ICML 2025 Papers — Page 13
International Conference on Machine Learning · 3257 papers
Generalization Principles for Inference over Text-Attributed Graphs with Large Language Models
Haoyu Peter Wang (Georgia Institute of Technology), Pan Li (Georgia Institute of Technology)
ClassificationGraph Neural NetworkLarge Language ModelPrompt EngineeringTextGraph
🎯 What it does: This paper proposes an unsupervised, zero-shot graph classification method called LLM-BP, which utilizes large language models for task-adaptive text embedding and Bayesian propagation aggregation of graph structures.
Generalized additive models via direct optimization of regularized decision stump forests
Magzhan Gabidolla (University of California), Miguel Á. Carreira-Perpiñán (University of California)
OptimizationTabularBenchmark
🎯 What it does: This paper proposes a general additive model (GAM) method for directly optimizing a fixed number of decision stump forests, avoiding the use of traditional boosting or bagging techniques.
Generalized Category Discovery via Reciprocal Learning and Class-Wise Distribution Regularization
Duo Liu (Shanghai Jiao Tong University), Weiran Huang
ClassificationKnowledge DistillationTransformerContrastive LearningImage
🎯 What it does: A recursive learning and category distribution regularization (RLCD) framework is proposed to improve the performance of base class discrimination and new class discovery in the general category discovery (GCD) task.
Generalized Interpolating Discrete Diffusion
Dimitri von Rütte (ETH Zurich), Thomas Hofmann (ETH Zurich)
GenerationData SynthesisTransformerLarge Language ModelDiffusion modelText
🎯 What it does: A Generalized Interpolated Discrete Diffusion (GIDD) framework is proposed, extending masked diffusion to arbitrary mixed distributions, providing closed-form forward/backward transitions and ELBO; and achieving self-correction under a masked + uniform noise mixture.
Generalized Random Forests Using Fixed-Point Trees
David Fleischer (McGill University), Archer Y. Yang (Mila - Quebec AI Institute)
OptimizationComputational EfficiencyTabular
🎯 What it does: A fixed-point iteration-based gradient-free tree splitting method is proposed to accelerate the training of Generalized Random Forest (GRF) and reduce the computational cost of gradient/Jacobian matrix estimation.
Generalized Smooth Bilevel Optimization with Nonconvex Lower-Level
Siqi Zhang (Nanjing University of Aeronautics and Astronautics), Feihu Huang (Nanjing University of Aeronautics and Astronautics)
OptimizationMeta LearningImage
🎯 What it does: A new penalty regularization normalized gradient algorithm (PNGBiO) and its stochastic version (S-PNGBiO) are proposed to solve bi-level optimization problems where both the upper and lower levels are generally smooth and the lower level is non-convex;
Generalized Venn and Venn-Abers Calibration with Applications in Conformal Prediction
Lars van der Laan (University of Washington), Ahmed Alaa (University of California Berkeley)
TabularBiomedical DataElectronic Health Records
🎯 What it does: A unified Venn and Venn-Abers calibration framework is proposed, which can extend any point calibrator that achieves perfect sample calibration to a set calibrator, and a Venn multi-calibration method is provided.
Generalizing Causal Effects from Randomized Controlled Trials to Target Populations across Diverse Environments
Baohong Li (Zhejiang University), Kun Kuang (Zhejiang University)
Tabular
🎯 What it does: This study investigates methods for extrapolating the causal effects of randomized controlled trials (RCTs) to target populations under different environments.
Generalizing from SIMPLE to HARD Visual Reasoning: Can We Mitigate Modality Imbalance in VLMs?
Simon Park (Princeton University), Sanjeev Arora (Princeton University)
TransformerSupervised Fine-TuningVision Language ModelImageTextMultimodalityChain-of-Thought
🎯 What it does: Three controllable algorithmic visual reasoning tasks (table reading, grid navigation, visual analogy) have been constructed, each with SIMPLE and HARD difficulty levels. The system evaluates the modal imbalance and vulnerability of VLM when transitioning from SIMPLE to HARD, and various supervision strategies (text, image, image→text, mixed, mixed+hard text, two-stage alignment) are employed to mitigate this issue.
Generating Hypotheses of Dynamic Causal Graphs in Neuroscience: Leveraging Generative Factor Models of Observed Time Series
Zachary C. Brown (Duke University), David Carlson (Duke University)
GenerationData SynthesisOptimizationRecurrent Neural NetworkGenerative Adversarial NetworkTime SeriesSequentialBiomedical Data
🎯 What it does: The REDCLIFF-S method is proposed for automatically generating scientific hypotheses in dynamic causal graphs from time series data.
Generation from Noisy Examples
Ananth Raman (Bridgewater-Raritan Regional High School), Vinod Raman (University of Michigan)
Generation
🎯 What it does: This paper introduces a noise example flow under the framework of generative theory, studying the generability of the generator under limited noise and providing a corresponding concept of generability.
Generative Audio Language Modeling with Continuous-valued Tokens and Masked Next-Token Prediction
Shu-wen Yang (National Taiwan University), Chao Wang (Amazon AGI)
GenerationData SynthesisTransformerDiffusion modelAudio
🎯 What it does: A framework for audio generation based on the Transformer decoder is proposed, using continuous-valued audio tokens and introducing a Masked Next Token Prediction (MNTP) task.
