget using the lazy greedy algorithm.
Flock: A Knowledge Graph Foundation Model via Learning on Random Walks
Jinwoo Kim (KAIST), Ismail Ilkan Ceylan (TU Wien AITHYRA)
Representation LearningRecurrent Neural NetworkGraph
🎯 What it does: Propose FLOCK, a knowledge graph foundation model that utilizes random walks and maintains probabilistic node-relation equivariance, for zero-shot link prediction.
floq: Training Critics via Flow-Matching for Scaling Compute in Value-Based RL
Bhavya Kumar Agrawalla (Carnegie Mellon University), Aviral Kumar (Carnegie Mellon University)
Reinforcement LearningFlow-based ModelBenchmarkOrdinary Differential Equation
🎯 What it does: Developed the floq method, parameterizing the Q-function as a time-varying velocity field, jointly trained with flow matching and TD targets, leveraging multi-step numerical integration to achieve scalable computational capacity;
FLoRG: Federated Fine-tuning with Low-rank Gram Matrices and Procrustes Alignment
Chuiyang Meng (University of British Columbia), Vincent W.S. Wong (University of British Columbia)
Federated LearningComputational EfficiencyRepresentation LearningSupervised Fine-TuningText
🎯 What it does: Proposed a federated fine-tuning framework called FLoRG based on a single low-rank Gram matrix, which eliminates the aggregation errors of traditional LoRA by aggregating Gram matrices.
Flow Actor-Critic for Offline Reinforcement Learning
Jongseong Chae (KAIST), Youngchul Sung (KAIST)
Flow-based ModelImageBenchmark
🎯 What it does: Proposed a flow model integrated offline reinforcement learning algorithm called Flow Actor-Critic (FAC), which simultaneously utilizes the flow behavior proxy to regularize the actor and penalize the critic, and trains the first-order flow actor and dual-tower critic within the same architecture.
Flow Along the $K$-Amplitude for Generative Modeling
weitao Du, Shengchao Liu (University of California, Berkeley)
GenerationTransformerFlow-based ModelAuto EncoderImage
🎯 What it does: Propose K-Flow, a generative model that performs flow matching in the K-amplitude space, combining Fourier, Wavelet, and PCA decompositions to achieve multi-scale frequency domain generation and editing.
Flow Autoencoders are Effective Protein Tokenizers
Rohit Dilip (California Institute of Technology), David Van Valen (California Institute of Technology)
Protein Structure PredictionTransformerDiffusion modelFlow-based ModelAuto EncoderBiomedical Data
🎯 What it does: Proposed and implemented an autoencoder called Kanzi based on flow matching, which discretizes protein structures into discrete tokens and reconstructs them.
Flow Caching for Autoregressive Video Generation
Yuexiao Ma (Xiamen University), Rongrong Ji (Xiamen University)
GenerationComputational EfficiencyFlow-based ModelVideoBenchmark
🎯 What it does: Propose the FlowCache framework to achieve efficient caching and KV compression for autoregressive video generation models
Flow Map Learning Via Non-Gradient Vector Flow
Mark Goldstein (Flatiron Institute), Rajesh Ranganath (Flatiron Institute)
GenerationConvolutional Neural NetworkDiffusion modelFlow-based ModelImageStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This study proposes the SGFlow method, which directly learns the flow map of probabilistic flow ODEs, eliminating the need for numerical integration during inference and significantly reducing sampling latency.
Flow Matching Policy Gradients
David McAllister (University of California Berkeley), Angjoo Kanazawa (University of California Berkeley)
OptimizationReinforcement LearningFlow-based ModelOrdinary Differential Equation
🎯 What it does: Proposes Flow Policy Optimization (FPO), an online RL algorithm that integrates flow matching into policy gradients, enabling the training of flow generation models (e.g., diffusion models) as continuous action policies.
Flow Matching with Injected Noise for Offline-to-Online Reinforcement Learning
Yongjae Shin, Youngchul Sung
Reinforcement LearningFlow-based ModelBenchmarkOrdinary Differential Equation
🎯 What it does: Propose a hybrid offline-online reinforcement learning method called FINO, combining flow matching with noise injection, which expands the policy distribution through noise injection and achieves efficient exploration via entropy-guided sampling.
