ICML 2025 Papers — Page 12
International Conference on Machine Learning · 3257 papers
Flexibility-conditioned protein structure design with flow matching
Vsevolod Viliuga (Heidelberg Institute for Theoretical Studies), Frauke Gräter (Max Planck Institute for Polymer Research)
Protein Structure PredictionFlow-based ModelBiomedical Data
🎯 What it does: This paper proposes a scalable protein structure generation framework based on flow matching (FliPS), which can synthesize protein backbones under a given flexible distribution for each residue and filter them using BackFlip;
Flexible and Efficient Grammar-Constrained Decoding
Kanghee Park (University of California San Diego), Loris D'Antoni (University of California San Diego)
OptimizationComputational EfficiencyAI Code AssistantText
🎯 What it does: A new grammar-constrained decoding (GCD) algorithm called GREATGRAMMA is proposed and implemented, which reduces offline preprocessing time by constructing a token spanner table while maintaining online masking efficiency.
Flexible Tails for Normalizing Flows
Tennessee Hickling (University of Bristol), Dennis Prangle (University of Bristol)
Flow-based ModelTabularTime SeriesFinance Related
🎯 What it does: A Tail Transform Flow (TTF) is proposed, which adds a non-Lipschitz transformation layer based on erfc to the traditional normalizing flow, transforming the Gaussian base into an adjustable heavy-tailed distribution, and applying this method to density estimation and variational inference.
Flexible, Efficient, and Stable Adversarial Attacks on Machine Unlearning
Zihan Zhou (Auburn University), Dejing Dou (Fudan University)
Adversarial AttackConvolutional Neural NetworkTransformerImageText
🎯 What it does: A dynamic delay poisoning attack (DDPA) targeting machine learning models is proposed, which specifies the attack target after model training and uses poisoned data to trigger malicious adjustments to model parameters when the model loses specific samples.
FlexiClip: Locality-Preserving Free-Form Character Animation
Anant Khandelwal (Microsoft)
GenerationData SynthesisDiffusion modelImageVideoOrdinary Differential Equation
🎯 What it does: We propose FlexiClip, a text-prompted free-form clip art animation framework that achieves smooth and coherent animations while maintaining visual identity and geometric consistency.
FlexiReID: Adaptive Mixture of Expert for Multi-Modal Person Re-Identification
Zhen Sun (Xiamen University), Rongrong Ji (Xiamen University)
RecognitionRetrievalTransformerMixture of ExpertsContrastive LearningImageTextMultimodality
🎯 What it does: The FlexiReID framework is proposed, supporting flexible pedestrian re-identification across seven retrieval modes in four modalities (RGB, infrared, sketch, text);
FlexTok: Resampling Images into 1D Token Sequences of Flexible Length
Roman Bachmann (Swiss Federal Institute of Technology Lausanne), Afshin Dehghan
GenerationCompressionTransformerRectified FlowImage
🎯 What it does: This paper studies the variable-length 1D image tokenizer FlexTok, which can compress images into 1-256 discrete tokens and supports autoregressive generation.
FlipAttack: Jailbreak LLMs via Flipping
Yue Liu (National University of Singapore), Bryan Hooi (National University of Singapore)
Adversarial AttackTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: A black-box LLM jailbreak method called FlipAttack is proposed, which utilizes left-side perturbations and flipping techniques to bypass security protections.
Floating-Point Neural Networks Can Represent Almost All Floating-Point Functions
Geonho Hwang (Gwangju Institute of Science and Technology), Sejun Park (Korea University)
🎯 What it does: The study investigates whether feedforward neural networks can approximate all floating-point functions under actual floating-point operators and parameters, and provides necessary and sufficient conditions.
FloE: On-the-Fly MoE Inference on Memory-constrained GPU
Yuxin Zhou (Zhejiang University), Lidan Shou (Zhejiang University)
OptimizationComputational EfficiencyMixture of ExpertsText
🎯 What it does: We designed and implemented FloE, an on-demand Mixture-of-Experts inference system for memory-constrained GPUs, capable of loading expert parameters into memory without significantly compromising generation quality.
Flopping for FLOPs: Leveraging Equivariance for Computational Efficiency
Georg Bökman, Fredrik Kahl (Chalmers University of Technology)
Computational EfficiencyConvolutional Neural NetworkTransformerImage
🎯 What it does: This study investigates a method to achieve equivariant networks using horizontal mirror (flopping) symmetry to reduce FLOPs while keeping the number of parameters unchanged.
Flow Matching for Denoised Social Recommendation
Yinxuan Huang (National University of Defense Technology), Ye Wang (National University of Defense Technology)
Recommendation SystemGraph Neural NetworkFlow-based ModelGraphOrdinary Differential Equation
🎯 What it does: This paper proposes a social recommendation framework called RecFlow based on flow matching, which uses flow matching to denoise the noise in social graphs.
Flow Matching for Few-Trial Neural Adaptation with Stable Latent Dynamics
Puli Wang (Zhejiang University), Gang Pan (Zhejiang University)
Domain AdaptationTransformerFlow-based ModelBiomedical Data
🎯 What it does: A few-shot trial neural adaptation framework based on flow matching (FDA) is proposed, which achieves source-independent domain alignment by learning flexible neural representations with stable latent dynamics, thereby enhancing the long-term decoding performance of brain-machine interfaces.