Generative Data Mining with Longtail-Guided Diffusion
David S Hayden, Siddhartha Srinivasa
GenerationData SynthesisVision Language ModelDiffusion modelImage
🎯 What it does: By combining the long-tail confidence signals of a pre-trained predictive model with a diffusion model, high-quality synthetic data can be generated for samples that are difficult for the model to recognize or are rare, without the need to retrain either model, and these data can be used to enhance the model's generalization performance.
Generative Human Trajectory Recovery via Embedding-Space Conditional Diffusion
Kaijun Liu (Nanyang Technological University), liang yu
GenerationData SynthesisGraph Neural NetworkTransformerDiffusion modelTime SeriesSequential
🎯 What it does: This paper proposes DiffMove, a conditionally diffusion model based on embedding space, designed to recover missing locations in sparse human trajectories.
Generative Intervention Models for Causal Perturbation Modeling
Nora Schneider (Helmholtz Munich), Andreas Krause (ETH Zurich)
OptimizationExplainability and InterpretabilityDrug DiscoveryBiomedical Data
🎯 What it does: A Generative Intervention Model (GIM) is proposed, which learns to map perturbation features to atomic interventions in a causal model, thereby predicting distribution changes under unknown perturbations.
Generative Modeling Reinvents Supervised Learning: Label Repurposing with Predictive Consistency Learning
Yang Li (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
SegmentationGenerationSupervised Fine-TuningDiffusion modelImageText
🎯 What it does: This paper proposes a new learning paradigm called Predictive Consistency Learning (PCL), which constructs multi-scale label prompts for progressive learning by gradually denoising complex labels and jointly modeling them with input data.
Generative Point Cloud Registration
Haobo Jiang (Nanyang Technological University), Jianmin Zheng (Nanyang Technological University)
GenerationData SynthesisOptimizationPrompt EngineeringDiffusion modelImagePoint Cloud
🎯 What it does: Using the deep conditional 2D generative model Match-ControlNet to generate RGB image pairs that are geometrically consistent with the source and target point clouds and texture compatible, and integrating the visual features of these synthetic images with geometric features to enhance point cloud registration accuracy.
Generative Social Choice: The Next Generation
Niclas Boehmer (Hasso Plattner Institute), Ariel D. Procaccia (Harvard University)
GenerationRecommendation SystemTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: A generative social choice framework with budget constraints and approximate queries is proposed to automatically generate a set of proportional representation statements from vast amounts of text.
GenMol: A Drug Discovery Generalist with Discrete Diffusion
Seul Lee (KAIST), Arash Vahdat (NVIDIA)
Drug DiscoveryTransformerDiffusion modelGraph
🎯 What it does: GenMol is proposed, a universal drug discovery framework that uses a single discrete diffusion model to generate SAFE sequences, capable of performing various drug discovery tasks such as de novo generation, fragment-constrained generation, goal-directed hit generation, and lead optimization simultaneously.
GenZSL: Generative Zero-Shot Learning Via Inductive Variational Autoencoder
Shiming Chen (Mohamed bin Zayed University of Artificial Intelligence), Fahad Shahbaz Khan (Mohamed bin Zayed University of Artificial Intelligence)
GenerationData SynthesisAuto EncoderImage
🎯 What it does: A generative zero-shot learning model GenZSL based on inductive variational autoencoders is proposed, which utilizes weak semantic vectors to induce visual features of unseen categories from similar known categories.
Geometric Algebra Planes: Convex Implicit Neural Volumes
Irmak Sivgin (Stanford University), Mert Pilanci (Stanford University)
SegmentationGenerationOptimizationNeural Radiance FieldVideo
🎯 What it does: A volume parameterization based on geometric algebra, GA-Planes, is proposed, which can be trained under convex or non-convex optimization.
Geometric and Physical Constraints Synergistically Enhance Neural PDE Surrogates
Yunfei Huang (Helmholtz Centre Hereon), David S. Greenberg (Helmholtz Centre Hereon)
Convolutional Neural NetworkTime SeriesPhysics Related
🎯 What it does: A neural PDE proxy that takes into account both physical constraints and geometric symmetries is proposed, along with a new input/output layer designed for the discretized staggered grid.
Geometric Contact Flows: Contactomorphisms for Dynamics and Control
Andrea Testa (Bosch Center for Artificial Intelligence), Leonel Rozo (Bosch Center for Artificial Intelligence)
OptimizationRobotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningGraphTime SeriesPhysics RelatedOrdinary Differential Equation
🎯 What it does: A new Geometric Contact Flow (GCF) framework is proposed for modeling and predicting complex dynamical systems, particularly those involving force exchange and dissipation.
Geometric Feature Embedding for Effective 3D Few-Shot Class Incremental Learning
Xiangqi Li (Institute of Computing Technology), Yongjun Xu (Institute of Computing Technology)
ClassificationRecognitionTransformerPoint Cloud
🎯 What it does: Proposes the 3D-FLEG method, which enhances text prompts by embedding geometric features in 3D few-shot class incremental learning, reducing reliance on high-quality prompts;
Geometric Generative Modeling with Noise-Conditioned Graph Networks
Peter Pao-Huang (Stanford University), Xiaojie Qiu (Stanford University)
GenerationData SynthesisGraph Neural NetworkTransformerDiffusion modelFlow-based ModelImagePoint Cloud
🎯 What it does: This paper proposes a Noise Conditioned Graph Network (NCGN) and its specific implementation, Dynamic Message Passing (DMP), to dynamically adjust the connectivity range and resolution of the graph based on different noise levels during the generation process in stream-based generative models (diffusion, flow-matching), thereby improving the quality of geometric shape generation (3D point clouds, spatial transcriptomics, images).