Flow Matching with Semidiscrete Couplings
Alireza Mousavi-Hosseini (Apple), marco cuturi
GenerationSuper ResolutionFlow-based ModelImage
🎯 What it does: Proposes using semi-discrete optimal transport (SD-OT) to pre-estimate potential vectors matching noise and data, and rapidly matches new noise to data points during flow matching (FM) training via maximum inner product search (MIPS), eliminating the high computational cost of batch OT; simultaneously provides unbiased convergence criteria, theoretical convergence analysis, and a generalization of Tweedie's formula, extending to guidance and consistency models.
Flow of Spans: Generalizing Language Models to Dynamic Span-Vocabulary via GFlowNets
Bo Xue (Shanghai Jiao Tong University), Zhouhan Lin (Shanghai Jiao Tong University)
GenerationData SynthesisTransformerLarge Language ModelReinforcement LearningFlow-based ModelTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes FoSS, a span generation framework based on generative flow networks (GFlowNets), which constructs a directed acyclic graph (DAG) state space using a dynamic span vocabulary to achieve multi-path generation.
Flow Straight and Fast in Hilbert Space: Functional Rectified Flow
Jianxin Zhang (University of Michigan), Clayton Scott (University of Michigan)
GenerationData SynthesisTransformerFlow-based ModelRectified FlowImagePhysics RelatedOrdinary Differential Equation
🎯 What it does: Proposed and implemented Functional Rectified Flow (FRF) in infinite-dimensional Hilbert space, achieving a generative model from discrete to continuous.
Flow-based Conformal Prediction for Multi-dimensional Time Series
Junghwan Lee (Georgia Institute of Technology), Yao Xie (Georgia Institute of Technology)
TransformerFlow-based ModelTime Series
🎯 What it does: Propose a conformal prediction method based on guided flow and classifier-free guidance, constructing adaptive and shape-flexible prediction sets for multi-dimensional time series.
Flow-Based Single-Step Completion for Efficient and Expressive Policy Learning
Prajwal Koirala (Iowa State University), Cody Fleming (Iowa State University)
Reinforcement LearningFlow-based ModelBenchmark
🎯 What it does: Proposed a single-step completion strategy (SSCP), achieving one-time action generation through flow-matching and completion vectors, and built SSCQL (offline RL) and GC-SSCP (goal-conditioned RL) algorithms based on this strategy.
Flow-Disentangled Feature Importance
Xingshu Chen (Sun Yat-sen University), Jin-Hong Du (University of Hong Kong)
Explainability and InterpretabilityFlow-based ModelBiomedical Data
🎯 What it does: Propose a general, model-agnostic feature importance evaluation framework called FDFI, which can accurately quantify feature contributions under feature correlations and provide confidence intervals;
Flow2GAN: Hybrid Flow Matching and GAN with Multi-Resolution Network for Few-step High-Fidelity Audio Generation
Zengwei Yao (Xiaomi Corp.), Daniel Povey (Xiaomi Corp.)
GenerationConvolutional Neural NetworkFlow-based ModelGenerative Adversarial NetworkAudio
🎯 What it does: Proposed Flow2GAN, a two-stage training framework that first learns generation capabilities using improved Flow Matching, then refines them with GAN to obtain high-quality audio generators for one, two, or four steps.
FlowAD: Ego-Scene Interactive Modeling for Autonomous Driving
Mingzhe Guo (Baidu Inc), Zhipeng Zhang (Shanghai Jiao Tong University)
Autonomous DrivingConvolutional Neural NetworkRecurrent Neural NetworkTransformerOptical FlowVideoPoint Cloud
🎯 What it does: Proposed a self-driven framework FlowAD based on scene flow, enhancing perception, planning, and VLM tasks through the introduction of ego-scene interaction modeling.
FlowAlign: Trajectory-Regularized, Inversion-Free Flow-based Image Editing
Jeongsol Kim (KAIST), Jong Chul Ye (KAIST)
Image TranslationImage HarmonizationDiffusion modelFlow-based ModelGaussian SplattingImageVideoBenchmark
🎯 What it does: Proposes FlowAlign, a flow model image editing framework that does not require a reverse process, utilizing endpoint similarity regularization to smooth the editing trajectory while maintaining structural consistency with the source image.
