ICLR 2025 Papers — Page 13
International Conference on Learning Representations · 3704 papers
Feedback Schrödinger Bridge Matching
Panagiotis Theodoropoulos (Georgia Institute of Technology), Evangelos Theodorou
Image TranslationDomain AdaptationImageStochastic Differential Equation
🎯 What it does: Proposed the Feedback Schrödinger Bridge Matching (FSBM) algorithm, which uses a small number of aligned samples to guide the distribution matching of non-aligned samples.
Fengbo: a Clifford Neural Operator pipeline for 3D PDEs in Computational Fluid Dynamics
Alberto Pepe (Huawei Research Center Germany & Austria), Joan Lasenby (University of Cambridge)
Explainability and InterpretabilityComputational EfficiencyConvolutional Neural NetworkPoint CloudPhysics Related
🎯 What it does: A neural operator pipeline called Fengbo, which is completely based on 3D Clifford algebra, is proposed for efficiently predicting pressure and velocity fields from geometric inputs in computational fluid dynamics.
Ferret-UI 2: Mastering Universal User Interface Understanding Across Platforms
Zhangheng LI, Zhe Gan (Apple)
RecognitionGenerationData SynthesisTransformerLarge Language ModelVision Language ModelImageMultimodality
🎯 What it does: A multi-platform universal UI understanding multimodal large language model, Ferret-UI 2, has been developed, supporting various screen types such as iPhone, Android, iPad, Webpage, and AppleTV, and achieving single-step user intent interaction.
Few for Many: Tchebycheff Set Scalarization for Many-Objective Optimization
Xi Lin (City University of Hong Kong), Qingfu Zhang (City University of Hong Kong)
OptimizationTabular
🎯 What it does: This paper proposes a method based on Tchebycheff set scalarization (TCH-Set) and its smoothed version (STCH-Set) to find only a few (e.g., 5) solutions in multi-objective optimization to cover a large number of objectives (e.g., 100+).
Few-Class Arena: A Benchmark for Efficient Selection of Vision Models and Dataset Difficulty Measurement
Bryan Bo Cao (Stony Brook University), Shubham Jain (Stony Brook University)
ClassificationRecognitionComputational EfficiencyContrastive LearningImageBenchmark
🎯 What it does: Proposes the Few-Class Arena benchmark and SimSS difficulty measurement for evaluating the performance of visual models in few-class scenarios.
Fewer May Be Better: Enhancing Offline Reinforcement Learning with Reduced Dataset
Yiqin Yang (Chinese Academy of Sciences), Bo XU
Reinforcement Learning
🎯 What it does: This study addresses the problem of data subset selection in offline reinforcement learning, proposing the ReDOR method which constructs a smaller effective dataset through gradient approximation and orthogonal matching pursuit to enhance algorithm performance.
Fictitious Synthetic Data Can Improve LLM Factuality via Prerequisite Learning
Yujian Liu (University of California Santa Barbara), Yang Zhang (MIT IBM Watson AI Lab)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: A fine-tuning strategy named PREREQ-TUNE is designed and implemented, which first learns the prerequisite knowledge required for the task through knowledge LoRA, then freezes that LoRA and uses skill LoRA to learn the skills of the task itself, thereby separating the learning of knowledge and skills, reducing the probability of hallucinations in LLMs.
Fiddler: CPU-GPU Orchestration for Fast Inference of Mixture-of-Experts Models
Keisuke Kamahori (University of Washington), Baris Kasikci (University of Washington)
OptimizationComputational EfficiencyTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: A hybrid CPU-GPU inference system named Fiddler is proposed, specifically designed for Mixture-of-Experts (MoE) language models in resource-constrained environments. It dynamically allocates expert layers to CPU or GPU to minimize inference latency and efficiently utilize limited GPU memory.
Field-DiT: Diffusion Transformer on Unified Video, 3D, and Game Field Generation
Kangfu Mei (Johns Hopkins University), Vishal M. Patel (Johns Hopkins University)
GenerationData SynthesisTransformerDiffusion modelImageVideoTextMultimodality
🎯 What it does: This paper proposes a unified diffusion field transformer (Field-DiT) that can generate videos, 3D views, and action-based game content through view sampling and autoregressive generation within the same network architecture.
FIG: Flow with Interpolant Guidance for Linear Inverse Problems
Yici Yan (University of Illinois), Zhizhen Zhao (University of Illinois)
RestorationSuper ResolutionDiffusion modelFlow-based ModelImageStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes a method called FIG, which uses a pre-trained flow matching or diffusion model to guide inverse sampling through measurement interpolation, addressing linear image inverse problems.
Filtered not Mixed: Filtering-Based Online Gating for Mixture of Large Language Models
Raeid Saqur (University of Toronto), Frank Rudzicz (Dalhousie University)
Large Language ModelMixture of ExpertsTime SeriesFinance Related
🎯 What it does: An online gating mechanism MoE-F based on continuous-time filters is proposed, capable of dynamically combining multiple pre-trained LLM experts for time series prediction.
Finally Rank-Breaking Conquers MNL Bandits: Optimal and Efficient Algorithms for MNL Assortment
Aadirupa Saha (University of Illinois), Pierre Gaillard (Univ. Grenoble Alpes)
OptimizationTabular
🎯 What it does: This study investigates the Assortment Optimization problem and proposes an efficient algorithm without strong default item constraints within the framework of the Multinomial Logit (MNL) model.