Geometric Hyena Networks for Large-scale Equivariant Learning
Artem Moskalev (Johnson and Johnson Innovative Medicine), Tommaso Mansi
Protein Structure PredictionConvolutional Neural NetworkBiomedical Data
🎯 What it does: Designed and implemented Geometric Hyena, a long convolution-based SE(3) equivariant network, to efficiently capture the global geometric context of large-scale molecules and proteins while maintaining rotational and translational equivariance.
Geometric Median (GM) Matching for Robust k-Subset Selection from Noisy Data
Anish Acharya (University of Texas), Inderjit S Dhillon
OptimizationImage
🎯 What it does: A robust data pruning method based on geometric median matching is proposed, which iteratively selects k subset samples to approximate the geometric median of the data;
Geometric Representation Condition Improves Equivariant Molecule Generation
Zian Li (Peking University), Muhan Zhang (Peking University)
GenerationDrug DiscoveryGraph Neural NetworkDiffusion modelGraph
🎯 What it does: Proposes the GeoRCG framework, which generates semantic representations using a pre-trained geometric encoder before molecular generation.
Geometric Resampling in Nearly Linear Time for Follow-the-Perturbed-Leader with Best-of-Both-Worlds Guarantee in Bandit Problems
Botao Chen (Kyoto University), Junya Honda (RIKEN AIP)
OptimizationReinforcement Learning
🎯 What it does: This paper studies the complexity and optimality of the Follow-the-Perturbed-Leader (FTPL) strategy in the K-armed bandit problem. A new technique called Conditional Geometric Resampling (CGR) is proposed for unbiased loss estimation, significantly reducing the average complexity per round.
Geometry Informed Tokenization of Molecules for Language Model Generation
Xiner Li (Texas A&M University), Shuiwang Ji (Texas A&M University)
GenerationDrug DiscoveryTransformerLarge Language ModelDiffusion modelGraph
🎯 What it does: Proposes the Geo2Seq method, which discretizes 3D molecular geometric structures into a one-dimensional token sequence invariant to SE(3), and utilizes large language models to generate 3D molecules.
Geometry-Informed Neural Networks
Arturs Berzins (LIT AI Lab, Institute for Machine Learning, Johannes Kepler University Linz), Johannes Brandstetter (Emmi AI GmbH)
GenerationOptimizationGenerative Adversarial NetworkMesh
🎯 What it does: This paper proposes Geometry-Informed Neural Networks (GINNs), a method for training shape-generating neural fields using geometric constraints and objectives without data.
GeoPixel: Pixel Grounding Large Multimodal Model in Remote Sensing
Akashah Shabbir (Mohamed bin Zayed University of Artificial Intelligence), Salman Khan (Australian National University)
Object DetectionSegmentationTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes GeoPixel, which enables pixel-level visual language dialogue and multi-object segmentation for high-resolution remote sensing images, generating fine-grained natural language descriptions.
GHOST: Generalizable One-Shot Federated Graph Learning with Proxy-Based Topology Knowledge Retention
Jiaru Qian (Wuhan University), Mang Ye (Wuhan University)
Federated LearningSafty and PrivacyGraph Neural NetworkGraph
🎯 What it does: A one-round communication federated graph learning framework GHOST is proposed, which utilizes proxy models to capture multi-dimensional knowledge from each client and integrates it into the global model, while reducing catastrophic forgetting through a topology knowledge retention mechanism.
GIVE: Structured Reasoning of Large Language Models with Knowledge Graph Inspired Veracity Extrapolation
Jiashu He (University of Pennsylvania), Alejandro Ribeiro (University of Pennsylvania)
TransformerLarge Language ModelPrompt EngineeringTextBiomedical DataBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: A structured reasoning framework named GIVE is designed, which integrates the internal knowledge of LLM with a limited external knowledge graph, guiding LLM to perform multi-step reasoning through 'veracity extrapolation'.
GLGENN: A Novel Parameter-Light Equivariant Neural Networks Architecture Based on Clifford Geometric Algebras
Ekaterina Filimoshina (HSE University), Dmitry Shirokov (Institute for Information Transmission Problems of the Russian Academy of Sciences)
🎯 What it does: A parameter-lightweight equivariant neural network architecture GLGENN is proposed, which utilizes geometric (Clifford) algebra to achieve equivariance for arbitrary pseudo-orthogonal transformations and implements parameter compression through weight sharing.
Global Context-aware Representation Learning for Spatially Resolved Transcriptomics
Yunhak Oh (KAIST), Chanyoung Park (KAIST)
Representation LearningGraph Neural NetworkAuto EncoderContrastive LearningBiomedical DataAlzheimer's Disease
🎯 What it does: A self-supervised representation learning framework named Spotscape is proposed to generate high-quality cell spot representations from spatially resolved transcriptomic data, achieving integration and alignment of single-slice and multi-slice data.
Global Convergence and Rich Feature Learning in $L$-Layer Infinite-Width Neural Networks under $\mu$ Parametrization
Zixiang Chen (University of California), Quanquan Gu (University of California)
OptimizationRepresentation LearningImage
🎯 What it does: This paper studies how an L-layer multilayer perceptron with maximum update parameterization (μP) can achieve significant feature learning during training while maintaining feature linear independence and ultimately converging to a global optimum under infinite width.
Global curvature for second-order optimization of neural networks
Alberto Bernacchia (MediaTek Research)
OptimizationTabular
🎯 What it does: The global curvature matrix structure of multilayer perceptrons is studied and derived, utilizing symmetry to achieve a computable second-order optimization preconditioning matrix.