FlowBind: Efficient Any-to-Any Generation with Bidirectional Flows
Yeonwoo Cha (KAIST), Seunghoon Hong (KAIST)
GenerationVision Language ModelFlow-based ModelImageTextMultimodalityPoint CloudAudio
🎯 What it does: Developed a flow model called FlowBind for arbitrary subset multimodal generation, capable of generating other modalities given any subset of modalities;
FlowCast: Advancing Precipitation Nowcasting with Conditional Flow Matching
Bernardo Perrone Ribeiro (University of Ljubljana), Jana Faganeli Pucer (University of Ljubljana)
GenerationComputational EfficiencyTransformerFlow-based ModelAuto EncoderImageTime SeriesOrdinary Differential Equation
🎯 What it does: Developed FlowCast, a full-probability rainfall nowcasting model that leverages Conditional Flow Matching (CFM) to directly generate high-dimensional rainfall images from noise in a compressed latent space, achieving fast sampling.
FlowCast: Trajectory Forecasting for Scalable Zero-Cost Speculative Flow Matching
Divya Jyoti Bajpai (Indian Institute of Technology Bombay), Manjesh Kumar Hanawal
GenerationComputational EfficiencyFlow-based ModelImageVideoTextOrdinary Differential Equation
🎯 What it does: Proposes FlowCast, a zero-cost inference framework based on Flow-Matching, which significantly accelerates the generation process by speculating through self-speed prediction and verifying at runtime;
Flower: A Flow-Matching Solver for Inverse Problems
Mehrsa Pourya (École Polytechnique Fédérale de Lausanne), Michael Unser (École Polytechnique Fédérale de Lausanne)
RestorationSuper ResolutionFlow-based ModelRectified FlowImage
🎯 What it does: Proposed a flow-based inverse problem solver named Flower, which generates posterior samples that satisfy measurement constraints through a three-step iterative process (flow-consistent target estimation, measurement-aware correction, and temporal progression) using a pre-trained flow model.
FlowGen: Synthesizing Diverse Flowcharts to Enhance and Benchmark MLLM Reasoning
Kaiwen Shi (Nanjing University), Gong Cheng (Nanjing University)
Data SynthesisLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodalityGraphBenchmark
🎯 What it does: Constructed a controllable flowchart generator named FlowGen, and used its generated large-scale multi-rendered, multi-structural samples to train and test the performance of multimodal large language models (MLLMs) on flowchart parsing and question-answering tasks.
Flowing Through States: Neural ODE Regularization for Reinforcement Learning
Mohamed Ghanem (CISPA Helmholtz Center for Information Security), Bernd Finkbeiner (CISPA Helmholtz Center for Information Security)
Reinforcement LearningOrdinary Differential Equation
🎯 What it does: Proposes FlowReg, a method that smooths state embeddings in reinforcement learning and aligns them with environmental dynamics through neural ODE regularization
FlowNIB: An Information Bottleneck Analysis of Bidirectional vs. Unidirectional Language Models
Md Kowsher (University of Central Florida), Chen Chen (University of Central Florida)
Explainability and InterpretabilityRepresentation LearningTransformerContrastive LearningText
🎯 What it does: This paper proposes the FlowNIB framework, which jointly estimates hierarchical mutual information between bidirectional and unidirectional language models using the information bottleneck theory.
FlowRL: Matching Reward Distributions for LLM Reasoning
Xuekai Zhu (Shanghai Jiao Tong University), Zhouhan Lin (Shanghai Jiao Tong University)
AI Code AssistantReinforcement LearningFlow-based ModelTextSequentialBenchmarkChain-of-Thought
🎯 What it does: Propose a novel reinforcement learning method called FlowRL, which converts rewards into target distributions using flow balance and achieves diverse reasoning paths through inverse KL divergence matching strategies.