Flow of Reasoning: Training LLMs for Divergent Reasoning with Minimal Examples
Fangxu Yu (University of California San Diego), Lianhui Qin (University of California San Diego)
TransformerLarge Language ModelReinforcement LearningTextChain-of-Thought
🎯 What it does: Proposes Flow of Reasoning (FOR), a method for fine-tuning large language models using the principles of GFlowNet to generate diverse reasoning paths.
Flow Q-Learning
Seohong Park (University of California), Sergey Levine (University of California)
Reinforcement LearningFlow-based ModelTabularBenchmarkOrdinary Differential Equation
🎯 What it does: Proposes Flow Q-Learning, an offline reinforcement learning method that utilizes flow matching models to train complex behavior distributions and achieves value maximization through first-order policy distillation.
Flow-based Domain Randomization for Learning and Sequencing Robotic Skills
Aidan Curtis (Massachusetts Institute of Technology), Nicole E Carey
Robotic IntelligenceReinforcement LearningFlow-based ModelMultimodality
🎯 What it does: This paper proposes the GoFlow method, which dynamically learns the sampling distribution of environmental parameters using regularized entropy and self-paced KL regularization to achieve more robust sim-to-real transfer.
Flow-field inference from neural data using deep recurrent networks
Timothy Doyeon Kim (Princeton University), Carlos D Brody
Recurrent Neural NetworkAuto EncoderTime SeriesSequentialStochastic Differential Equation
🎯 What it does: The FINDR method is proposed, which infers low-dimensional nonlinear stochastic dynamics in neural network data using deep recursive variational autoencoders.
Flow-of-Options: Diversified and Improved LLM Reasoning by Thinking Through Options
Lakshmi Nair (Flagship Pioneering), J. Mark Kim (Flagship Pioneering)
GenerationDrug DiscoveryTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringImageTabular
🎯 What it does: This paper proposes the Flow-of-Options (FoO) reasoning framework and combines it with case-based reasoning to construct an agent system capable of automating various tasks such as machine learning, reinforcement learning, and image generation.
FlowAR: Scale-wise Autoregressive Image Generation Meets Flow Matching
Sucheng Ren (Johns Hopkins University), Liang-Chieh Chen (ByteDance)
GenerationData SynthesisTransformerFlow-based ModelAuto EncoderImage
🎯 What it does: A framework for image generation called FlowAR is proposed, which combines scale autoregression with flow matching, capable of generating high-quality images at each scale while maintaining a two-dimensional structure.
FlowDrag: 3D-aware Drag-based Image Editing with Mesh-guided Deformation Vector Flow Fields
Gwanhyeong Koo (Korea Advanced Institute of Science and Technology), Chang D. Yoo (Korea Advanced Institute of Science and Technology)
Image TranslationRestorationDiffusion modelImageVideoMeshBenchmark
🎯 What it does: By controlling the image through drag points for fine editing, the issue of geometric inconsistency in traditional drag editing is resolved;
Flowing Datasets with Wasserstein over Wasserstein Gradient Flows
Clément Bonet (ENSAE), Anna Korba
Domain AdaptationOptimizationFlow-based ModelImage
🎯 What it does: This paper proposes modeling gradient flows on random measure spaces using the Wasserstein over Wasserstein (WoW) distance, thereby achieving the flow and optimization of labeled datasets.
Fluctuations of the largest eigenvalues of transformed spiked Wigner matrices
Aro Lee (Korea Advanced Institute of Science and Technology), Ji Oon Lee (Korea Advanced Institute of Science and Technology)
Data SynthesisAnomaly DetectionOptimizationComputational EfficiencyTabularTime SeriesPhysics Related
🎯 What it does: The study investigates the limiting distribution of the maximum eigenvalue after element-wise transformations of the pulse Wigner matrix and proves that it satisfies the BBP phase transition; it converges to a Gaussian distribution in the supercritical region and to the GOE Tracy-Widom distribution in the subcritical region.
Focal-SAM: Focal Sharpness-Aware Minimization for Long-Tailed Classification
Sicong Li (Institute of Information Engineering), Qingming Huang (University of Chinese Academy of Sciences)
ClassificationOptimizationComputational EfficiencySupervised Fine-TuningImage
🎯 What it does: This paper proposes a long-tail classification method named Focal-SAM, which achieves fine control over the loss landscape of each category by introducing a focal-weighted sharpness penalty within the Sharpness-Aware Minimization (SAM) framework, while maintaining computational efficiency.
FOCoOp: Enhancing Out-of-Distribution Robustness in Federated Prompt Learning for Vision-Language Models
Xinting Liao (Zhejiang University), Tat-Seng Chua (National University of Singapore)
Domain AdaptationOptimizationFederated LearningTransformerPrompt EngineeringVision Language ModelImageMultimodality
🎯 What it does: This paper proposes the FOCoOp framework, which combines global, local, and OOD prompts to achieve prompt learning for federated vision-language models, and enhances OOD robustness and classification performance through bi-level distributionally robust optimization and semi-balanced optimal transport.
Focus On This, Not That! Steering LLMs with Adaptive Feature Specification
Tom A. Lamb (University of Oxford), Francesco Pinto (University of Chicago)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Proposes Focus Instruction Tuning (FIT), allowing LLM to focus on or ignore specified features based on natural language instructions during inference, dynamically controlling the output;
Forest-of-Thought: Scaling Test-Time Compute for Enhancing LLM Reasoning
Zhenni Bi (Huawei Noah's Ark Lab), Yunhe Wang (Huawei Noah's Ark Lab)
TransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: The Forest-of-Thought (FoT) framework is proposed, which integrates multiple reasoning trees and enhances the complex reasoning ability of LLMs through three mechanisms: sparse activation, dynamic self-correction, and consensus decision-making.
