ICML 2026 Papers — Page 24
International Conference on Machine Learning · 6554 papers
Gated Relational Alignment via Confidence-based Distillation for Efficient VLMs
Yanlong Chen (ETH Zurich), Yawei Li (ETH Zurich)
Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerVision Language ModelImageTextMultimodality
🎯 What it does: A unified GRACE framework is designed, combining knowledge distillation with quantization-aware training (QAT) to achieve efficient INT4 inference for Vision-Language models.
Gateways to Tractability for Satisfiability in Pearl’s Causal Hierarchy
Robert Ganian (TU Wien), Simon Wietheger (TU Wien)
OptimizationExplainability and InterpretabilityComputational EfficiencyReinforcement LearningPrompt EngineeringContrastive LearningReview/Survey Paper
🎯 What it does: This paper explores the parameterized complexity of the satisfiability problem in the Pearl Causal Hierarchy (PCH), and for the first time identifies the decidability threshold that can reach various subproblems of satisfiability;
Gauge-Equivariant Graph Networks via Self-Interference Cancellation
Yoonhyuk Choi (Sookmyung Women's University), Jiwoo Kang (Sookmyung Women's University)
Representation LearningGraph Neural NetworkContrastive LearningGraphBenchmark
🎯 What it does: Propose a Gauge-Equivariant Graph Network (GESC) based on complex representation, achieving efficient message passing on heterogeneous graphs through self-interference cancellation and phase-aware gating.
GAUSS: Graph-Assisted Uncertainty Quantification using Structure and Semantics for Long-Form Generation in LLMs
Karthik Somayaji NS (University of California, Santa Barbara), Peng Li (University of California, Santa Barbara)
GenerationExplainability and InterpretabilityGraph Neural NetworkTransformerLarge Language ModelContrastive LearningTextGraph
🎯 What it does: Propose the GAUSS framework, which splits long texts generated by LLMs into atomic facts and constructs semantic graphs, quantifying the uncertainty of generation through graph alignment.
Gaussian Mean Field Variational Inference can Overestimate Predictive Variance
James Odgers (University of Technology Nuremberg), Vincent Fortuin (University of Technology Nuremberg)
OptimizationExplainability and InterpretabilityComputational EfficiencyRepresentation LearningDiffusion modelScore-based ModelFlow-based ModelRectified FlowAuto EncoderGenerative Adversarial NetworkContrastive LearningGaussian SplattingTabularTime SeriesSequentialBenchmark
🎯 What it does: This paper explores the impact of mean-field variational inference (MFVI) on predictive variance in Bayesian linear regression through theoretical analysis and experimental validation, and reveals that it may produce excessive uncertainty under the training distribution.
GaussTrace: Provenance Analysis of 3D Gaussian Splatting Models with Evidence-based LLM Reasoning
Haoliang Han (Hong Kong Baptist University), Renjie Wan (Hong Kong Baptist University)
Explainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelDiffusion modelScore-based ModelContrastive LearningGaussian SplattingImagePoint CloudMeshChain-of-Thought
🎯 What it does: Construct a framework called GaussTrace based on evidence reasoning to trace the evolutionary relationships of 3D Gaussian Splatting (3DGS) models and generate interpretable directed Provenance graphs.
GCIB: Graph Contrastive Information Bottleneck for Multi-Behavior Recommendation
Likang Wu (Tianjin University), Lei Zhang (Anhui University)
Recommendation SystemGraph Neural NetworkContrastive LearningGraph
🎯 What it does: Propose a multi-behavior recommendation framework GCIB, which denoises auxiliary behavior graphs using graph information bottleneck and enhances target behavior representations through cross-behavior contrastive learning;
GDPO: Group reward-Decoupled Normalization Policy Optimization for Multi-reward RL Optimization
Shih-Yang Liu (NVIDIA), Pavlo Molchanov (NVIDIA)
OptimizationReinforcement LearningText
🎯 What it does: This paper investigates the limitations of the traditional GRPO method in multi-reward reinforcement learning, and proposes GDPO by separating reward normalization to avoid the reward folding problem, thereby improving training stability and performance.