FlowSearcher: Synthesizing Memory-Guided Agentic Workflows for Web Information Seeking
Keyi Xiang (Agency for Science, Technology and Research), Haiyan Yin (Agency for Science, Technology and Research)
RetrievalLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Designed and implemented the FlowSearcher framework, utilizing memory-driven agent workflow synthesis for web information retrieval, replacing the traditional ReAct linear workflow;
FlowSymm: Physics–Aware, Symmetry–Preserving Graph Attention for Network Flow Completion
Ege Demirci (UC Santa Barbara), Ambuj Singh (UC Santa Barbara)
OptimizationGraph Neural NetworkGraphPhysics Related
🎯 What it does: Developed a physics-constrained graph attention network called FLOWSYMM for recovering missing traffic flows in networks while maintaining flow conservation.
Fluent Alignment with Disfluent Judges: Post-training for lower-resource languages
David Samuel (University of Oslo), Andrey Kutuzov (University of Oslo)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Proposed a self-adversarial reinforcement learning method for post-training on low-resource languages, which can maintain model fluency without using target language instruction data.
FLUX-Reason-6M & PRISM-Bench: A Million-Scale Text-to-Image Reasoning Dataset and Comprehensive Benchmark
Rongyao Fang (Chinese University of Hong Kong), Hongsheng Li (Chinese University of Hong Kong)
Data SynthesisVision Language ModelDiffusion modelImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Constructed the FLUX-Reason-6M dataset comprising 6 million images and 20 million bilingual descriptions, and designed the seven-track PRISM-Bench evaluation benchmark.
Fly-CL: A Fly-Inspired Framework for Enhancing Efficient Decorrelation and Reduced Training Time in Pre-trained Model-based Continual Representation Learning
Heming Zou (Tsinghua University), Xiangyang Ji (Tianjin University)
Computational EfficiencyRepresentation LearningConvolutional Neural NetworkTransformerImage
🎯 What it does: A continuous learning framework named Fly-CL based on the moth olfactory circuit was constructed, which uses sparse random projection and top-k operations to decorrelate features from a frozen pre-trained model, and then adopts streaming ridge regression for efficient class prototype learning.
FlyPrompt: Brain-Inspired Random-Expanded Routing with Temporal-Ensemble Experts for General Continual Learning
Hongwei Yan (Tsinghua University), Yi Zhong (Tsinghua University)
TransformerPrompt EngineeringMixture of ExpertsImage
🎯 What it does: Designed and implemented a general continual learning framework called FlyPrompt based on the fruit fly olfactory memory system, addressing the core challenges of expert routing and enhancing expert capabilities.
FM4NPP: A Scaling Foundation Model for Nuclear and Particle Physics
David Keetae Park (Brookhaven National Laboratory), Yihui Ren (Brookhaven National Laboratory)
Representation LearningContrastive LearningPoint CloudPhysics Related
🎯 What it does: Proposed and trained an expandable self-supervised foundation model, FM4NPP, for processing sparse spatial point data from particle detectors, achieving multiple downstream tasks (track finding, particle identification, noise labeling) by freezing weights and combining with a lightweight adapter.
FMIP: Joint Continuous-Integer Flow For Mixed-Integer Linear Programming
Hongpei Li (Shanghai University of Finance and Economics), Yinyu Ye (Princeton University)
OptimizationGraph Neural NetworkFlow-based ModelTabular
🎯 What it does: Proposed a joint continuous-integer flow generation framework named FMIP for predicting high-quality solutions to mixed integer linear programming (MILP), incorporating a global guidance mechanism during inference;
FOCUS: Efficient Keyframe Selection for Long Video Understanding
Zirui Zhu (National University of Singapore), Yang You (National University of Singapore)
Computational EfficiencyVision Language ModelVideoMultimodalityBenchmark
🎯 What it does: Proposed a training-agnostic, plug-and-play keyframe selection module called FOCUS, which helps multimodal large language models reduce the number of visual tokens when processing ultra-long videos.
Follow-Your-Preference: Towards Preference-Aligned Image Inpainting
Yutao Shen (University Of Tokyo), Jack Ma
RestorationGenerationDiffusion modelFlow-based ModelImage
🎯 What it does: This study, without altering the model structure, utilizes Direct Preference Optimization (DPO) and public reward models to construct preference data, aligning BrushNet and FLUX.1 Fill with preferences to improve image restoration effects;
Follow-Your-Shape: Shape-Aware Image Editing via Trajectory-Guided Region Control
Zeqian Long (University of Illinois Urbana-Champaign), Yue Ma (HKUST)
Image TranslationDiffusion modelFlow-based ModelImageBenchmark
🎯 What it does: Propose a training- and mask-free image editing framework that enables large-scale object shape transformation while preserving the background intact.