Find A Winning Sign: Sign Is All We Need to Win the Lottery
Junghun Oh (Seoul National University), Kyoung Mu Lee (Seoul National University)
OptimizationConvolutional Neural NetworkImage
🎯 What it does: An improved learning rate reset (LRR) method named AWS is proposed to find sparse subnetworks that can be trained from any random initialization to performance comparable to that of fully parameterized networks, and it is demonstrated that the signed mask can transfer the generalization potential of the subnetwork to new initialized networks.
Finding Shared Decodable Concepts and their Negations in the Brain
Cory Daniel Efird, Alona Fyshe (University of Alberta)
Vision Language ModelContrastive LearningImageMultimodalityMagnetic Resonance Imaging
🎯 What it does: By mapping brain fMRI signals to the multimodal embedding space of CLIP through contrastive learning and applying an improved DBSCAN clustering to cross-subject linear decoding parameters, shared decodable semantic concepts (SDC) were discovered.
Fine-Grained Verifiers: Preference Modeling as Next-token Prediction in Vision-Language Alignment
Chenhang Cui (National University of Singapore), Tat-Seng Chua (National University of Singapore)
GenerationRecommendation SystemOptimizationTransformerReinforcement LearningVision Language ModelMultimodality
🎯 What it does: Proposes the FiSAO self-alignment method, which utilizes the visual encoder of VLLM to perform fine-grained scoring for each generated token and fine-tunes the model through reinforcement learning to enhance visual-language alignment;
Fine-Tuning Attention Modules Only: Enhancing Weight Disentanglement in Task Arithmetic
Ruochen Jin (East China Normal University), Li Shen (University of Pennsylvania)
TransformerSupervised Fine-TuningImageText
🎯 What it does: Only fine-tune the attention module in the Transformer to improve weight decoupling and model performance in arithmetic tasks.
Fine-tuning can Help Detect Pretraining Data from Large Language Models
Hengxiang Zhang (Southern University of Science and Technology), Hongxin Wei (Southern University of Science and Technology)
Anomaly DetectionTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This study investigates a method for detecting pre-training data by fine-tuning large language models with a small amount of unseen data, using score function deviation.
Fine-Tuning Discrete Diffusion Models via Reward Optimization with Applications to DNA and Protein Design
Chenyu Wang (Massachusetts Institute of Technology), Tommaso Biancalani (Genentech)
OptimizationDrug DiscoveryReinforcement LearningDiffusion modelSequentialBiomedical Data
🎯 What it does: Reward optimization fine-tuning of pre-trained discrete diffusion models is performed, proposing the DRAKES algorithm, which enables the model to generate sequences that meet task objectives while maintaining a natural distribution.
Fine-tuning with Reserved Majority for Noise Reduction
Shuyang Jiang (Fudan University), Yu Wang (Shanghai Jiao Tong University)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This study investigates the redundancy of LoRA parameters and proposes the PREFT framework and NORM method to eliminate noise and enhance PEFT effectiveness.
FIRING-Net: A filtered feature recycling network for speech enhancement
Xinmeng Xu (Wuhan University), Weiping Tu (Wuhan University)
RestorationConvolutional Neural NetworkTransformerAudio
🎯 What it does: A Filter-Recycle-Interguide (FRI) framework is proposed, which recycles speech information using filtered noise features and mutually guides it with retained target features, achieving finer separation of speech and noise.
First-Person Fairness in Chatbots
Tyna Eloundou (OpenAI), Adam Tauman Kalai (OpenAI)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: By conducting name counterfactual experiments on chatbot responses and using the Language Model Research Assistant (LMRA) for large-scale quantitative and qualitative bias assessment, a scalable first-person fairness evaluation method is proposed and implemented.
Fitting Networks with a Cancellation Trick
Jiashun Jin (Carnegie Mellon University), Jingming Wang (University of Virginia)
Score-based ModelGraph
🎯 What it does: This paper proposes a new network model called logit-DCBM and designs a recursive SCORE (R-SCORE) algorithm based on this model for community detection.
Flash Inference: Near Linear Time Inference for Long Convolution Sequence Models and Beyond
Costin-Andrei Oncescu (Harvard University), Sham M. Kakade (Harvard University)
OptimizationComputational EfficiencyConvolutional Neural NetworkSequential
🎯 What it does: A framework named Flash Inference is proposed, which can improve the inference time of Long Convolution Sequence Models (LCSM) and similar architectures from quadratic to nearly linear (O(L log²L)).
FlashMask: Efficient and Rich Mask Extension of FlashAttention
Guoxia Wang (Baidu Inc.), Haifeng Wang (Baidu Inc.)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes FLASHMASK, which extends FlashAttention by introducing a column-sparse mask representation that supports more complex masking patterns and achieves linear memory complexity.
FlashRNN: I/O-Aware Optimization of Traditional RNNs on modern hardware
Korbinian Pöppel (Johannes Kepler University), Sepp Hochreiter (Johannes Kepler University)
OptimizationComputational EfficiencyRecurrent Neural NetworkSequential
🎯 What it does: The FlashRNN library is proposed, which aims for efficient implementation of traditional RNNs (LSTM, GRU, sLSTM) on GPUs by integrating matrix multiplication and activation operations, supporting multi-head parallelism, providing implementations in both CUDA and Triton, and using the integer constraint solver ConstrINT for automatic tuning.