Global Optimization with a Power-Transformed Objective and Gaussian Smoothing
Chen Xu (Shenzhen MSU-BIT University)
OptimizationAdversarial AttackGaussian SplattingImage
🎯 What it does: This paper proposes a new global optimization method called GS-PowerOpt, which first amplifies the weight of the global maximum point of the objective function through power transformation (or exponential power transformation), then applies Gaussian smoothing to the transformed function, and uses zero-order stochastic gradient ascent iteration to find the maximum point.
Global-Local Dirichlet Processes for Clustering Grouped Data in the Presence of Group-Specific Idiosyncratic Variables
Arhit Chakrabarti (Texas A&M University), Bani Mallick
OptimizationBiomedical Data
🎯 What it does: A global-local Dirichlet process (GLocal DP) model is proposed for clustering grouped data with shared variables and group-specific variables simultaneously.
GMAIL: Generative Modality Alignment for generated Image Learning
Shentong Mo (Carnegie Mellon University), Sukmin Yun (Hanyang University)
GenerationRetrievalTransformerVision Language ModelDiffusion modelContrastive LearningImageMultimodality
🎯 What it does: Proposes the GMAIL framework, which uses cross-modal alignment to map generated images and real images into the same latent space, and fine-tunes the visual language model based on this.
Goal-Oriented Skill Abstraction for Offline Multi-Task Reinforcement Learning
Jinmin He (Chinese Academy of Sciences), Jian Cheng (Chinese Academy of Sciences)
Robotic IntelligenceTransformerReinforcement LearningSequential
🎯 What it does: In the offline multi-task reinforcement learning framework, Goal-Oriented Skill Abstraction (GO-Skill) is proposed, which constructs a reusable skill library through goal-oriented skill extraction, discrete vector quantization, skill enhancement, and focal loss, and combines skills hierarchically to complete multiple tasks;
Going Deeper into Locally Differentially Private Graph Neural Networks
Longzhu He (Beijing University of Posts and Telecommunication), Sen Su (Beijing University of Posts and Telecommunication)
ClassificationSafty and PrivacyGraph Neural NetworkGraph
🎯 What it does: A privacy-preserving graph neural network framework named UPGNET is designed and implemented, which enhances the learning effectiveness of node classification tasks by injecting noise into node features under the local differential privacy (LDP) mechanism, followed by aggregation and regularization.
GoIRL: Graph-Oriented Inverse Reinforcement Learning for Multimodal Trajectory Prediction
Muleilan Pei (Hong Kong University of Science and Technology), Shaojie Shen (Hong Kong University of Science and Technology)
Autonomous DrivingGraph Neural NetworkReinforcement LearningMultimodality
🎯 What it does: A new graph-oriented inverse reinforcement learning framework (GoIRL) is proposed for multimodal trajectory prediction, aimed at addressing the trajectory prediction problem of surrounding agents in autonomous driving.
GPEN: Global Position Encoding Network for Enhanced Subgraph Representation Learning
Nannan Wu (Tianjin University), Wenjun Wang (Hainan Tropical Ocean University)
Representation LearningGraph Neural NetworkGraph
🎯 What it does: This paper proposes a novel subgraph representation learning framework called GPEN, which utilizes a tree structure for global position encoding and combines boundary-aware convolution to enhance the accuracy of subgraph representation.
GPTAQ: Efficient Finetuning-Free Quantization for Asymmetric Calibration
Yuhang Li (Yale University), Priyadarshini Panda (Yale University)
CompressionComputational EfficiencyTransformerLarge Language ModelImageText
🎯 What it does: This paper proposes GPTAQ, a quantization method that does not require fine-tuning, achieving asynchronous calibration by ensuring the consistency of outputs after layer-aligned quantization with the outputs of the full-precision model.
GRADEO: Towards Human-Like Evaluation for Text-to-Video Generation via Multi-Step Reasoning
Zhun Mou (Tsinghua University), Jiaya Jia (Hong Kong University of Science and Technology)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoTextMultimodalityChain-of-Thought
🎯 What it does: This paper proposes a multi-dimensional video evaluation framework called GRADEO, and trains an evaluation model capable of multi-step reasoning, GRADEO, by constructing the GRADEO-Instruct dataset.
Gradient Aligned Regression via Pairwise Losses
Dixian Zhu (Stanford University), Livnat Jerby (Stanford University)
Convolutional Neural NetworkRecurrent Neural NetworkTabular
🎯 What it does: This paper proposes Gradient Aligned Regression (GAR), which improves regression models in the label space through two pairwise losses (difference matching and normalized difference matching).
Gradient Boosting Reinforcement Learning
Benjamin Fuhrer (NVIDIA), Gal Dalal (NVIDIA)
Reinforcement LearningTabular
🎯 What it does: This paper proposes the Gradient Boosting Reinforcement Learning (GBRL) framework, which uses Gradient Boosting Trees (GBT) as function approximators in reinforcement learning, alternately building trees and interacting with the environment in real-time.