Foundation Model Insights and a Multi-Model Approach for Superior Fine-Grained One-shot Subset Selection
Zhijing Wan (Wuhan University), Shin'ichi Satoh (National Institute of Informatics)
ClassificationContrastive LearningImage
🎯 What it does: This paper studies the use of Foundation Models (FMs) as Information Extractors (IE) for one-time subset selection and proposes a multi-model fusion method called RAM-APL to enhance the subset selection performance for fine-grained image classification.
Foundation Molecular Grammar: Multi-Modal Foundation Models Induce Interpretable Molecular Graph Languages
Michael Sun (Massachusetts Institute of Technology), Jie Chen (Massachusetts Institute of Technology)
GenerationData SynthesisExplainability and InterpretabilityDrug DiscoveryGraph Neural NetworkLarge Language ModelPrompt EngineeringMultimodalityGraphChain-of-Thought
🎯 What it does: A framework is proposed that utilizes a multimodal foundation model (MMFM) to automatically learn an interpretable molecular graph grammar (FMG), and uses this grammar for molecular generation and property prediction.
FOUNDER: Grounding Foundation Models in World Models for Open-Ended Embodied Decision Making
Yucen Wang (Nanjing University), De-Chuan Zhan (Nanjing University)
Robotic IntelligenceReinforcement LearningAuto EncoderWorld ModelVideoText
🎯 What it does: Proposes the FOUNDER framework, embedding the fund model into the world model to achieve open-ended task learning based on text or video without rewards.
Fourier Position Embedding: Enhancing Attention’s Periodic Extension for Length Generalization
Ermo Hua (Tsinghua University), Bowen Zhou (Tsinghua University)
RetrievalOptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper analyzes the frequency domain properties of RoPE from the perspective of discrete signal processing, pointing out that spectral damage caused by linear layers, activation functions, and undertrained low-frequency components leads to this issue. It then proposes Fourier Position Embedding (FoPE) — using Fourier series in each dimension and setting low-frequency components to zero to enhance the periodic extension and length generalization of attention, and conducts pre-training, continued pre-training, and fine-tuning experiments on multi-scale models.
FourierMamba: Fourier Learning Integration with State Space Models for Image Deraining
Dong Li (University of Science and Technology of China), Zheng-Jun Zha (University of Science and Technology of China)
RestorationImage
🎯 What it does: Proposes the FourierMamba framework, which uses Mamba to correlate different frequencies in the Fourier space to achieve raindrop removal.
Fragments to Facts: Partial-Information Fragment Inference from LLMs
Lucas Rosenblatt (New York University), Bill Howe (University of Washington)
Safty and PrivacyAdversarial AttackLarge Language ModelTextBiomedical Data
🎯 What it does: This paper proposes the PIFI threat model and demonstrates a privacy leakage attack that infers sensitive segments based solely on a small amount of unordered fragments on LLM.
FrameBridge: Improving Image-to-Video Generation with Bridge Models
Yuji Wang (Tsinghua University), Jianfei Chen (Tsinghua University)
GenerationData SynthesisDiffusion modelAuto EncoderImageVideoStochastic Differential Equation
🎯 What it does: This paper proposes the FrameBridge framework, modeling frame-to-frame generation from images to videos as a data-to-data bridge process, replacing the traditional noise-to-data diffusion generation.
Fraud-Proof Revenue Division on Subscription Platforms
Abheek Ghosh (University of Oxford), Giannis Tyrovolas (University of Oxford)
Recommendation SystemOptimizationTabularAudio
🎯 What it does: Design a revenue distribution mechanism to combat fraud on subscription platforms, propose three types of axioms to resist manipulation, and evaluate the satisfaction of existing rules and new rules with these axioms.
Free Process Rewards without Process Labels
Lifan Yuan (University of Illinois Urbana Champaign), Hao Peng
Reinforcement Learning from Human FeedbackReinforcement LearningText
🎯 What it does: This paper proposes an Implicit Reward Model (PRM) that does not require stepwise labeling and can be implemented by training a standard Outcome Reward Model (ORM);
FreeMesh: Boosting Mesh Generation with Coordinates Merging
Jian Liu (Hong Kong University of Science and Technology), Chunchao Guo (Tencent Hunyuan)
GenerationCompressionTransformerPoint CloudMesh
🎯 What it does: This paper proposes a training-free evaluation metric PTME based on entropy theory, and improves the mesh tokenization compression rate and generation quality through coordinate rearrangement and merging techniques.
Freeze-Omni: A Smart and Low Latency Speech-to-speech Dialogue Model with Frozen LLM
Xiong Wang (Tencent), Long MA
TransformerLarge Language ModelSupervised Fine-TuningMultimodalityAudio
🎯 What it does: Proposes Freeze-Omni, a multimodal LLM that freezes LLMs to achieve low-latency speech-to-speech dialogue.
From Black Boxes to Transparent Minds: Evaluating and Enhancing the Theory of Mind in Multimodal Large Language Models
Xinyang Li (University of Science and Technology Beijing), Huimin Ma (University of Science and Technology Beijing)
Explainability and InterpretabilityTransformerLarge Language ModelVision Language ModelVideoTextMultimodality
🎯 What it does: A multi-modal theory of mind (ToM) evaluation dataset, GridToM, was constructed, and an interpretive analysis of the internal representations of multi-modal large language models was conducted without training intervention to enhance their ToM capabilities.