Flat Reward in Policy Parameter Space Implies Robust Reinforcement Learning
Hyun Kyu Lee (Ulsan National Institute of Science and Technology), Sung Whan Yoon (Ulsan National Institute of Science and Technology)
Reinforcement LearningSequential
🎯 What it does: This paper studies the relationship between the flatness of the reward function in reinforcement learning and robustness, proving that a flatter reward landscape can enhance robustness to action, transition probability, and reward perturbations, and validates this theory through algorithms such as SAM+PPO.
Flavors of Margin: Implicit Bias of Steepest Descent in Homogeneous Neural Networks
Nikolaos Tsilivis (New York University), Julia Kempe (New York University)
OptimizationTabular
🎯 What it does: This paper studies the implicit bias of steepest descent algorithms (including gradient descent, sign gradient descent, and coordinate descent) for arbitrary norms when aligning homogeneous deep neural networks. It proves that this bias will promote the increase of geometric boundaries in the later stages of training and converge to the (generalized) KKT points of the maximum boundary problem. It also provides a generalization stagnation measure based on Bregman divergence and experimentally verifies its connection to Adam.
FlexCAD: Unified and Versatile Controllable CAD Generation with Fine-tuned Large Language Models
Zhanwei Zhang (Zhejiang University), Jiang Bian (Microsoft Research)
GenerationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This paper presents FlexCAD, a unified controllable CAD generation model that utilizes LLM to achieve multi-level control of CAD generation (such as CAD, sketch-extrusion, sketch, face, loop, curve, etc.) through structured text representation and hierarchical-aware masking.
FlexPrefill: A Context-Aware Sparse Attention Mechanism for Efficient Long-Sequence Inference
Xunhao Lai (Peking University), Xun Zhou (ByteDance Inc)
Computational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: Introducing a mechanism called FlexPrefill that dynamically adjusts the sparse attention pattern during the pre-filling phase of long sequence reasoning;
FlickerFusion: Intra-trajectory Domain Generalizing Multi-agent Reinforcement Learning
Woosung Koh (Yonsei University), Se-Young Yun (KAIST AI)
Reinforcement LearningTabularBenchmark
🎯 What it does: In multi-agent reinforcement learning, the FLICKERFUSION method is proposed to address the zero-shot domain generalization problem caused by changes in the number of entities during inference.
FLIP: Flow-Centric Generative Planning as General-Purpose Manipulation World Model
Chongkai Gao (National University of Singapore), Lin Shao (National University of Singapore)
GenerationRobotic IntelligenceTransformerDiffusion modelFlow-based ModelWorld ModelImageVideoTextMultimodality
🎯 What it does: We propose FLIP, a generative planning framework based on image streams that can generate long-term operational plans from language descriptions and guide low-level control.
FLOPS: Forward Learning with OPtimal Sampling
Tao Ren (Peking University), Yijie Peng (Peking University)
OptimizationTransformerReinforcement LearningPrompt EngineeringImage
🎯 What it does: This study investigates the process of estimating gradients in forward learning solely through forward queries, proposing an optimal query allocator based on the likelihood ratio method, which adaptively allocates the number of queries for each sample according to the variance of the gradient estimate.
Flow Distillation Sampling: Regularizing 3D Gaussians with Pre-trained Matching Priors
Lin-Zhuo Chen (Nanjing University), Yao Yao (Nanjing University)
RestorationOptimizationKnowledge DistillationGaussian SplattingOptical FlowPoint Cloud
🎯 What it does: This paper proposes the Flow Distillation Sampling (FDS) technique, which utilizes the matching prior of a pre-trained optical flow model to supervise the geometric optimization of 3D Gaussian Splatting (3DGS/2DGS), thereby improving the quality of geometric reconstruction and new view rendering under sparse perspectives.
Flow matching achieves almost minimax optimal convergence
Kenji Fukumizu (Institute of Statistical Mathematics Preferred Networks), Masanori Koyama (Preferred Networks University of Tokyo)
GenerationData SynthesisOptimizationDiffusion modelFlow-based Model
🎯 What it does: This paper provides a rigorous analysis of the convergence properties of Flow Matching generative models, proving that they can achieve nearly optimal convergence rates in p-Wasserstein distance under large sample sizes.
Flow Matching with Gaussian Process Priors for Probabilistic Time Series Forecasting
Marcel Kollovieh (Technical University of Munich), Stephan Günnemann (Technical University of Munich)
Flow-based ModelTime Series
🎯 What it does: A TSFlow model based on conditional flow matching is proposed for time series forecasting, and a Gaussian process prior is utilized to improve the generation process.
Flow Matching with General Discrete Paths: A Kinetic-Optimal Perspective
Neta Shaul (Weizmann Institute of Science), Ricky T. Q. Chen (Meta FAIR)
Data SynthesisOptimizationFlow-based ModelImageText
🎯 What it does: A discrete flow matching framework based on continuous-time Markov chains is proposed, allowing the use of any discrete probability path (corruption process), and obtaining the corresponding velocity through dynamic energy minimization;
Flow-based Variational Mutual Information: Fast and Flexible Approximations
Caleb Dahlke (University of Michigan), Jason Pacheco (University of Arizona)
Flow-based Model
🎯 What it does: A variational mutual information estimation method based on regularized flows (JVF and NVF) is proposed, and its performance is evaluated across various tasks.