Gradient Descent Converges Arbitrarily Fast for Logistic Regression via Large and Adaptive Stepsizes
Ruiqi Zhang (University of California), Peter Bartlett
OptimizationTabular
🎯 What it does: In this paper, the author studies the convergence properties of gradient descent (GD) when using adaptive large step sizes on linearly separable data, proving that after reaching the burn-in period bounded by the margin, the average iterations of GD can achieve exponential fast convergence;
Gradient Flow Provably Learns Robust Classifiers for Orthonormal GMMs
Hancheng Min (University of Pennsylvania), Rene Vidal
ClassificationOptimizationAdversarial Attack
🎯 What it does: This paper studies how to learn classifiers with provable robustness under the orthonormal Gaussian mixture model using standard learning methods, without the need for defensive mechanisms.
Gradient Inversion of Multimodal Models
Omri Ben Hemo (Ben Gurion University of the Negev), Asaf Shabtai (Ben Gurion University of the Negev)
Federated LearningAdversarial AttackTransformerVision Language ModelImageMultimodality
🎯 What it does: This paper studies gradient inversion attacks on multimodal visual-language document visual question answering (DQA) models in a federated learning environment, proposing the GI-DQA method.
Gradient-based Explanations for Deep Learning Survival Models
Sophie Hanna Langbein (Leibniz Institute for Prevention Research and Epidemiology), Marvin N. Wright (Leibniz Institute for Prevention Research and Epidemiology)
Explainability and InterpretabilityMultimodalityTabular
🎯 What it does: A temporal gradient interpretability method for deep survival models is proposed and implemented, including GradSHAP(t), etc.;
GradPS: Resolving Futile Neurons in Parameter Sharing Network for Multi-Agent Reinforcement Learning
Haoyuan Qin (Xiamen University), Cheng Wang (Xiamen University)
Reinforcement Learning
🎯 What it does: The GradPS method is proposed, which identifies and clones 'ineffective neurons' using gradient conflict, sharing parameters according to conflict groups, thereby improving sample efficiency and policy diversity in multi-agent reinforcement learning.
Gradual Transition from Bellman Optimality Operator to Bellman Operator in Online Reinforcement Learning
Motoki Omura (University of Tokyo), Tatsuya Harada (RIKEN)
Reinforcement LearningTabular
🎯 What it does: The study transitions the Bellman optimal operator to the Bellman operator gradually to enhance sample efficiency in online RL and reduce estimation bias.
GRAIL: Graph Edit Distance and Node Alignment using LLM-Generated Code
Samidha Verma (Indian Institute of Technology Delhi), Sayan Ranu (Indian Institute of Technology Delhi)
OptimizationExplainability and InterpretabilityGraph Neural NetworkLarge Language ModelPrompt EngineeringGraph
🎯 What it does: This paper presents GRAIL, a method that uses large language models to automatically generate programs for approximating the graph edit distance (GED);
GRAM: A Generative Foundation Reward Model for Reward Generalization
Chenglong Wang (Northeastern University), JingBo Zhu
Recommendation SystemReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: A Generative Reward Model (GRAM) is proposed, which learns the input-response relationship using unlabeled data through unsupervised pre-training and supervised fine-tuning in two stages, and then fine-tunes with a small amount of labeled preference data.
Grammar-Forced Translation of Natural Language to Temporal Logic using LLMs
William H English, Rickard Ewetz (University of Florida)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: The GraFT framework is proposed, which breaks down the process of converting natural language (NL) to temporal logic (TL) into two steps: the induction of atomic propositions (AP) based on a masked language model and sequence-to-sequence translation based on grammatical constraints, significantly reducing the search space and improving learning efficiency.
Graph Adaptive Autoregressive Moving Average Models
Moshe Eliasof (University of Cambridge), Carola-Bibiane Schönlieb
Graph Neural NetworkGraphTime Series
🎯 What it does: A framework called GRAMA is proposed, which converts graph structures into time series and learns selectable autoregressive moving average (ARMA) models on them, seamlessly integrating with any GNN architecture;
Graph Attention is Not Always Beneficial: A Theoretical Analysis of Graph Attention Mechanisms via Contextual Stochastic Block Models
Zhongtian Ma (Northwestern Polytechnical University), Zhen Wang (Northwestern Polytechnical University)
ClassificationGraph Neural NetworkGraph
🎯 What it does: A systematic analysis of the theoretical properties of the graph attention mechanism under the Contextual Stochastic Block Model (CSBM) framework reveals its sensitivity to structural noise and feature noise, explores its role in the over-smoothing problem, and based on this, designs a multi-layer GAT structure that can achieve perfect node classification.
Graph Diffusion for Robust Multi-Agent Coordination
Xianghua Zeng (Beihang University), Zhiyuan LIN
Robotic IntelligenceGraph Neural NetworkTransformerReinforcement LearningDiffusion modelGraph
🎯 What it does: This paper proposes an offline multi-agent coordination framework MCGD based on graph diffusion to overcome the robustness issues caused by traditional diffusion methods that neglect coordination structures.
Graph Generative Pre-trained Transformer
Xiaohui Chen (Tufts University), Liping Liu
GenerationData SynthesisDrug DiscoveryGraph Neural NetworkTransformerReinforcement LearningGraph
🎯 What it does: A self-regressive graph generation model G2PT based on Transformer is proposed, which uses tokenized graph sequence representation and can directly predict the graph structure by generating the next token.
Graph Inverse Style Transfer for Counterfactual Explainability
Bardh Prenkaj (Technical University of Munich), Gjergji Kasneci (Technical University of Munich)
GenerationExplainability and InterpretabilityGraph Neural NetworkTransformerGraph
🎯 What it does: A reverse generation framework GIST based on graph style transfer is proposed, which generates graph counterfactual explanations that conform to structure and semantics using a backward backtracking approach.