Following the Navigation: Enhancing Small Language Models Contextual Reasoning with LLM Guidance
Xiaoqi Ni (University of Science and Technology of China), Jianye HAO
Computational EfficiencyKnowledge DistillationLarge Language ModelTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose a training-agnostic Navigation framework that compresses the reasoning experience of large models into reusable templates, helping small models better locate and utilize key information in context reasoning tasks.
FoNE: Precise Single-Token Number Embeddings via Fourier Features
Tianyi Zhou (University of Southern California), Vatsal Sharan (University of Southern California)
Representation LearningTransformerTextTabular
🎯 What it does: Propose a single-token numerical embedding method called FoNE based on Fourier features, achieving precise arithmetic reasoning on a small Transformer trained from scratch.
Fore-Mamba3D: Mamba-based Foreground-Enhanced Encoding for 3D Object Detection
Zhiwei Ning (Shanghai Jiao Tong University), Wei Liu (Shanghai Jiao Tong University)
Object DetectionAutonomous DrivingPoint Cloud
🎯 What it does: Proposes Fore-Mamba3D, a 3D object detection backbone network focusing on foreground voxels, combining foreground sampling, region-to-global sliding window, and semantic-geometric fusion.
Foresight Diffusion: Improving Sampling Consistency in Predictive Diffusion Models
Yu Zhang (Tsinghua University), Mingsheng Long (Tsinghua University)
GenerationRobotic IntelligenceTransformerDiffusion modelVideoTime SeriesPhysics Related
🎯 What it does: Proposes the Foresight Diffusion framework, decoupling conditional understanding from the target denoising process, and enhancing sampling consistency of predictive diffusion models through a pre-trained deterministic prediction flow.
Forest-Based Graph Learning for Semi-Supervised Node Classification
Jin Li (Hong Kong University of Science and Technology), Hui Xiong (Hong Kong University of Science and Technology)
ClassificationGraph Neural NetworkGraph
🎯 What it does: Proposes a graph learning framework FGL based on forests for semi-supervised node classification, utilizing multiple spanning trees to achieve global information propagation.
ForestPersons: A Large-Scale Dataset for Under-Canopy Missing Person Detection
Deokyun Kim (ETRI), Jihun Cha (KAIST)
Object DetectionConvolutional Neural NetworkImageBenchmark
🎯 What it does: Proposed the ForestPersons large-scale dataset, collecting 96,482 low-altitude aerial perspective images under forest canopies, annotated with 204,078 occluded person instances.
Forget Forgetting: Continual Learning in a World of Abundant Memory
Dongkyu Cho (New York University), Sungmin Cha (New York University)
Computational EfficiencyImageText
🎯 What it does: Proposes the Weight Space Consolidation method to balance stability and plasticity in continuous learning scenarios with sufficient VRAM.
Forget Many, Forget Right: Scalable and Precise Concept Unlearning in Diffusion Models
Kaiyuan Deng (University of Arizona), Xiaolong Ma (University of Arizona)
GenerationDiffusion modelImage
🎯 What it does: Propose the ScaPre framework to achieve large-scale concept forgetting, utilizing closed-form updates in diffusion models to precisely erase multiple target concepts.
Formal Mechanistic Interpretability: Automated Circuit Discovery with Provable Guarantees
Itamar Hadad (Hebrew University of Jerusalem), Shahaf Bassan (Hebrew University of Jerusalem)
OptimizationExplainability and InterpretabilityImageBenchmark
🎯 What it does: Proposes an automatic circuit discovery framework based on neural network verification, which can provide provable robustness and various minimization guarantees on continuous input domains and patch domains.
Formalising Human-in-the-Loop: Computational Reductions, Failure Modes, and Legal-Moral Responsibility
Maurice Chiodo (University of Cambridge), John Burden (University of Cambridge)
ClassificationExplainability and Interpretability
🎯 What it does: Proposes a formal framework based on reducibility for classifying human-in-the-loop (HITL) systems, and conducts a systematic analysis by integrating failure modes with legal liability.