Flow: Modularized Agentic Workflow Automation
Boye Niu (Sydney AI Centre, University of Sydney), Tongliang Liu (Sydney AI Centre, University of Sydney)
Large Language ModelAgentic AITextMultimodality
🎯 What it does: This paper proposes a multi-agent framework called Flow, based on large language models, which can dynamically adjust workflows during task execution, achieving task parallelization and fault tolerance.
FlowDec: A flow-based full-band general audio codec with high perceptual quality
Simon Welker (Signal Processing University of Hamburg), YI-CHIAO WU
CompressionFlow-based ModelAudio
🎯 What it does: An end-to-end neural audio codec named FlowDec is proposed, utilizing Conditional Flow Matching for post-processing, supporting 48kHz full-band audio and achieving high perceptual quality at low bit rates of 4–7.5 kbit/s.
Fluid: Scaling Autoregressive Text-to-image Generative Models with Continuous Tokens
Lijie Fan (Google DeepMind), Yonglong Tian (MIT)
GenerationData SynthesisTransformerDiffusion modelImageText
🎯 What it does: The study explores the scalability of autoregressive models in text-to-image generation, systematically evaluating the impact of continuous/discrete tokens and random/grid generation orders. Based on this, a large-scale randomly ordered continuous token model called Fluid was constructed and trained, achieving new zero-shot FID and GenEval benchmark records.
Follow My Instruction and Spill the Beans: Scalable Data Extraction from Retrieval-Augmented Generation Systems
Zhenting Qi (Harvard University), Himabindu Lakkaraju (Harvard University)
RetrievalSafty and PrivacyAdversarial AttackTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: This paper studies the risks of extracting text from non-parametric repositories of Retrieval-Augmented Generation (RAG) systems through Prompt-Injected Data Extraction.
Following the Human Thread in Social Navigation
Luca Scofano (Sapienza University of Rome), Fabio Galasso (Sapienza University of Rome)
Robotic IntelligenceReinforcement LearningImage
🎯 What it does: This paper proposes the Social Dynamics Adaptation (SDA) model, which utilizes human trajectory information available during training to assist robots in social navigation, and infers social dynamics during deployment through state-action history.
For Better or For Worse? Learning Minimum Variance Features With Label Augmentation
Muthu Chidambaram (Duke University), Rong Ge (Duke University)
ClassificationOptimizationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: This study investigates the theoretical and experimental characteristics of two label augmentation methods: label smoothing and Mixup. It demonstrates that they encourage the model to learn low-variance features by reducing the variance of model outputs, and validates this phenomenon in image classification tasks.
ForecastBench: A Dynamic Benchmark of AI Forecasting Capabilities
Ezra Karger (Forecasting Research Institute Federal Reserve Bank of Chicago), Philip Tetlock
Large Language ModelPrompt EngineeringTextTime SeriesBenchmark
🎯 What it does: A dynamic and continuously updated forecasting benchmark, ForecastBench, has been proposed and implemented, along with the public release of a leaderboard and a forecasting question-answer dataset.
Forewarned is Forearmed: Harnessing LLMs for Data Synthesis via Failure-induced Exploration
Qintong Li (Hong Kong University), Lingpeng Kong (Hong Kong University)
Data SynthesisSafty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: A data synthesis framework named REVERSEGEN is proposed, which generates targeted training samples using the failure cases of the target model, thereby improving the model's performance on tasks related to safety, honesty, and mathematical reasoning.
Forget the Data and Fine-Tuning! Just Fold the Network to Compress
Dong Wang (Graz University of Technology), Olga Saukh (ETH Zurich)
CompressionConvolutional Neural NetworkImage
🎯 What it does: Without using training data and fine-tuning, a model folding technique is proposed to compress the model by clustering and merging similar channels within the same network.
Forgetting Transformer: Softmax Attention with a Forget Gate
Zhixuan Lin (Mila and Universite de Montreal), Aaron Courville (Mila and Universite de Montreal)
TransformerLarge Language ModelText
🎯 What it does: This paper proposes the Forgetting Transformer (FoX), which incorporates a learnable forget gate into the softmax attention of the Transformer, allowing the attention weights to decay over time and adapt based on the data.
Forking Paths in Neural Text Generation
Eric J Bigelow, Tomer Ullman (Harvard University)
GenerationTransformerLarge Language ModelText
🎯 What it does: This paper proposes the 'Forking Tokens Hypothesis' and develops the Forking Paths Analysis framework, which reveals key fork points in the LLM generation path and their impact on the final answer by gradually resampling during the generation process and statistically analyzing the subsequent text of each token.
FormalAlign: Automated Alignment Evaluation for Autoformalization
Jianqiao Lu (University of Hong Kong), Zhijiang Guo (Hong Kong University of Science and Technology)
TransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
🎯 What it does: This paper proposes the FORMALALIGN framework, which achieves automated evaluation of the semantic alignment of mathematical proofs from natural language to formal language, significantly reducing the workload of manual verification.
Formation of Representations in Neural Networks
Liu Ziyin (Massachusetts Institute of Technology), Tomaso A Poggio
Representation LearningConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper proposes the Canonical Representation Hypothesis (CRH) and the Polynomial Alignment Hypothesis (PAH), theoretically explaining the alignment relationships among representations, gradients, and weights in neural networks, and experimentally validating the universality of these alignments across various networks.