From Complex to Atomic: Enhancing Augmented Generation via Knowledge-Aware Dual Rewriting and Reasoning
Jinyu Wang (Microsoft Research Asia), Jiang Bian (Microsoft Research Asia)
GenerationRetrievalTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: An iterative RAG framework based on knowledge atomization, query generation, atomic retrieval, and atomic selection (KAR-RAG) is proposed to achieve adaptive decomposition and reasoning for complex multi-hop questions.
From Crowdsourced Data to High-quality Benchmarks: Arena-Hard and Benchbuilder Pipeline
Tianle Li (University of California), Ion Stoica (University of California)
TransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Proposes the BenchBuilder automation process to generate high-quality, challenging LLM benchmarks (Arena-Hard-Auto) from crowdsourced data.
From Debate to Equilibrium: Belief‑Driven Multi‑Agent LLM Reasoning via Bayesian Nash Equilibrium
Xie Yi, Bo Han (Hong Kong Baptist University)
Computational EfficiencyTransformerLarge Language ModelReinforcement LearningTextBenchmarkChain-of-Thought
🎯 What it does: The ECON framework is proposed, which achieves collaborative reasoning among multiple LLMs through Bayesian Nash equilibrium, avoiding the high costs of traditional multi-round communication.
From Feature Interaction to Feature Generation: A Generative Paradigm of CTR Prediction Models
Mingjia Yin, Enhong Chen
GenerationRecommendation SystemAuto EncoderTabular
🎯 What it does: A supervised feature generation framework is proposed, transforming the CTR model from discriminative feature interaction to generative feature generation.
From Individual Experience to Collective Evidence: A Reporting-Based Framework for Identifying Systemic Harms
Jessica Dai (University of California Berkeley), Benjamin Recht (University of California Berkeley)
TabularBiomedical DataFinance Related
🎯 What it does: A framework based on individual report databases is proposed to identify systemic harm using sequential hypothesis testing.
From Jack of All Trades to Master of One: Specializing LLM-based Autoraters to a Test Set
Mara Finkelstein (Google), Markus Freitag (Google)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Designed an LLM automatic evaluator (Specialist) that requires only multi-turn prompts and no fine-tuning.
From Kernels to Features: A Multi-Scale Adaptive Theory of Feature Learning
Noa Rubin (Hebrew University of Jerusalem), Moritz Helias (RWTH Aachen University)
Representation LearningImage
🎯 What it does: This paper constructs a multi-scale adaptive feature learning theory that unifies previous perspectives on kernel rescaling and kernel adaptation, providing precise statistical outputs of networks under standard and mean-field scaling.
From Language Models over Tokens to Language Models over Characters
Tim Vieira (ETH Zürich), Ryan Cotterell (ETH Zürich)
GenerationOptimizationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: A theoretical and algorithmic framework is proposed to transition from token-based language models to character-based interfaces, addressing the prompt boundary issue and achieving character-level probability estimation and conditional generation.
From Local Details to Global Context: Advancing Vision-Language Models with Attention-Based Selection
Lincan Cai (Beijing Institute of Technology), Jian Liang (Kuaishou Technology)
ClassificationDomain AdaptationRepresentation LearningTransformerLarge Language ModelVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: By utilizing the attention map from DINO to guide image cropping and performing cropping in the feature space to retain global semantics, combined with soft matching to filter fine-grained descriptions generated by LLM, the ABS (Attention-Based Selection) method is proposed to enhance the zero-shot and OOD generalization performance of CLIP.
From Logits to Hierarchies: Hierarchical Clustering made Simple
Emanuele Palumbo (ETH Zurich), Julia E Vogt
Computational EfficiencyImage
🎯 What it does: A lightweight hierarchical clustering method L2H based on pre-trained non-hierarchical clustering model logits is proposed, which can generate high-quality hierarchies without fine-tuning.
From Low Rank Gradient Subspace Stabilization to Low-Rank Weights: Observations, Theories, and Applications
AJAY KUMAR JAISWAL, Zhangyang Wang (University of Texas at Austin)
CompressionOptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper reveals through theoretical and empirical analysis the alignment relationship between gradient subspaces and Hessian matrix features, demonstrating that the weight matrices of large language models (LLMs) gradually converge to low-rank subspaces during the pre-training process. Based on this, a unified low-rank compression and parameter-efficient fine-tuning method called WeLore is proposed.
From Mechanistic Interpretability to Mechanistic Biology: Training, Evaluating, and Interpreting Sparse Autoencoders on Protein Language Models
Etowah Adams (Columbia University), Mohammed AlQuraishi (Columbia University)
Explainability and InterpretabilityProtein Structure PredictionAuto EncoderBiomedical Data
🎯 What it does: Train sparse autoencoders (SAE) to extract interpretable features from the residual flow of the protein language model ESM-2, interpret the features using the InterProt visualization tool, and evaluate their predictive performance in downstream tasks such as secondary structure, subcellular localization, thermal stability, and CHO expression using linear detectors.
From Passive to Active Reasoning: Can Large Language Models Ask the Right Questions under Incomplete Information?