FormalML: A Benchmark for Evaluating Formal Subgoal Completion in Machine Learning Theory
Xiao-Wen Yang (Nanjing University), Xiaoxing Ma (Nanjing University)
TransformerLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: This paper proposes a benchmark called FormalML for evaluating the completion of formal subgoals in machine learning theory and constructs 4,937 subgoal problems;
Forward-Learned Discrete Diffusion: Learning how to noise to denoise faster
Grigory Bartosh (University of Amsterdam), Javier Zazo (Microsoft Research)
GenerationData SynthesisReinforcement LearningDiffusion modelImageTextGraph
🎯 What it does: Proposes Forward-Learned Discrete Diffusion (FLDD), achieving few-step discrete diffusion generation by learning a variable non-Markovian forward noise process;
Fostering Video Reasoning via Next-Event Prediction
Haonan Wang (National University of Singapore), Tianyu Pang (National University of Singapore)
Representation LearningData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelContrastive LearningVideoMultimodalityBenchmark
🎯 What it does: Propose the Next Event Prediction (NEP) learning task, construct the V1-33K video dataset and FutureBench evaluation benchmark, and investigate how NEP enhances the temporal reasoning capabilities of multimodal large models.
Foundation Models for Causal Inference via Prior-Data Fitted Networks
Yuchen Ma (LMU Munich), Stefan Feuerriegel (LMU Munich)
TransformerTabular
🎯 What it does: Construct and train a base model called CausalFM based on Prior-Data Fitted Networks (PFNs) for Bayesian inference in various causal inference scenarios (backdoor, frontdoor, and instrumental variable).
Foundation Visual Encoders Are Secretly Few-Shot Anomaly Detectors
Guangyao Zhai (Technical University of Munich), Benjamin Busam (Technical University of Munich)
Data SynthesisAnomaly DetectionRepresentation LearningTransformerContrastive LearningImage
🎯 What it does: This paper proposes FOUNDAD, a few-shot anomaly detection method based on a foundational visual encoder, which identifies abnormal regions by mapping features back to the natural image manifold through a non-linear projector.
Foundational Automatic Evaluators: Scaling Multi-Task Generative Evaluator Training for Reasoning-Centric Domains
Austin Xu, Shafiq Joty (Salesforce AI Research)
Data-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmark
🎯 What it does: By constructing a multi-task, multi-domain evaluation dataset containing 2.5M samples and adopting an iterative rejection sampling supervised fine-tuning (RS-SFT) method, we trained general-purpose automatic evaluators FARE with 8B and 20B parameter scales;
FRABench and UFEval: Unified Fine-grained Evaluation with Task and Aspect Generalization
Shibo Hong (Fudan University), Yixin Cao (Fudan University)
TransformerLarge Language ModelSupervised Fine-TuningImageTextMultimodalityBenchmark
🎯 What it does: Propose UFEval, a unified fine-grained evaluator that supports four categories of multimodal tasks and achieves task and dimension generalization.
Fractional-Order Spiking Neural Network
Chengjie Ge (University Of Science And Technology Of China), Zheng-Jun Zha (Nanyang Technological University)
Spiking Neural NetworkImageVideoGraphOrdinary Differential Equation
🎯 What it does: Propose an Spiking Neural Network framework f-SNN that replaces integer-order ODEs with fractional-order ODEs, leveraging fractional-order calculus to capture the long-term memory characteristics of neurons.
Fracture-GS: Dynamic Fracture Simulation with Physics-Integrated Gaussian Splatting
Xiaogang Wang (Southwest University), Kai Xu (Beijing Information Science and Technology University)
GenerationGaussian SplattingImagePhysics Related
🎯 What it does: Proposes a unified framework for simulating and rendering extreme mechanical collisions and dynamic fractures, named Fracture‑GS, which can achieve implicit reconstruction of colliding objects, particle sampling, collision MPM computation, and high-quality Gaussian rendering of fractured particles from multi-view images.
FragFM: Hierarchical Framework for Efficient Molecule Generation via Fragment-Level Discrete Flow Matching
Joongwon Lee (KAIST), Woo Youn Kim (KAIST)
GenerationDrug DiscoveryFlow-based ModelAuto EncoderGraph
🎯 What it does: Propose FragFM, a hierarchical molecular graph generation framework that first generates molecules at the fragment level and then losslessly reconstructs them to the atomic level using a fine-grained autoencoder.