Forte : Finding Outliers with Representation Typicality Estimation
Debargha Ganguly (Case Western Reserve University), Vipin Chaudhary (Google Research)
Anomaly DetectionTransformerContrastive LearningImage
🎯 What it does: A self-supervised representation-based unsupervised anomaly detection framework called Forte is proposed, which utilizes four point-level statistics (precision, recall, density, coverage) and non-parametric density estimators to detect OOD and generated images.
FOSP: Fine-tuning Offline Safe Policy through World Models
Chenyang Cao (Tsinghua University), Xueqian Wang
Safty and PrivacyRobotic IntelligenceReinforcement LearningWorld ModelImage
🎯 What it does: A framework for offline safe policy online fine-tuning based on world models (FOSP) is proposed, achieving safe offline to online transfer for visual robotic tasks.
Foundation Models Secretly Understand Neural Network Weights: Enhancing Hypernetwork Architectures with Foundation Models
Jeffrey Gu (Stanford University), Serena Yeung-Levy (Stanford University)
OptimizationRepresentation LearningTransformerPrompt EngineeringPoint CloudAudio
🎯 What it does: This study investigates how to integrate large pre-trained foundational models (such as ViT, CLIP, DINO) into hypernetwork architectures to enhance the performance of generalizable implicit neural representations (INR).
Fourier Head: Helping Large Language Models Learn Complex Probability Distributions
Nate Gillman (Brown University), Chen Sun (Google DeepMind)
Data SynthesisTransformerLarge Language ModelReinforcement LearningContrastive LearningTime SeriesSequential
🎯 What it does: A Fourier Head based on Fourier series is proposed and implemented to allow large language models to retain continuity priors after discretization, improving the modeling and sampling quality of non-linguistic continuous values.
Fourier Sliced-Wasserstein Embedding for Multisets and Measures
Tal Amir (Technion Israel Institute of Technology), Nadav Dym (Technion Israel Institute of Technology)
OptimizationPoint CloudBiomedical Data
🎯 What it does: A Fourier Sliced-Wasserstein (FSW) embedding method is proposed, which embeds finite-support multisets and measures into Euclidean space, approximately preserving the sliced Wasserstein distance.
Fragment and Geometry Aware Tokenization of Molecules for Structure-Based Drug Design Using Language Models
Cong Fu (Texas A&M University), Shuiwang Ji (University of Illinois)
Drug DiscoveryTransformerLarge Language ModelGraph
🎯 What it does: This paper proposes a language model-based molecular fragment and geometric-aware serialization method called Frag2Seq for structure-based drug design.
Frame-Voyager: Learning to Query Frames for Video Large Language Models
Sicheng Yu (ByteDance), Qianru Sun (Singapore Management University)
TransformerLarge Language ModelReinforcement LearningVision Language ModelVideoText
🎯 What it does: A plugin method called FRAME-VOYAGER is proposed to learn the most informative frame combinations from videos based on given text queries, thereby enhancing the video understanding capabilities of Video-LLM.
Framer: Interactive Frame Interpolation
Wen Wang (Zhejiang University), Chunhua Shen (Zhejiang University of Technology)
GenerationData SynthesisAutonomous DrivingSupervised Fine-TuningDiffusion modelVideo
🎯 What it does: An interactive frame interpolation framework called Framer has been developed, allowing users to control smooth transitions between two frames by specifying keypoint trajectories.
FreCaS: Efficient Higher-Resolution Image Generation via Frequency-aware Cascaded Sampling
Zhengqiang ZHANG, Lei Zhang (Hong Kong Polytechnic University)
GenerationData SynthesisComputational EfficiencyDiffusion modelImage
🎯 What it does: This paper proposes a frequency-aware cascading sampling framework named FreCaS, which efficiently generates high-resolution images (e.g., 4096×4096) using existing diffusion models without the need for retraining.
Fréchet Wavelet Distance: A Domain-Agnostic Metric for Image Generation
Lokesh Veeramacheneni (University of Bonn), Juergen Gall (University of Bonn)
GenerationData SynthesisComputational EfficiencyImageAgriculture Related
🎯 What it does: Proposed the Fréchet Wavelet Distance (FWD), a domain-free, untrained network evaluation metric for generative models based on Wavelet Packet Transform.
FreDF: Learning to Forecast in the Frequency Domain
Hao Wang (Zhejiang University), Dacheng Tao (Nanyang Technological University)
TransformerTime Series
🎯 What it does: This study investigates the learning target bias caused by label autocorrelation in direct forecasting (DF) methods and proposes the FreDF framework for frequency domain learning prediction, compatible with any forecasting model.
Free Hunch: Denoiser Covariance Estimation for Diffusion Models Without Extra Costs
Severi Rissanen (Aalto University), Arno Solin (Aalto University)
RestorationSuper ResolutionDiffusion modelImageOrdinary Differential Equation
🎯 What it does: Proposes the Free Hunch framework, achieving covariance estimation for diffusion model denoisers with no additional training and low computational cost;
FreeCG: Free the Design Space of Clebsch-Gordan Transform for Machine Learning Force Fields
Shihao Shao (Peking University), Qinghua Cui (Peking University)
Graph Neural NetworkTransformerGraphPhysics Related
🎯 What it does: A geometric neural network named FreeCG is proposed, which constructs abstract edges that can be freely designed using variable Clebsch–Gordan (CG) transformations, significantly enhancing the expressive power of machine learning force fields (MLFF).