Gecko: A Simulation Environment with Stateful Feedback for Refining Agent Tool Calls
Zeyu Zhang (Australian National University), Liang Zheng (Australian National University)
Autonomous DrivingRobotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose Gecko, a stateful simulation environment, to verify, simulate responses, and provide task-level feedback for tool calls generated by LLMs before actual tool execution, thereby achieving safe and efficient tool usage;
GEM-FI: Gated Evidential Mixtures with Fisher Modulation
Marco Mustafa Mohammed (University of Kurdistan), Pietro Lio
ClassificationAnomaly DetectionConvolutional Neural NetworkTransformerMixture of ExpertsContrastive LearningImage
🎯 What it does: Propose a single forward inference evidence deep learning framework called GEM-FI, which utilizes learned energy signals for gating, combined with a lightweight multi-head mixture and Fisher information regularization to enhance the model's distance-awareness and multi-modal uncertainty expression.
GEM: Geometric Entropy Mixing for Optimal LLM Data Curation
Yue Min (Wizard Quant), Yujun Li (Wizard Quant)
OptimizationKnowledge DistillationRepresentation LearningData-Centric LearningTransformerLarge Language ModelMixture of ExpertsContrastive LearningTextRetrieval-Augmented Generation
🎯 What it does: Propose the GEM framework, which transforms LLM data clustering into a variational optimization problem on the unit hypersphere, and achieves provable convergence through the MM algorithm;
GEM: Geometric Erasure by Contrastive Velocity Matching in Rectified Flows
Jonas Henry Grebe (Technical University of Darmstadt), Marcus Rohrbach (Technical University of Darmstadt)
GenerationSafty and PrivacyTransformerSupervised Fine-TuningDiffusion modelRectified FlowContrastive LearningImageTextMultimodality
🎯 What it does: Propose the GEM method for rectified flow text-to-image generation models (such as FLUX, SD3), utilizing teacher-guided velocity matching and geometric contrast loss to efficiently erase undesirable concepts (such as nudity, gore, copyrighted characters, etc.).
GemDepth: Geometry-Embedded Features for 3D-Consistent Video Depth
Yuecheng Liu (Huazhong University Of Science And Technology), Xin Yang (Huazhong University Of Science And Technology)
Pose EstimationDepth EstimationConvolutional Neural NetworkTransformerDiffusion modelScore-based ModelAuto EncoderGenerative Adversarial NetworkContrastive LearningOptical FlowImageVideoPoint Cloud
🎯 What it does: Propose the GemDepth framework, which utilizes a geometric embedding module (GEM) and an alternating spatiotemporal Transformer (ASTT) to achieve 3D-consistent video depth estimation;
GEMQ: Global Expert-Level Mixed-Precision Quantization for MoE LLMs
Jianing Deng (University of Pittsburgh), Jingtong Hu (University of Pittsburgh)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringMixture of ExpertsText
🎯 What it does: This paper proposes a global expert-level mixed-precision quantization method called GEMQ, which systematically addresses the expert bit-width allocation and routing offset issues in MoE LLMs;
GenAlign: Towards Unified Alignment Framework of MLLMs via Generative Reward Model
Jingyu Zhang (Ant Group), shiwen cui
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringVision Language ModelGenerative Adversarial NetworkContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: Propose the GenAlign framework, which combines generative reward models (GRM) with multimodal large language models (MLLM) alignment, achieving reasoning-based preference judgment based on adaptive rubric.