Frame Guidance: Training-Free Guidance for Frame-Level Control in Video Diffusion Models
Sangwon Jang (KAIST), Sung Ju Hwang (DeepAuto.ai)
GenerationDiffusion modelAuto EncoderVideo
🎯 What it does: Propose a training-agnostic Frame Guidance framework that achieves controllable video generation by applying gradient guidance to a few frames, supporting multiple conditions such as keyframes, style, looping, depth, and sketches.
FrameThinker: Learning to Think with Long Videos via Multi-Turn Frame Spotlighting
Zefeng He (Shanghai AI Laboratory), Yu Cheng (Chinese University of Hong Kong)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelVision-Language-Action ModelVideoChain-of-Thought
🎯 What it does: Developed a long video reasoning framework called FrameThinker that proactively selects frames in multiple rounds and performs reasoning.
Frayed RoPE and Long Inputs: A Geometric Perspective
Davis Wertheimer (IBM Research), Naigang Wang (IBM Research)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Investigate the geometric mechanism behind RoPE's failure in long sequences, propose the RoPE-ID scheme to achieve parameter-free long-context reasoning.
FREAK: A Fine-grained Hallucination Evaluation Benchmark for Advanced MLLMs
Zhihan Yin (Wangxuan Institute of Computer Technology, Peking University), Dongyan Zhao (State Key Laboratory of General Artificial Intelligence)
Large Language ModelImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Constructed FREAK—a fine-grained hallucination evaluation benchmark for multimodal large language models (MLLMs), focusing on the misrecognition and fabrication of details in images;
Free Energy Mixer
Jiecheng Lu (Georgia Institute of Technology), Shihao Yang (Georgia Institute of Technology)
ClassificationComputational EfficiencyTransformerImageTextTime Series
🎯 What it does: Propose Free Energy Mixer (FEM) to replace standard attention for implementing per-channel value-aware posterior readout.
Free Lunch for Stabilizing Rectified Flow Inversion
Chenru Wang (Westlake University), Chi Zhang (Westlake University)
GenerationFlow-based ModelRectified FlowImageBenchmarkOrdinary Differential Equation
🎯 What it does: Proposed a training-free, plug-and-play velocity field correction method, Proximal-Mean Inversion (PMI) and mimic-CFG, enhancing the inverse and editing performance of Rectified Flow generative models.
FreeAdapt: Unleashing Diffusion Priors for Ultra-High-Definition Image Restoration
Xiaoan Liu (Wuhan University), Dongdong Yue (Wuhan University)
RestorationSuper ResolutionSupervised Fine-TuningDiffusion modelAuto EncoderImage
🎯 What it does: Propose the FreeAdapt framework, which utilizes pre-trained latent diffusion models to achieve ultra-high-resolution image restoration.
FreeKV: Boosting KV Cache Retrieval for Efficient LLM Inference
Guangda Liu (Shanghai Jiao Tong University), Jieru Zhao (Shanghai Jiao Tong University)
Computational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposes the FreeKV framework, combining algorithmic and system-level improvements to achieve an efficient, training-free solution for KV cache retrieval.
FreeViS: Training-free Video Stylization with Inconsistent References
Jiacong Xu (Johns Hopkins University), Vishal M. Patel (Johns Hopkins University)
Image TranslationGenerationTransformerVision Language ModelDiffusion modelOptical FlowImageVideoTextMultimodality
🎯 What it does: Propose the FreeViS framework to achieve training-free video style transfer, leveraging multiple inconsistent references, indirect high-frequency compensation, and optical flow guidance to enhance style details and temporal consistency.
FreqKV: Key-Value Compression in Frequency Domain for Context Window Extension
Jushi Kai (Shanghai Jiao Tong University), Zhouhan Lin (Shanghai Jiao Tong University)
CompressionComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Proposed a method to compress the KV cache of large language models in the frequency domain to achieve context window expansion.
Frequency Bands in RoPE: Base Frequency and Context Length Shape the Interpolation–Extrapolation Trade-off
Yui Oka (NTT, Inc.), Kuniko Saito (NTT, Inc.)