Graph Minimum Factorization Distance and Its Application to Large-Scale Graph Data Clustering
Jicong Fan (Chinese University of Hong Kong)
OptimizationGraph Neural NetworkGraphStochastic Differential Equation
🎯 What it does: This paper proposes the Minimum Factorization Distance of graphs (MMFD) and its generalization MFD, which are used to measure the similarity between graphs, and based on this, designs a large-scale clustering algorithm (MMFD-KM, MFD-KD) that operates in linear time.
Graph Neural Network Generalization With Gaussian Mixture Model Based Augmentation
Yassine ABBAHADDOU, Michalis Vazirgiannis
ClassificationGraph Neural NetworkGraph
🎯 What it does: A Gaussian Mixture Model-based graph data augmentation method called GRATIN is proposed to enhance the generalization and robustness of graph neural networks in graph classification tasks.
Graph World Model
Tao Feng (University of Illinois Urbana Champaign), Jiaxuan You (University of Illinois Urbana Champaign)
GenerationRetrievalRecommendation SystemOptimizationGraph Neural NetworkLarge Language ModelPrompt EngineeringDiffusion modelWorld ModelImageTextMultimodalityGraphTabularRetrieval-Augmented Generation
🎯 What it does: A unified graph world model (GWM) is proposed to handle multimodal structured and unstructured data, achieving prediction, generation, and planning across six domain tasks.
Graph-Assisted Stitching for Offline Hierarchical Reinforcement Learning
Seungho Baek (Sungkyunkwan University), Yusung Kim (Sungkyunkwan University)
Graph Neural NetworkReinforcement LearningGraph
🎯 What it does: A graph-structured offline hierarchical reinforcement learning framework (Graph-Assisted Stitching, GAS) has been designed and implemented, which generates sub-goals using graph search instead of high-level policies, and achieves long-term task completion through trajectory stitching.
Graph-Based Algorithms for Diverse Similarity Search
Piyush Anand (Microsoft), Haike Xu (MIT)
RetrievalOptimizationGraph Neural NetworkGraph
🎯 What it does: This paper proposes a graph-based approximate nearest neighbor search algorithm that achieves efficient querying while ensuring result diversity.
Graph-constrained Reasoning: Faithful Reasoning on Knowledge Graphs with Large Language Models
Linhao Luo (Monash University), Shirui Pan (Griffith University)
Graph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningTextGraphBenchmark
🎯 What it does: This paper proposes a Graph-Constrained Reasoning (GCR) framework, which encodes knowledge graphs into prefix trees using KG-Trie and constrains generation during the LLM decoding process, enabling LLMs to perform reliable reasoning on graphs and eliminate hallucinations.
Graph-Supported Dynamic Algorithm Configuration for Multi-Objective Combinatorial Optimization
Robbert Reijnen (Eindhoven University of Technology), Yingqian Zhang (Eindhoven University of Technology)
OptimizationGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: A dynamic algorithm configuration framework GS-MODAC based on Graph Neural Networks (GNN) and Deep Reinforcement Learning (DRL) is proposed to adjust the parameters of evolutionary algorithms (such as NSGA-II, MOPSO) in real-time for Multi-Objective Combinatorial Optimization (MOCO).
Graph4MM: Weaving Multimodal Learning with Structural Information
Xuying Ning (University of Illinois Urbana Champaign), Jingrui He (University of Illinois Urbana Champaign)
ClassificationGenerationGraph Neural NetworkTransformerLarge Language ModelVision Language ModelMultimodalityGraph
🎯 What it does: This paper proposes the Graph4MM framework, which utilizes a multimodal graph structure to integrate multi-hop neighbor information and cross-modal representations in the base model.
GraphCL: Graph-based Clustering for Semi-Supervised Medical Image Segmentation
Mengzhu Wang (Hebei University of Technology), Jingcai Guo (Hong Kong Polytechnic University)
SegmentationGraph Neural NetworkImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: GraphCL is proposed, a clustering method based on graph neural networks, designed to utilize both labeled and unlabeled data in semi-supervised medical image segmentation; it models structural information and clustering by constructing instance graphs, performing structure-aware alignment, and implementing a non-k clustering loss; it also combines bidirectional Copy-Paste augmentation within a teacher-student framework.
GraphGPT: Generative Pre-trained Graph Eulerian Transformer
Qifang Zhao (Alibaba Inc), Xiaoxiao Xu (Alibaba Inc)
GenerationRepresentation LearningGraph Neural NetworkTransformerGraphBenchmark
🎯 What it does: This paper presents GraphGPT—a self-supervised generative pre-trained graph model based on Graph Euler Transformer (GET), which can directly input any graph into a standard Transformer for learning after serializing it through an Euler path.
Gravity-Bench-v1: A Benchmark on Gravitational Physics Discovery for Agents
Nolan Koblischke (University of Toronto), Mohamad Ali-Dib (New York University)
OptimizationRobotic IntelligenceReinforcement Learning from Human FeedbackAgentic AITabularBenchmarkPhysics Related
🎯 What it does: This paper presents Gravity-Bench-v1, an AI scientific discovery benchmark based on a physical simulation environment, designed to evaluate agents' capabilities in observation planning, data analysis, and reasoning in two-body gravitational dynamics experiments.
Great Models Think Alike and this Undermines AI Oversight
Shashwat Goel, Jonas Geiping
Supervised Fine-TuningText
🎯 What it does: A CAPA quantitative model for functional similarity is proposed, and this metric is used to study biases in AI supervision, weak-to-strong training gains, and error correlations caused by capability improvements.