GenCircuit-RL: Reinforcement Learning from Hierarchical Verification for Genetic Circuit Design
Noah Flynn (University of California, Berkeley)
Drug DiscoveryAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBiomedical Data
🎯 What it does: Propose the GenCircuit-RL framework, which uses reinforcement learning combined with hierarchical verification rewards to train language models to generate gene circuit code that complies with the SBOL standard;
GenDA: Generative Data Assimilation on Complex Urban Areas via Classifier-Free Diffusion Guidance
Francisco Giral (Universidad Politécnica de Madrid), Soledad Le Clainche (Universidad Politécnica de Madrid)
RestorationGenerationData SynthesisOptimizationGraph Neural NetworkDiffusion modelScore-based ModelGenerative Adversarial NetworkContrastive LearningMeshGraphTabularTime SeriesPhysics Related
🎯 What it does: Propose a generative data assimilation framework called GenDA based on graph neural networks, which uses multi-scale graph diffusion models and classifier-free guidance (CFG) to recover high-resolution wind fields in complex cities from sparse sensor observations;
GENEB: Why Genomic Models Are Hard to Compare
Daria Ledneva (Moscow Independent Research Institute of Artificial Intelligence), Denis Kuznetsov (Moscow Independent Research Institute of Artificial Intelligence)
Representation LearningTransformerSupervised Fine-TuningContrastive LearningBiomedical DataBenchmark
🎯 What it does: Propose the GENEB benchmark to uniformly evaluate the performance of 40 genomic foundation models across 100 tasks (covering 13 functional categories);
General Analysis of LMO-based Optimizers: Beyond Bounded Variance
Egor Shulgin (King Abdullah University of Science and Technology), Eduard Gorbunov (Mohammed Bin Zayed University of Artificial Intelligence)
OptimizationTextTabular
🎯 What it does: This paper proposes a unified convergence theory for analyzing linear minimization operator (LMO) optimizers with momentum (such as normalized SGDM, momentum-based SignSGD, and Muon). The theory is based on the ABC (expected smoothness) assumption, overcoming the limitations of the traditional uniform variance model in subsampling scenarios.
General and Efficient Steering of Unconditional Diffusion Models
Qingsong Wang (University of California San Diego), Yusu Wang (University of California San Diego)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: Propose NA-RFM, a post-diffusion gradient-free control method, which achieves conditional guidance of unconditional diffusion models during sampling by using offline PCA alignment and recursive feature machine (RFM) directions.
General Covariant Action Modeling: Constructing Generalized Manifolds via Spatio-Temporal Decoupling
Huaihai Lyu (Chinese Academy of Sciences), Changsheng Xu (Chinese Academy of Sciences)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerVision-Language-Action ModelFlow-based ModelRectified FlowAuto EncoderVideoSequential
🎯 What it does: Proposed a general covariant action modeling framework called GAM based on spatiotemporal decoupling, constructing an action manifold.
General Quantification of Covariate and Concept Shifts
Hongbo Chen (South China University of Technology), Li C Xia
Domain AdaptationData-Centric LearningImageTabular
🎯 What it does: Propose the γ* concept shift definition and jointly provide a general error bound with X shift, while offering estimable shift magnitudes and the DataShifts algorithm;
General Synthetic-Powered Inference
Meshi Bashari (Technion Iit), Yaniv Romano (Technion Iit)
Data SynthesisFederated LearningSafty and PrivacyComputational EfficiencyRepresentation LearningProtein Structure PredictionLarge Language ModelImageTextTabularBiomedical DataBenchmark
🎯 What it does: Developed a general framework called GESPI, which can securely utilize synthetic data in statistical inference while ensuring error rate control.
Generalist Graph Anomaly Detection via Prototype-Based Distillation
Yiming Xu (Xi'an Jiaotong University), Chao Shen (Xi'an Jiaotong University)
Anomaly DetectionComputational EfficiencyKnowledge DistillationRepresentation LearningGraph Neural NetworkContrastive LearningGraph
🎯 What it does: Proposed ProMoS, an unsupervised general graph anomaly detection framework that can achieve zero-shot detection on unseen graphs
Generalizable and Actionable Parts Pose Estimation with Symmetry Annotation-Free Learning Strategy
wenxiao chen, Dan Guo (Hefei University of Technology)
Pose EstimationGraph Neural NetworkAuto EncoderContrastive LearningPoint CloudGraph
🎯 What it does: Propose the SAFAG framework to achieve cross-category 6D pose estimation of parts without relying on symmetric annotations, and realize multi-hypothesis reasoning through two-stage candidate-to-final quaternion regression and adaptive symmetric learning.
Generalizable and Composable Multi-Model Embedding Translation
Beining Yang (University of Edinburgh), Yang Cao (University of Edinburgh)
RetrievalDomain AdaptationRepresentation LearningTransformerPrompt EngineeringMixture of ExpertsContrastive LearningTextMultimodality
🎯 What it does: Study the cross-model embedding translation problem, propose a translation method based on geometric error bounds and confidence metrics, and design a hierarchical Mixture-of-Experts (H-MoE) architecture, enabling embeddings to maintain high robustness in OOD, multi-model hybrid, and chained translation scenarios.