Zhanke Zhou (Hong Kong Baptist University), Bo Han (Hong Kong Baptist University)
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkChain-of-Thought
🎯 What it does: This paper proposes the AR-Bench benchmark to evaluate the active reasoning capabilities of large language models (LLMs) in scenarios with missing information, and systematically assesses the performance of existing LLMs in active reasoning.
From Pixels to Perception: Interpretable Predictions via Instance-wise Grouped Feature Selection
Moritz Vandenhirtz (ETH Zurich), Julia E Vogt
ClassificationExplainability and InterpretabilityTransformerSupervised Fine-TuningImage
🎯 What it does: By performing instance-level binarization on images and sparsifying them based on semantically meaningful regions, interpretable predictions are achieved; at the same time, the sparsity threshold is dynamically adjusted during inference based on classifier confidence.
From RAG to Memory: Non-Parametric Continual Learning for Large Language Models
Bernal Jiménez Gutiérrez (Ohio State University), Yu Su (Ohio State University)
RetrievalTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes HippoRAG 2, an improvement over traditional RAG, which constructs a non-parametric continuous learning framework based on concept-context dense-sparse coding fusion, query-triple mapping, memory filtering recognition, and personalized PageRank.
From Spectrum-free towards Baseline-view-free: Double-track Proximity Driven Multi-view Clustering
Shengju Yu (National University of Defense Technology), Yi Zhang (National University of Defense Technology)
Tabular
🎯 What it does: A dual-track approach to multi-view clustering is proposed, combining the geometric relationships of anchor-point-sample and anchor-point-anchor.
From Theory to Practice: Rethinking Green and Martin Kernels for Unleashing Graph Transformers
Yoon Hyeok Lee (Samsung Electronics), Bosun Hwang
Representation LearningGraph Neural NetworkTransformerGraphBenchmark
🎯 What it does: This paper proposes structure encoding based on Green kernel and Martin kernel (GKSE, MKSE) and applies it to the Graph Transformer GRIT to enhance graph representation learning effectiveness.
From Thousands to Billions: 3D Visual Language Grounding via Render-Supervised Distillation from 2D VLMs
Ang Cao (University of Michigan), Alexander Sax (Meta)
SegmentationKnowledge DistillationTransformerVision Language ModelGaussian SplattingPoint Cloud
🎯 What it does: This paper proposes an unsupervised training framework LIFT-GS that distills knowledge from 2D visual language models to 3D point clouds through differentiable rendering, achieving 3D visual-language alignment.
From Token to Rhythm: A Multi-Scale Approach for ECG-Language Pretraining
Fuying Wang (University of Hong Kong), Lequan Yu (University of Hong Kong)
TransformerContrastive LearningMultimodalityTime SeriesBiomedical DataElectrocardiogram
🎯 What it does: This paper proposes a multi-scale ECG-language pre-training model MELP, which learns clinically interpretable ECG representations through cross-modal supervision at three levels: token, beat, and rhythm.
From Uncertain to Safe: Conformal Adaptation of Diffusion Models for Safe PDE Control
Peiyan Hu (Westlake University), Tailin Wu (Westlake University)
OptimizationSafty and PrivacyDiffusion modelTime SeriesPhysics Related
🎯 What it does: This paper proposes the SafeDiffCon method, which combines diffusion models with conformal prediction to meet safety constraints in PDE control tasks through post-training and fine-tuning during inference, thereby enhancing control performance.
From Weight-Based to State-Based Fine-Tuning: Further Memory Reduction on LoRA with Parallel Control
Chi Zhang (National University of Singapore), Qianxiao Li (National University of Singapore)
CompressionOptimizationComputational EfficiencyTransformerSupervised Fine-TuningImageText
🎯 What it does: Proposes a parameter tuning framework that transitions from weight-based LoRA to network state-based tuning, utilizing parallel control to reduce the storage of forward activations, further compressing GPU memory and computation time.
FRUGAL: Memory-Efficient Optimization by Reducing State Overhead for Scalable Training
Philip Zmushko (Yandex), Samuel Horváth (Mohamed bin Zayed University of Artificial Intelligence)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes a memory-efficient optimization framework named FRUGAL, which divides the parameter space into subspaces that require state information and those that do not, using stateful optimizers like Adam for the former and stateless optimizers like signSGD for the latter.
FSL-SAGE: Accelerating Federated Split Learning via Smashed Activation Gradient Estimation
Srijith Nair (Ohio State University), Jia Liu (Ohio State University)
Federated LearningComputational EfficiencyImageText
🎯 What it does: This paper proposes FSL-SAGE, a federated split learning framework that estimates server gradients through an auxiliary model;
FSTLLM: Spatio-Temporal LLM for Few Shot Time Series Forecasting
YUE JIANG, Gao Cong (Nanyang Technological University)
Graph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningTime Series
🎯 What it does: This paper proposes the FSTLLM framework, which combines large language models with spatiotemporal graph neural networks to address the few-shot time series forecasting problem.
Fully Dynamic Embedding into $\ell_p$ Spaces
Kiarash Banihashem (University of Maryland), Tong Yu (Adobe)
Graph
🎯 What it does: A fully dynamic algorithm is proposed for embedding graph metrics into ℓp space, supporting edge insertion and deletion while maintaining efficient embeddings after each update.
Fully Dynamic Euclidean Bi-Chromatic Matching in Sublinear Update Time
Gramoz Goranci (University of Vienna), Da Wei Zheng
Time Series
🎯 What it does: This paper presents the first fully dynamic bi-color matching algorithm that achieves sublinear update time in Euclidean two-dimensional space for rapid estimation of the W1-Wasserstein distance.