FreeVS: Generative View Synthesis on Free Driving Trajectory
Qitai Wang (University of Chinese Academy of Sciences), Zhaoxiang Zhang (Chinese Academy of Sciences)
GenerationData SynthesisAutonomous DrivingDiffusion modelAuto EncoderVideoPoint CloudBenchmark
🎯 What it does: A fully generative view synthesis method named FreeVS is proposed, which can synthesize new camera views along free driving trajectories.
FreqPrior: Improving Video Diffusion Models with Frequency Filtering Gaussian Noise
Yunlong Yuan (Fudan University), Li Zhang (Fudan University)
GenerationData SynthesisComputational EfficiencyDiffusion modelVideoBenchmark
🎯 What it does: Designed and implemented FreqPrior, a technique that improves the noise prior of video diffusion models through frequency domain noise filtering.
Frequency-Guided Masking for Enhanced Vision Self-Supervised Learning
Amin Karimi Monsefi (Ohio State University), Rajiv Ramnath (Ohio State University)
SegmentationKnowledge DistillationRepresentation LearningConvolutional Neural NetworkTransformerImage
🎯 What it does: A frequency-domain based self-supervised learning framework called FOLK is proposed, utilizing adaptive frequency masking and teacher-student knowledge distillation for pre-training.
FreSh: Frequency Shifting for Accelerated Neural Representation Learning
Adam Kania (Jagiellonian University), Przemysław Spurek (IDEAS NCBR)
Representation LearningHyperparameter SearchNeural Radiance FieldImageVideo
🎯 What it does: The FreSh method is proposed, which automatically selects the best embedding hyperparameters by matching the frequency spectrum of the model's initial output with the target signal, avoiding the heavy grid search.
From an LLM Swarm to a PDDL-empowered Hive: Planning Self-executed Instructions in a Multi-modal Jungle
Kaustubh Vyas (Huawei Technologies), Jeff Z. Pan (University of Edinburgh)
Recommendation SystemOptimizationExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextMultimodalityBenchmarkAudio
🎯 What it does: The HIVE system is proposed, which constructs a model capability knowledge graph (C-KG). It combines LLM and PDDL planning to automatically parse multimodal user instructions, select appropriate model sets, and execute them according to an interpretable action plan, supporting user-defined constraints.
From Artificial Needles to Real Haystacks: Improving Retrieval Capabilities in LLMs by Finetuning on Synthetic Data
Zheyang Xiong (University of Wisconsin Madison), Dimitris Papailiopoulos (Microsoft Research)
RetrievalTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextChain-of-Thought
🎯 What it does: A method for fine-tuning large language models (LLMs) on fully synthetic retrieval tasks is proposed, significantly improving their performance in long-context retrieval and reasoning tasks.
From Attention to Activation: Unraveling the Enigmas of Large Language Models
Prannay Kaul (Huawei), Jiankang Deng (Huawei)
OptimizationTransformerLarge Language ModelText
🎯 What it does: This study investigates two phenomena in autoregressive Transformers: the dominance of the first token's attention and the abnormal activation of hidden states, and proposes corresponding solutions.
From Commands to Prompts: LLM-based Semantic File System for AIOS
Zeru Shi (Rutgers University), Yongfeng Zhang (Rutgers University)
RetrievalOptimizationTransformerLarge Language ModelPrompt EngineeringTextMultimodality
🎯 What it does: Designed and implemented a Large-scale Language Model-based Semantic File System (LSFS), achieving a closed loop from natural language prompts to file retrieval, updates, rollbacks, and other operations through the construction of a semantic index, vector database, and scalable syscall and API layers.
From Decoupling to Adaptive Transformation: a Wider Optimization Space for PTQ
Zhaojing Wen (Hikvision Research Institute), Jiang Zhu (Hikvision Research Institute)
Super ResolutionOptimizationTransformerSupervised Fine-TuningImage
🎯 What it does: This paper proposes AdaQTransform, a post-training quantization (PTQ) method that decouples the weight quantization step size and applies adaptive linear transformations to the quantized output, making the quantized model closer to the FP32 output.
From Exploration to Mastery: Enabling LLMs to Master Tools via Self-Driven Interactions
Changle Qu (Renmin University of China), Ji-Rong Wen (Renmin University of China)
TransformerLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: A self-driven framework named DRAFT is proposed, which interacts with external tools through LLM to generate feedback, iteratively rewriting tool documentation to enhance LLM's understanding and usage capabilities of the tools.
From Few to Many: Self-Improving Many-Shot Reasoners Through Iterative Optimization and Generation
Xingchen Wan (Google Cloud AI Research), Sercan O Arik
OptimizationTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: This paper studies the effects of multi-sample context learning in long-context large language models and proposes a self-improving iterative optimization and generation algorithm called BRIDGE.
From GNNs to Trees: Multi-Granular Interpretability for Graph Neural Networks
Jie Yang (Zhejiang University), Shunyu Liu (Nanyang Technological University)
Explainability and InterpretabilityGraph Neural NetworkGraph
🎯 What it does: Transforming ordinary GNNs into a hierarchical tree structure to achieve interpretable graph classification at multiple levels (multi-granularity).