Generalization and Scaling Laws for Mixture-of-ExpertsTransformers
Mansour MAYAKI (Universite Lumiere Lyon 2)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: This paper proposes a theoretical framework for Mixture-of-Experts Transformer, deriving the approximation, generalization, and neural network scaling laws, and verifying the theoretical predictions through experiments.
Generalization Bounds for Discrete Diffusion: Statistical Advantage of Masking
Zixuan Zhang (Georgia Institute of Technology), Minshuo Chen (Northwestern University)
GenerationData SynthesisTransformerDiffusion modelScore-based ModelSequential
🎯 What it does: This paper studies the generalization properties of discrete diffusion models from a statistical learning perspective, deriving the impact of different forward corruption kernels (masking/absorption vs. uniform replacement) on distribution estimation error and sample complexity.
Generalization Bounds for Out-of-distribution Generalization
Xin Zou (Wuhan University), Weiwei Liu (Fudan University)
Domain AdaptationOptimizationFederated LearningExplainability and InterpretabilitySupervised Fine-TuningContrastive Learning
🎯 What it does: This paper provides lower and upper bounds for out-of-distribution (OOD) generalization under multi-source distributions within the framework of statistical learning theory, and proves that simple mean empirical risk minimization (ERM) nearly achieves the optimal target domain overfitting rate when the RCS and empirical RCS conditions are satisfied.
Generalization of Gibbs and Langevin Monte Carlo Algorithms in the Interpolation Regime
Andreas Maurer (Istituto Italiano di Tecnologia), Massimiliano Pontil (Istituto Italiano di Tecnologia)
OptimizationExplainability and InterpretabilityComputational EfficiencyImageStochastic Differential Equation
🎯 What it does: Studies the generalization performance of Gibbs and Langevin Monte Carlo in the overparameterized interpolation regime, and provides an empirical error upper bound based on training data;
Generalized Boundary FDR Control under Arbitrary Dependence: An Approach on Closure Principle
Yifan Zhang (Shanghai Jiao Tong University), Haojie Ren (Shanghai Jiao Tong University)
Drug DiscoveryTabularBiomedical DataFinance Related
🎯 What it does: Propose the k-bFDR metric and design the Domino algorithm based on the closure principle to achieve boundary FDR control for p-values and e-values under any dependency structure;
Generalized Correctness Models: Learning Calibrated and Cross-Model Correctness Predictors from Historical Patterns
Hanqi Xiao (University of North Carolina Chapel Hill), Mohit Bansal (University of North Carolina Chapel Hill)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Studied a cross-model confidence calibration method based on historical prediction data, called Generalized Correctness Models (GCM), used to predict the correctness of answers from multiple large language models (LLMs).
Generalized Discrete Diffusion with Self-Correction
Linxuan Wang (Purdue University), Qifan Song (Purdue University)
GenerationData SynthesisTransformerLarge Language ModelDiffusion modelContrastive LearningText
🎯 What it does: Proposed a self-correcting discrete diffusion model (SCDD), which eliminates the residual masking step by explicitly using state transitions and a forward process with only uniform transitions, significantly improving the quality and efficiency of parallel generation.
Generalized Linear Bandits with Memory
Heesang Ann (Seoul National University), Min-hwan Oh (Seoul National University)
Reinforcement Learning
🎯 What it does: Studied generalized linear bandits with memory, and proposed a block algorithm based on shrinking confidence intervals, improving the regret bounds of linear models and extending them to generalized linear models.
Generalized Schrödinger Bridge on Graphs
Panagiotis Theodoropoulos (Georgia Institute of Technology), Jaemoo Choi (Georgia Institute of Technology)
OptimizationComputational EfficiencyGraph Neural NetworkReinforcement LearningScore-based ModelContrastive LearningGraphTabularTime SeriesStochastic Differential Equation
🎯 What it does: Proposes a general Schr"odinger bridge framework (GSBoG) for learning continuous-time Markov chain (CTMC) control policies on arbitrary graphs, achieving probabilistic mass transport from source to target distributions, and optimizing state-dependent operational costs during the transport process.