Gridded Transformer Neural Processes for Spatio-Temporal Data
Matthew Ashman (University of Cambridge), Richard E. Turner (University of Cambridge)
TransformerTime Series
🎯 What it does: This paper proposes 'gridded TNP'—a neural process framework that maps unstructured spatiotemporal data to a structured pseudo-label grid, and utilizes efficient Transformers (ViT, Swin) for encoding, processing, and decoding on this grid.
Griffin: Towards a Graph-Centric Relational Database Foundation Model
Yanbo Wang (Institute for Artificial Intelligence Peking University), Muhan Zhang (Institute for Artificial Intelligence Peking University)
ClassificationGraph Neural NetworkSupervised Fine-TuningGraphTabularTime Series
🎯 What it does: Griffin is a foundational model for relational databases that unifies encoders and decoders, capable of handling various prediction tasks such as classification and regression within the same framework.
GrokFormer: Graph Fourier Kolmogorov-Arnold Transformers
GuoguoAi, Hui Yan (Nanjing University of Science and Technology)
ClassificationGraph Neural NetworkTransformerGraph
🎯 What it does: The GrokFormer graph Transformer model is proposed, which integrates self-attention with learnable graph Fourier Kolmogorov–Arnold networks (KAN) filters to achieve adaptive modeling of graph frequency.
Grokking at the Edge of Linear Separability
Alon Beck (Tel Aviv University), Yohai Bar-Sinai
ClassificationOptimizationTabular
🎯 What it does: In a high-dimensional binary logistic regression task, the authors investigated the phenomenon of 'grokking' through theoretical and numerical experiments, which refers to the non-monotonic and delayed generalization behavior of training and testing losses.
Grokking Beyond the Euclidean Norm of Model Parameters
Tikeng Notsawo Pascal Junior, Guillaume Rabusseau (Universite de Montreal)
Image
🎯 What it does: This paper studies the phenomenon of delayed generalization—grokking—occurring in the gradient descent training of neural networks, and proves that, in addition to traditional ℓ2 regularization, sparse (ℓ1) or low-rank (nuclear norm) and even domain-specific regularization can induce this phenomenon.
Grokking in the Wild: Data Augmentation for Real-World Multi-Hop Reasoning with Transformers
Roman Abramov (Technical University of Munich), Gjergji Kasneci (Technical University of Munich)
TransformerLarge Language ModelTextGraph
🎯 What it does: By performing data augmentation in real-world multi-hop reasoning tasks, implicit reasoning is achieved using Transformers.
GRU: Mitigating the Trade-off between Unlearning and Retention for LLMs
Yue Wang (Hong Kong Baptist University), Bo Han (Hong Kong Baptist University)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper proposes the Gradient Rectified Unlearning (GRU) framework, which mitigates the negative impact on non-target data performance when eliminating target data by projecting and correcting the unlearning gradient during LLM training. Additionally, it introduces the Task Vector Rectified Unlearning (TRU) scheme for scenarios without retained data.
GS-Bias: Global-Spatial Bias Learner for Single-Image Test-Time Adaptation of Vision-Language Models
Zhaohong Huang (Xiamen University), Rongrong Ji (Xiamen University)
Domain AdaptationTransformerVision Language ModelContrastive LearningImage
🎯 What it does: A testing-time adaptive method named GS-Bias is proposed, which directly adjusts the output logits of the CLIP model by learning global and spatial biases during testing on a single image, in order to enhance zero-shot generalization performance.
GSM-$\infty$: How Do your LLMs Behave over Infinitely Increasing Reasoning Complexity and Context Length?
Yang Zhou (Carnegie Mellon University), Beidi Chen (Carnegie Mellon University)
Large Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes an infinitely expandable long-context reasoning benchmark, GSM-∞, which automatically generates elementary arithmetic problems of arbitrary complexity and context length through a computation graph generator, thereby evaluating the reasoning capabilities of LLMs.
GTR: A General, Multi-View, and Dynamic Framework for Trajectory Representation Learning
Xiangheng Wang (Zhejiang University), Yunjun Gao (Zhejiang University)
Representation LearningGraph Neural NetworkTransformerMixture of ExpertsTime SeriesSequential
🎯 What it does: Proposes the GTR framework, which combines a multi-view encoder (road network view + free space grid view) with position and time embeddings to learn a global multi-dimensional representation of trajectories;
Guarantees of a Preconditioned Subgradient Algorithm for Overparameterized Asymmetric Low-rank Matrix Recovery
Paris Giampouras (University of Warwick), Rene Vidal
Anomaly DetectionOptimizationTabular
🎯 What it does: A over-parameterized preconditioned subgradient algorithm (OPSA) is proposed for robust low-rank matrix recovery in the presence of outliers, unknown rank, and asymmetric matrices.
GuardAgent: Safeguard LLM Agents via Knowledge-Enabled Reasoning
Zhen Xiang (University of Georgia), Bo Li (University of Chicago)
Safty and PrivacyAI Code AssistantTransformerLarge Language ModelAgentic AITextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes GuardAgent, a knowledge-driven reasoning LLM monitoring agent that generates executable safety code for dynamically checking and limiting the behavior of other LLM agents to meet user safety requests.
Guardians of Image Quality: Benchmarking Defenses Against Adversarial Attacks on Image Quality Metrics
Aleksandr Gushchin (Moscow State University), Anastasia Antsiferova (Moscow State University)
CompressionAdversarial AttackImageBenchmark
🎯 What it does: This paper conducts a systematic evaluation of defense methods for no-reference image quality assessment (NR-IQA) metrics based on neural networks under adversarial attacks and establishes the first comprehensive benchmark.