TransformerLarge Language ModelText
🎯 What it does: This paper analyzes the frequency bands of RoPE, investigates the patterns generated by changes in the base frequency θ and training length L_train, and verifies their performance under different LLMs and position interpolation methods.
Frequency-aware Dynamic Gaussian Splatting
Qiaowei Miao (Zhejiang University), Yawei Luo (Zhejiang University)
GenerationOptimizationComputational EfficiencyNeural Radiance FieldGaussian SplattingVideo
🎯 What it does: Address motion blur in 4D reconstruction caused by conflicts between high-frequency details and high-frequency motion.
Frequency-Balanced Retinal Representation Learning with Mutual Information Regularization
Seunghoon Lee (VUNO Inc), Doohyun Park (VUNO Inc)
Representation LearningAuto EncoderImageBiomedical Data
🎯 What it does: Propose a retinal image pre-training framework named RetMAE, which learns frequency-balanced visual features by incorporating high-frequency mutual information (MI) regularization into the Masked Autoencoder (MAE).
Frequency-Domain Better than Time-Domain for Causal Structure Recovery in Dynamical Systems on Networks
Mohammed Tuhin Rana (University of Minnesota), Murti Salapaka
Computational EfficiencyGraphTime Series
🎯 What it does: Studied the use of frequency-domain Wiener filter (WF) for causal structure recovery in dynamic networks, and proposed a fast FFT computation method and the Wiener-Phase algorithm based on WF phase.
Fresh in memory: Training-order recency is linearly encoded in language model activations
Dmitrii Krasheninnikov (University of Cambridge), David Krueger (Mila, University of Montreal)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
🎯 What it does: Observing the training order in the linear encoding of model activation through sequential fine-tuning of Llama-3.2-1B.
FRIEDA: Benchmarking Multi-Step Cartographic Reasoning in Vision-Language Models
Jiyoon Pyo (University of Minnesota Twin Cities), Yao-Yi Chiang (University of Minnesota Twin Cities)
Large Language ModelVision Language ModelMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Constructed the FRIEDA benchmark, focusing on multi-map, multi-step map reasoning tasks to evaluate the ability of large vision-language models in reading geographic information charts.
From ``Sure" to ``Sorry": Detecting Jailbreak in Large Vision Language Model via JailNeurons
Yuyou Gan (Zhejiang University), Shouling Ji (Zhejiang University)
Anomaly DetectionSafty and PrivacyVision Language ModelImageTextMultimodality
🎯 What it does: Proposes a method (JDJN) for detecting jailbreak attacks in vision-language models by locating 'JailNeuron' and aggregating multi-layer hidden states
From Abstract to Contextual: What LLMs Still Cannot Do in Mathematics
Bowen Cao (Chinese University of Hong Kong), Furu Wei (Microsoft)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBenchmark
🎯 What it does: Propose the ContextMATH benchmark to systematically evaluate the ability of large language models in contextual (narrative) mathematical reasoning, analyzing the model's performance in extracting and solving mathematical problems from scenarios.
From Assistant to Independent Developer — Are GPTs Ready for Software Development?
Dezhi Ran (Peking University), Tao Xie (Peking University)
AI Code AssistantTransformerLarge Language ModelAgentic AITextBenchmark
🎯 What it does: This paper proposes and implements APPFORGE, a benchmark for evaluating the ability of large language models to complete real Android application development from scratch.
From Assumptions to Actions: Turning LLM Reasoning into Uncertainty-Aware Planning for Embodied Agents
SeungWon Seo (Korea University), HyeongYeop Kang (Korea University)
Large Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: Proposed the Planner-Composer-Evaluator (PCE) framework, which structures implicit assumptions generated by LLMs into decision trees, and selects actions with or without communication by evaluating scene probability, gain, and cost, thereby improving planning performance in decentralized partially observable multi-agent environments.
From atom to space: A region-based readout function for spatial properties of materials
Jiawen Zou (Fudan University), Bo Yan (Fudan University)
Graph Neural NetworkTransformerGraphPhysics Related
🎯 What it does: Proposed a spatial node-based readout function called SpatialRead, improving the prediction of spatially separable attributes (e.g., gas adsorption, pore size) by graph neural networks.