Fully Heteroscedastic Count Regression with Deep Double Poisson Networks
Spencer Young (Delicious AI), Hua Wei (Arizona State University)
ImageTextPoint CloudTabular
🎯 What it does: The Deep Double Poisson Network (DDPN) is proposed, achieving complete heteroscedastic uncertainty quantification in count regression and enhancing performance through integration.
FunBO: Discovering Acquisition Functions for Bayesian Optimization with FunSearch
Virginia Aglietti (Google DeepMind), Silvia Chiappa (Google DeepMind)
OptimizationHyperparameter SearchLarge Language ModelTabular
🎯 What it does: Using large language models in conjunction with FunSearch, we automatically explore and discover new acquisition functions (AF) that can enhance the performance of Bayesian optimization, represented in readable Python code.
Function Encoders: A Principled Approach to Transfer Learning in Hilbert Spaces
Tyler Ingebrand (University of Texas at Austin), ufuk topcu
Pose EstimationDomain AdaptationMeta LearningTransformerAuto EncoderImageTabular
🎯 What it does: Geometric characterization of transfer learning in Hilbert space, proposing three types of transfer (interpolation, linear extrapolation, Hilbert space extrapolation), and achieving rapid adaptation to any type of transfer using function encoders.
Function-Space Learning Rates
Edward Milsom (University of Bristol), Laurence Aitchison (University of Bristol)
OptimizationHyperparameter SearchTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes an efficient method for estimating and controlling the learning rate of the function space of neural network layers (i.e., the impact of parameter updates on the network output function) and introduces the FLeRM (Function-space Learning Rate Matching) technique based on this, which facilitates hyperparameter (learning rate) transfer when expanding model size (width, depth, initialization scale, LoRA dimensions, etc.).
Function-to-Style Guidance of LLMs for Code Translation
Longhui Zhang (Harbin Institute of Technology), Min Zhang (Harbin Institute of Technology)
AI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: This paper proposes the F2STRANS two-stage functionality-to-style guidance framework to enhance the functional correctness and readability of large language models in code translation.
Functional Alignment Can Mislead: Examining Model Stitching
Damian Smith (University of Southampton), Antonia Marcu (University of Southampton)
ClassificationRepresentation LearningConvolutional Neural NetworkImage
🎯 What it does: The researchers explored whether functional alignment can truly reflect the similarity of information captured by networks through a model stitching method, demonstrating that even when models are functionally aligned, they may represent entirely different input features.
Fundamental Bias in Inverting Random Sampling Matrices with Application to Sub-sampled Newton
Chengmei Niu (Huazhong University of Science and Technology), Michael W. Mahoney (University of California)
OptimizationImage
🎯 What it does: This study investigates the bias problem of the inverse of random sampling matrices, proposes a bias-correction sampling technique, and applies it to improve the convergence performance of the subsampled Newton method.
Fundamental limits of learning in sequence multi-index models and deep attention networks: high-dimensional asymptotics and sharp thresholds
Emanuele Troiani (Ecole Polytechnique Federale de Lausanne), Lenka Zdeborova (Ecole Polytechnique Federale de Lausanne)
TransformerTextSequential
🎯 What it does: This paper studies the learning theory of deep self-attention networks (composed of multiple layers of self-attention), first mapping them to the sequence multi-index model (SMI), and then deriving the Bayes-optimal performance and the performance of the AMP (GAMP) algorithm in the high-dimensional limit, providing the weak recovery thresholds for each layer and revealing the order of hierarchical learning.
Fundamental Limits of Visual Autoregressive Transformers: Universal Approximation Abilities
Yifang Chen (University of Chicago), Zhao Song (University of California)
GenerationData SynthesisTransformerFlow-based ModelImage
🎯 What it does: Theoretical analysis of the Visual Autoregressive Transformer (VAR) and FlowAR is conducted, proving that a single-layer single-head VAR Transformer can achieve any continuous mapping under the Lipschitz condition, thus serving as a universal approximator.
FuseUNet: A Multi-Scale Feature Fusion Method for U-like Networks
Quansong He (Sichuan University), Tao He (Sichuan University)
RestorationSegmentationConvolutional Neural NetworkImageBiomedical DataMagnetic Resonance ImagingOrdinary Differential Equation
🎯 What it does: Viewing the skip connections of U-Net as discrete nodes of numerical integration, we propose the FuseUNet multi-scale feature fusion method, which achieves adaptive high-order discretization using linear multi-step methods and neural memory ODEs (nmODEs) for prediction-correction.
Fusing Reward and Dueling Feedback in Stochastic Bandits
Xuchuang Wang (University of Massachusetts), Adam Wierman (California Institute of Technology)
OptimizationReinforcement LearningTabular
🎯 What it does: A random multi-armed bandit (DR-MAB) framework is proposed, where reward feedback and comparison feedback can be used simultaneously, and two fusion algorithms are designed.
G-Adaptivity: optimised graph-based mesh relocation for finite element methods
James Rowbottom (University of Cambridge), Chris Budd (University of Bath)
OptimizationGraph Neural NetworkMesh
🎯 What it does: A grid relocation method based on graph neural networks, G-Adaptivity, has been developed to directly minimize FEM errors, significantly improving the accuracy of finite element solutions while keeping the number of grid nodes unchanged.