From Isolated Conversations to Hierarchical Schemas: Dynamic Tree Memory Representation for LLMs
Alireza Rezazadeh (Center for Advanced AI), Yujia Bao (Center for Advanced AI)
Large Language ModelTextRetrieval-Augmented Generation
🎯 What it does: This paper presents MemTree, a long-term memory representation method based on a dynamic tree structure that can update, aggregate, and retrieve information in real-time during large language model dialogues or document question-answering.
From Layers to States: A State Space Model Perspective to Deep Neural Network Layer Dynamics
Qinshuo Liu (University of Hong Kong), Guodong Li (University of Hong Kong)
ClassificationObject DetectionSegmentationConvolutional Neural NetworkTransformerImage
🎯 What it does: The outputs of each layer in the deep network are treated as a continuous state sequence, and a lightweight S6LA module is proposed to achieve cross-layer information aggregation through a selective state space model (S6).
From Lazy to Rich: Exact Learning Dynamics in Deep Linear Networks
Clémentine Carla Juliette Dominé (University College London), Andrew M Saxe
🎯 What it does: This study investigates the learning dynamics of deep linear networks under different λ-balanced initializations, providing an analytical solution for the gradient flow and revealing a continuous transition from the lazy (kernel regression) to the rich (feature learning) paradigm through this analytical solution.
From Models to Microtheories: Distilling a Model's Topical Knowledge for Grounded Question-Answering
Nathaniel Weir (Johns Hopkins University), Peter Clark (Allen Institute for AI)
OptimizationKnowledge DistillationTransformerLarge Language ModelText
🎯 What it does: Through a problem-driven process, the internal knowledge of the language model is distilled into interpretable 'micro-theories' (a set of core knowledge sentences), and their ability to support answers to topics is tested using a text entailment mechanism.
From Pixels to Tokens: Byte-Pair Encoding on Quantized Visual Modalities
Wanpeng Zhang (Peking University), Zongqing Lu (Peking University)
GenerationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageMultimodality
🎯 What it does: A byte pair encoding (BPE)-based image tokenizer is proposed, and a two-stage multimodal large language model (MLLM) training framework is designed around it.
From Probability to Counterfactuals: the Increasing Complexity of Satisfiability in Pearl's Causal Hierarchy
Julian Dörfler (Saarland University), Maciej Liskiewicz (University of Lübeck)
🎯 What it does: This paper studies the computational complexity of the satisfiability problem under three types of reasoning (probabilistic, intervention, counterfactual) in Pearl's Causal Hierarchy (PCH), focusing on basic and linear languages that allow the use of summation (Σ) operations.
From Promise to Practice: Realizing High-performance Decentralized Training
Zesen Wang (KTH Royal Institute of Technology), Mikael Johansson (KTH Royal Institute of Technology)
OptimizationTransformerImageText
🎯 What it does: Implement and evaluate decentralized training on multi-node GPU clusters, propose a decentralized Adam variant that covers communication and computation along with an accumulation mechanism, and construct a runtime model to guide topology and parameter configuration.
From Risk to Uncertainty: Generating Predictive Uncertainty Measures via Bayesian Estimation
Nikita Kotelevskii (Mohammed Bin Zayed University of Artificial Intelligence), Maxim Panov (Mohammed Bin Zayed University of Artificial Intelligence)
ClassificationAnomaly DetectionImage
🎯 What it does: A prediction uncertainty framework based on point risk decomposition and strict proper scoring rules is proposed, unifying various uncertainty measures that can be obtained through Bayesian estimation.
From Search to Sampling: Generative Models for Robust Algorithmic Recourse
Prateek Garg (Indian Institute of Technology Bombay), Sunita Sarawagi (Indian Institute of Technology Bombay)
Recommendation SystemOptimizationTransformerGenerative Adversarial NetworkTabular
🎯 What it does: A generative feedback model called GenRe is proposed, aimed at jointly optimizing effective algorithmic feedback methods to help individuals adversely affected by automated model decisions improve their features, thereby achieving more favorable outcomes.
From Sparse Dependence to Sparse Attention: Unveiling How Chain-of-Thought Enhances Transformer Sample Efficiency
Kaiyue Wen (Stanford University), Jingzhao Zhang (Tsinghua University)
Computational EfficiencyTransformerTabularChain-of-Thought
🎯 What it does: Study the impact of Chain of Thought (CoT) on the sample efficiency and attention sparsity of Transformers in learning parity functions.
From Tokens to Lattices: Emergent Lattice Structures in Language Models
Bo Xiong (Stanford University), Steffen Staab (University of Stuttgart)
TransformerLarge Language ModelTextBiomedical Data
🎯 What it does: Through the framework of Formal Concept Analysis (FCA), this study investigates and constructs the concept lattice of implicit learning in pre-trained Masked Language Models (MLM), revealing how it recovers object-attribute relationships and hierarchical structures from token-level probability distributions, and validates this method on three types of domain data.
From Tokens to Words: On the Inner Lexicon of LLMs
Guy Kaplan (Hebrew University of Jerusalem), Roy Schwartz (Hebrew University of Jerusalem)
TransformerLarge Language ModelTextBiomedical Data
🎯 What it does: This study investigates how large language models internally reorganize subword sequences into complete words and explores the hierarchical mechanisms and internal vocabulary involved in this process.