Guided Search Strategies in Non-Serializable Environments with Applications to Software Engineering Agents
Karina Zainullina (Nebius), Boris Yangel (Nebius)
TransformerLarge Language ModelReinforcement LearningAgentic AITextBenchmark
🎯 What it does: In non-serializable RL environments (such as Docker containers), two guided search strategies, one-step lookahead and trajectory selection, are used to improve the success rate of large language models on SWE-bench Verified.
Guided Structural Inference: Leveraging Priors with Soft Gating Mechanisms
Aoran Wang (University of Luxembourg), Jun Pang (University of Luxembourg)
Graph Neural NetworkAuto EncoderGraph
🎯 What it does: This paper proposes a Soft-Gated Structural Inference (SGSI) framework that seamlessly integrates partial prior knowledge (known edges, sparsity, node degree constraints) into the graph structure learning of VAE through three main mechanisms: soft gating, clone-fix, and adaptive regularization.
Guided Zeroth-Order Methods for Stochastic Non-convex Problems with Decision-Dependent Distributions
Yuya Hikima (NTT Corporation), Akinori Fujino (NTT Corporation)
OptimizationTabular
🎯 What it does: Two 'guided' optimization algorithms based on zeroth-order methods are proposed, specifically designed to address non-convex stochastic optimization problems under decision-related unknown distributions.
GuidedQuant: Large Language Model Quantization via Exploiting End Loss Guidance
Jinuk Kim (Seoul National University), Hyun Oh Song (Seoul National University)
TransformerLarge Language ModelText
🎯 What it does: A post-training quantization method called GuidedQuant is proposed, which utilizes the final loss gradient to guide quantization while maintaining weight correlation within output channels. An efficient non-uniform scalar quantization algorithm, LNQ, is also introduced.
Gumiho: A Hybrid Architecture to Prioritize Early Tokens in Speculative Decoding
Jinze Li (Advanced Micro Devices), Emad Barsoum (Advanced Micro Devices)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: A hybrid model called Gumiho, which combines sequence Transformer and parallel MLP, is proposed to enhance the efficiency of Speculative Decoding during inference.
H-Tuning: Toward Low-Cost and Efficient ECG-based Cardiovascular Disease Detection with Pre-Trained Models
Rushuang Zhou (City University of Hong Kong), Yining Dong (City University of Hong Kong)
ClassificationComputational EfficiencyKnowledge DistillationSupervised Fine-TuningTime SeriesBiomedical DataElectrocardiogram
🎯 What it does: The H-Tuning framework is proposed, which first uses low-cost mix-order optimization and LoRA low-rank adaptation for efficient fine-tuning of large pre-trained models, and then compresses the fine-tuned teacher model into a very small student model through knowledge distillation, for multi-label cardiovascular disease (CVD) detection on mobile ECG.
Habitizing Diffusion Planning for Efficient and Effective Decision Making
Haofei Lu (Tsinghua University), Dongqi Han (Microsoft Research Asia)
OptimizationComputational EfficiencyReinforcement LearningDiffusion modelAuto EncoderTabularBenchmark
🎯 What it does: This paper proposes the Habi framework, which transforms slow, precise diffusion planning models into fast, habitual decision-making strategies.
HALoS: Hierarchical Asynchronous Local SGD over Slow Networks for Geo-Distributed Large Language Model Training
Geon-Woo Kim (University of Texas at Austin), Aditya Akella (University of Texas at Austin)
Large Language ModelText
🎯 What it does: A hierarchical asynchronous local SGD framework named HALoS is proposed to address communication bottlenecks and hardware heterogeneity in geographically distributed LLM training.
Handling Imbalanced Pseudolabels for Vision-Language Models with Concept Alignment and Confusion-Aware Calibrated Margin
Yuchen Wang (Harbin Institute of Technology), Xinyang Chen (Harbin Institute of Technology)
ClassificationRecognitionTransformerLarge Language ModelPrompt EngineeringVision Language ModelImage
🎯 What it does: In unsupervised, semi-supervised, and transductive zero-shot learning tasks for VLM, a solution to the pseudo-label imbalance problem is proposed through concept alignment and confusion-based calibrated margins.
HaploVL: A Single-Transformer Baseline for Multi-Modal Understanding
Rui Yang (University of Hong Kong), Hengshuang Zhao (University of Hong Kong)
Knowledge DistillationRepresentation LearningTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: A single Transformer early fusion multimodal model called HaploVL has been designed and implemented, along with a staged training scheme based on knowledge distillation.
Hardware and Software Platform Inference
Cheng Zhang (Imperial College London), Ilia Shumailov (Google DeepMind)
RecognitionOptimizationConvolutional Neural NetworkTransformerLarge Language ModelImage
🎯 What it does: This paper proposes a Hardware and Software Platform Inference (HSPI) method that can identify the GPU architecture and software stack of a black-box model solely based on its input-output behavior.
HarmoniCa: Harmonizing Training and Inference for Better Feature Caching in Diffusion Transformer Acceleration
Yushi Huang (Hong Kong University of Science and Technology), Jun Zhang (Hong Kong University of Science and Technology)
GenerationComputational EfficiencyTransformerDiffusion modelImage
🎯 What it does: The HarmoniCa framework is proposed to accelerate the generation process of the Diffusion Transformer through feature caching.