G-Designer: Architecting Multi-agent Communication Topologies via Graph Neural Networks
Guibin Zhang (Tongji University), Dawei Cheng (Tongji University)
Graph Neural NetworkLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: Design an adaptive, scalable, and attack-resistant communication topology for LLM agents, and propose G-Designer to automate the generation of task-specific multi-agent networks.
G-Sim: Generative Simulations with Large Language Models and Gradient-Free Calibration
Samuel Holt (University of Cambridge), Mihaela van der Schaar (University of Cambridge)
GenerationOptimizationExplainability and InterpretabilityTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: A hybrid framework G-Sim is proposed, which utilizes large language models to automatically generate simulator structures and empirically match their parameters through gradient-independent and likelihood-independent calibration techniques, thereby constructing interpretable and intervenable simulators.
Galileo: Learning Global & Local Features of Many Remote Sensing Modalities
Gabriel Tseng (Mila - Quebec AI Institute), David Rolnick (Mila - Quebec AI Institute)
ClassificationSegmentationTransformerContrastive LearningMultimodality
🎯 What it does: Proposes the Galileo model, which learns multi-scale features of remote sensing through a multi-modal Transformer and dual-objective self-supervised learning.
Gamma Distribution PCA-Enhanced Feature Learning for Angle-Robust SAR Target Recognition
Chong Zhang (Xidian University), Mengke Li (Shenzhen University)
RecognitionObject DetectionConvolutional Neural NetworkTransformerImage
🎯 What it does: A Gamma Distribution-based Principal Component Analysis (Γ PCA) method is proposed to extract features that are insensitive to observation angles, thereby enhancing the robustness of Synthetic Aperture Radar (SAR) target recognition.
Gandalf the Red: Adaptive Security for LLMs
Niklas Pfister (Lakera), Mateo Rojas-Carulla (Lakera)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes a dynamic security utility threat model D-SEC, constructs a crowd-based red team platform Gandalf, and collects 279k adaptive prompt attack samples; it jointly evaluates the security and usability of LLM defenses using this dataset and user data, further exploring the impact of three strategies: domain restrictions, defense depth, and adaptive defenses on the security-usability trade-off.
GANQ: GPU-Adaptive Non-Uniform Quantization for Large Language Models
Pengxiang Zhao (University of Hong Kong), Xiaoming Yuan (University of Hong Kong)
GenerationCompressionOptimizationTransformerLarge Language ModelGenerative Adversarial NetworkText
🎯 What it does: This paper proposes GANQ, a GPU-adaptive post-training non-uniform quantization framework for weight quantization of large-scale language models that is compatible with LUT-based mpGEMM.
Gap-Dependent Bounds for Federated $Q$-Learning
Haochen Zhang (Pennsylvania State University), Lingzhou Xue (Pennsylvania State University)
Federated LearningReinforcement Learning
🎯 What it does: The paper presents a 'gap-dependent' theoretical analysis of the non-zero sub-optimality gap in Federated Q-learning (FedQ-Hoeffding), providing upper bounds on the logarithmic level (log T) of scheduling costs (communication cost) and regret in online Federated Reinforcement Learning (FRL).
GAPrompt: Geometry-Aware Point Cloud Prompt for 3D Vision Model
Zixiang Ai (Wangxuan Institute of Computer Technology Peking University), Jiahuan Zhou (Wangxuan Institute of Computer Technology Peking University)
ClassificationTransformerPrompt EngineeringPoint Cloud
🎯 What it does: A geometric-aware point cloud prompting method called GAPrompt is designed to efficiently fine-tune the model through learnable point cloud prompts and point displacement promptors while keeping the pre-trained 3D vision model unchanged.
Gaussian Mixture Flow Matching Models
Hansheng Chen (Stanford University), Sai Bi (Adobe Research)
GenerationData SynthesisFlow-based ModelImageStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: A Gaussian Mixture Flow Matching (GMFlow) model is proposed, which predicts flow velocity using a multimodal Gaussian mixture distribution to improve the quality of few-step sampling.
GaussMark: A Practical Approach for Structural Watermarking of Language Models
Adam Block (Columbia University), Ayush Sekhari (MIT)
GenerationLarge Language ModelText
🎯 What it does: This paper studies a structural watermarking method based on Gaussian perturbation called GaussMark, which is used to embed imperceptible watermarks during text generation by LLMs and provides a detection mechanism.
GaussMarker: Robust Dual-Domain Watermark for Diffusion Models
Kecen Li (Institute of Automation, Chinese Academy of Sciences), Cheng Hong (Ant Group)
GenerationData SynthesisAnomaly DetectionDiffusion modelImage
🎯 What it does: This paper proposes and implements GaussMarker, a no-fine-tuning method for dual-domain watermarking in diffusion models that utilizes both spatial and frequency domains simultaneously, and incorporates a learnable Gaussian Noise Restorer (GNR) to enhance robustness against attacks such as rotation and cropping.
GCAL: Adapting Graph Models to Evolving Domain Shifts
Ziyue Qiao (Great Bay University), Hui Xiong (Hong Kong University of Science and Technology)
Domain AdaptationOptimizationGraph Neural NetworkContrastive LearningGraphTime Series
🎯 What it does: Proposes the GCAL framework, achieving continuous adaptation on graph data with continuous unsupervised domain drift, utilizing a dual-layer optimization of adaptation and memory replay.