Fugatto 1: Foundational Generative Audio Transformer Opus 1
Rafael Valle (NVIDIA), Bryan Catanzaro (NVIDIA)
GenerationData SynthesisTransformerLarge Language ModelAudio
🎯 What it does: A general audio generation and transformation model called Fugatto is proposed, which implements the combination and interpolation of instructions, tasks, and models during inference through ComposableART.
Fully-inductive Node Classification on Arbitrary Graphs
Jianan Zhao (Mila - Québec AI Institute), Jian Tang (Mila - Québec AI Institute)
ClassificationGraph Neural NetworkGraph
🎯 What it does: The GraphAny model is proposed, achieving fully transferable node classification and enabling inference on any graph.
Functional Homotopy: Smoothing Discrete Optimization via Continuous Parameters for LLM Jailbreak Attacks
Zi Wang (University of Wisconsin Madison), Somesh Jha (University of Wisconsin Madison)
OptimizationAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A functional homotopy method is proposed, coupling the continuous space of model parameters with the discrete input space. It constructs an optimization sequence from easy to difficult in the parameter space using gradients, and then searches in the discrete space through a temperature-increasing approach, smoothing the input generation problem of LLMs.
Fundamental Limitations on Subquadratic Alternatives to Transformers
Josh Alman (Columbia University), Hantao Yu (Columbia University)
Transformer
🎯 What it does: Proves that any sub-quadratic time attention alternative cannot complete the document similarity task, and demonstrates that a single-head Transformer can accomplish these tasks.
Fundamental Limits of Prompt Tuning Transformers: Universality, Capacity and Efficiency
Jerry Yao-Chieh Hu (Northwestern University), Han Liu (Northwestern University)
TransformerPrompt Engineering
🎯 What it does: The research highlights the statistical and computational limits of Prompt Tuning in Transformer models and demonstrates that a single-head single-layer Transformer can approximate any Lipschitz sequence-to-sequence function.
G-LLaVA: Solving Geometric Problem with Multi-Modal Large Language Model
Jiahui Gao (University of Hong Kong), Lingpeng Kong (University of Hong Kong)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: This paper proposes G-LLAVA, a multimodal large language model for geometric problems, which utilizes a text LLM to automatically generate a high-quality geometric dataset Geo170K, thereby enhancing the model's understanding and reasoning capabilities regarding geometric shapes.
GALA: Geometry-Aware Local Adaptive Grids for Detailed 3D Generation
Dingdong Yang (Simon Fraser University), Hao Zhang (Simon Fraser University)
GenerationData SynthesisTransformerDiffusion modelPoint CloudMesh
🎯 What it does: We propose GALA—a geometry-aware local adaptive mesh that utilizes a sparse octree to cover surfaces and constructs locally aligned and anisotropically scaled meshes at each leaf node for high-quality 3D shape reconstruction and generation.
GameArena: Evaluating LLM Reasoning through Live Computer Games
Lanxiang Hu (University of California), Hao Zhang (University of California)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: By conducting three interactive games (Akinator, Taboo, Bluffing) with human players, we collected and analyzed step-by-step reasoning data from LLMs to construct a dynamic benchmark called GameArena.
GameGen-X: Interactive Open-world Game Video Generation
Haoxuan Che (Hong Kong University of Science and Technology), Hao Chen (Hong Kong University of Science and Technology)
GenerationData SynthesisTransformerMixture of ExpertsDiffusion modelVideoTextMultimodality
🎯 What it does: GameGenX is proposed, a diffusion transformer model capable of generating and interactively controlling open-world game videos.
GANDALF: Generative AttentioN based Data Augmentation and predictive modeLing Framework for personalized cancer treatment
Aishwarya Jayagopal (National University of Singapore), Vaibhav Rajan (National University of Singapore)
GenerationDrug DiscoveryTransformerDiffusion modelAuto EncoderBiomedical Data
🎯 What it does: This paper proposes the GANDALF framework, which generates patient-like gene mutation samples on cell line data through a generative attention mechanism and assigns pseudo-labels for drug responses, thereby alleviating the issue of scarce patient annotation data and further training drug response prediction models.
Gap Preserving Distillation by Building Bidirectional Mappings with A Dynamic Teacher
Yong Guo (Max Planck Institute for Informatics), Jian Chen (South China University of Technology)
ClassificationKnowledge DistillationConvolutional Neural NetworkReinforcement LearningImage
🎯 What it does: This paper studies a Gap Preserving Distillation (GPD) method, which utilizes a dynamic teacher constructed from the student model and enforces parameter sharing to maintain an appropriate performance gap between the student and teacher, thereby significantly enhancing the effectiveness of knowledge distillation.
Gap-Dependent Bounds for Q-Learning using Reference-Advantage Decomposition
Zhong Zheng (Pennsylvania State University), Lingzhou Xue (Pennsylvania State University)
Reinforcement Learning
🎯 What it does: This paper studies the gap-dependent bounds of two important online Q-learning algorithms (UCB-Advantage and Q-EarlySettled-Advantage) based on reference advantage decomposition in finite-time tabular Markov Decision Processes (MDPs).
Gated Delta Networks: Improving Mamba2 with Delta Rule
Songlin Yang (Massachusetts Institute of Technology), Ali Hatamizadeh (NVIDIA)
TransformerText
🎯 What it does: This paper proposes a new linear Transformer—Gated DeltaNet, which constructs an efficient memory management mechanism by integrating gated decay with the Delta update rule, and provides a hardware-friendly chunked parallel training algorithm.