GEFA: A General Feature Attribution Framework Using Proxy Gradient Estimation
Yi Cai (Freie Universitaet Berlin), Gerhard Wunder (Freie Universitaet Berlin)
Explainability and InterpretabilityImageText
🎯 What it does: A general feature attribution framework GEFA based on proxy gradient estimation is proposed, which can explain any model in a black-box environment.
General agents need world models
Jonathan Richens (Google DeepMind), David Abel (Google DeepMind)
Robotic IntelligenceReinforcement LearningAgentic AIWorld ModelSequential
🎯 What it does: This paper proves that any agent with the ability to generalize multi-step goal tasks (i.e., a goal-conditioned agent that meets a given reward bound) must learn a predictable world model of the environment, and provides an algorithm for extracting this model from the agent's policy;
General framework for online-to-nonconvex conversion: Schedule-free SGD is also effective for nonconvex optimization
Kwangjun Ahn (Microsoft Research), Ashok Cutkosky (Boston University)
Optimization
🎯 What it does: A general framework for converting online to non-convex optimization is proposed, and it is proven that under this framework, Schedule-Free SGD (stochastic gradient descent without learning rate scheduling) can achieve optimal iteration complexity on non-smooth, non-convex objective functions.
Generalists vs. Specialists: Evaluating LLMs on Highly-Constrained Biophysical Sequence Optimization Tasks
Angelica Chen (New York University), Nathan C. Frey (Genentech)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: An evaluation framework called LLOME is proposed for the optimization task of biomolecular sequences with high constraints, and its effectiveness is validated through a novel Ehrlich function test set.
Generalizable Multi-Camera 3D Object Detection from a Single Source via Fourier Cross-View Learning
Xue Zhao (Shanghai Jiao Tong University), Nanyang Ye (Shanghai Jiao Tong University)
Object DetectionDomain AdaptationAutonomous DrivingTransformerPoint Cloud
🎯 What it does: This paper addresses multi-camera 3D object detection trained on single-source data and proposes the FCVL framework to enhance the model's generalization performance in unseen domains.
Generalization and Robustness of the Tilted Empirical Risk
Gholamali Aminian (Alan Turing Institute), Samuel N. Cohen (University of Oxford)
Tabular
🎯 What it does: This study investigates the generalization error and robustness of Tilted Empirical Risk Minimization (TERM) under negative tilt and when the loss function satisfies the (1+ε)-th moment being finite, providing a unified approach and information-theoretic framework for data-driven selection of the tilt parameter.
Generalization Analysis for Controllable Learning
Yi-Fan Zhang (Southeast University), Min-Ling Zhang (Southeast University)
Transformer
🎯 What it does: This paper constructs a unified theoretical framework for controllable learning and presents a new vector contraction inequality for vector-valued function classes, resulting in a generalization error upper bound that is independent of the number of task objectives (only logarithmic level); it also provides specific generalization bounds for embedded and hypernetwork-based controllable learning methods, along with a Rademacher complexity module for controllers and predictors (such as FNNs and Transformers);
Generalization Analysis for Supervised Contrastive Representation Learning under Non-IID Settings
Nong Minh Hieu (Singapore Management University), Antoine Ledent (Singapore Management University)
ClassificationRepresentation LearningContrastive LearningImage
🎯 What it does: This paper studies the generalization behavior of supervised contrastive learning (CRL) under non-i.i.d. settings, proposing a theoretical framework based on U-statistics and subsampling risk, along with corresponding generalization bounds.
Generalization Bounds via Meta-Learned Model Representations: PAC-Bayes and Sample Compression Hypernetworks
Benjamin Leblanc (Université Laval), Pascal Germain (Université Laval)
CompressionMeta LearningImage
🎯 What it does: This paper proposes a meta-learning framework based on hypernetworks, deriving a non-vacuous generalization upper bound for each downstream predictor using PAC-Bayes and sample compression theory.
Generalization in Federated Learning: A Conditional Mutual Information Framework
Ziqiao Wang (Tongji University), Yongyi Mao (University of Ottawa)
Federated LearningSafty and PrivacyConvolutional Neural NetworkImage
🎯 What it does: A generalized framework for federated learning based on Conditional Mutual Information (CMI) is proposed and implemented, deriving a theoretical upper bound for the two-level generalization error and providing an upper bound for the rapid convergence of the evaluative CMI.
Generalization of noisy SGD in unbounded non-convex settings
Leello Tadesse Dadi (École Polytechnique Fédérale de Lausanne), Volkan Cevher (École Polytechnique Fédérale de Lausanne)
OptimizationSafty and PrivacyStochastic Differential Equation
🎯 What it does: This study investigates the generalization ability of iterative noise gradient schemes under smooth non-convex loss and establishes a time-independent information-theoretic generalization bound.
Generalization Performance of Ensemble Clustering: From Theory to Algorithm
Xu Zhang (Southeast University), Yuheng Jia (Southeast University)
OptimizationTabular
🎯 What it does: This paper proposes a theoretical analysis framework for ensemble clustering and designs a new ensemble clustering algorithm based on this framework. First, the upper bounds for the generalization error and overfitting risk of ensemble clustering are provided, proving them to be O(√(log n m) + 1/√n), along with sufficient conditions for consistency (m ≫ log n). Secondly, the error is decomposed into a Bias-Diversity structure, indicating that minimizing the bias of the base distribution and maximizing diversity can reduce error, and equating diversity enhancement to robust min-max optimization. Finally, an algorithm based on high-confidence elements for K* approximation and weight learning is implemented, using reduced gradient descent for solving.