Generalizing Multi-Scale Time-Series Modeling with a Single Operator
Cheonwoo Lee (Korea Advanced Institute of Science and Technology), Jaemin Yoo (Seoul National University)
Computational EfficiencyRepresentation LearningTransformerAuto EncoderContrastive LearningGaussian SplattingTime Series
🎯 What it does: Proposes a theoretical framework for the unified scale operation family, and based on this, designs a learnable discrete Gaussian kernel (SIGMA), achieving multi-scale and distance-aware modeling of time series.
Generalizing Stochastic Smoothing for Differentiation and Gradient Estimation
Felix Petersen (DiffLogic Inc), Stefano Ermon (Stanford University)
OptimizationComputational EfficiencyRepresentation LearningData-Centric LearningDiffusion modelScore-based ModelAuto EncoderContrastive LearningImagePoint CloudTabularBiomedical DataBenchmarkPhysics RelatedStochastic Differential Equation
🎯 What it does: A general stochastic smoothing framework is proposed for non-differentiable black-box functions, enabling gradient estimation and differentiability.
Generation Enhances Understanding in Unified Multimodal Models via Multi-Representation Generation
Zihan Su (Tsinghua University), Xiangxiang Chu (Alibaba Group)
SegmentationGenerationDepth EstimationRepresentation LearningTransformerSupervised Fine-TuningVision Language ModelDiffusion modelAuto EncoderContrastive LearningImageTextMultimodality
🎯 What it does: Propose the post-training method UniMRG, enabling the unified multimodal model (UMM) to learn various intrinsic visual representations (pixel reconstruction, depth, segmentation) in generation tasks, thereby enhancing its visual understanding capability.
Generative Adaptation of Dynamics to Environmental Shifts via Weight-space Diffusion
Ruikun Li (Tsinghua University), Yong Li (Tsinghua University)
GenerationData SynthesisDomain AdaptationComputational EfficiencyMeta LearningGraph Neural NetworkTransformerPrompt EngineeringMixture of ExpertsDiffusion modelAuto EncoderTime SeriesSequentialPhysics Related
🎯 What it does: Propose DynaDiff, which directly generates environment-specific dynamic prediction models on observed short sequences using weight space diffusion models, achieving rapid adaptation with zero gradient fine-tuning.
Generative Augmented Inference
Cheng Lu (Washington University in St. Louis), Heng Zhang (Arizona State University)
Explainability and InterpretabilityRepresentation LearningData-Centric LearningReinforcement Learning from Human FeedbackLarge Language ModelContrastive LearningTextTabularElectronic Health Records
🎯 What it does: Proposed a new statistical inference framework called Generative Augmented Inference (GAI), which treats the labels generated by large language models as auxiliary features rather than proxies for target variables;
Generative Inverse Design with Abstention via Diagonal Flow Matching
Miguel de Campos (Technische Universitaet Berlin), Hanno Gottschalk (Technische Universitaet Berlin)
GenerationData SynthesisOptimizationScore-based ModelFlow-based ModelGenerative Adversarial NetworkImagePoint CloudMeshGraphTabularOrdinary Differential Equation
🎯 What it does: A new reversible conditional flow matching model called Diagonal Flow Matching (Diag-CFM) is proposed and implemented in the inverse design task, which can efficiently generate diverse designs that meet target performance between design parameters and performance labels, and provide forward prediction functionality.
Generative Large Neighborhood Search: Scalable Set Cover Optimization via Discrete Diffusion
Achref Jaziri (Goethe University), Bruno Adrien De Backer
OptimizationGraph Neural NetworkTransformerDiffusion modelTabularBenchmark
🎯 What it does: Propose Generative Large Neighborhood Search (GLNS), which uses the predicted trajectory from a discrete diffusion model to guide neighborhood selection in large set cover problems, and combines local decomposition to achieve scalable repair processes.
Generative Modeling of Discrete Latent Structures via Dynamic Policy Gradients
Stefan Ivanovic (University of Illinois at Urbana-Champaign), Mohammed El-Kebir (University of Illinois at Urbana-Champaign)
GenerationData SynthesisReinforcement LearningDiffusion modelScore-based ModelGraphTabularBiomedical Data
🎯 What it does: Propose the GReinSS framework, which directly maximizes the likelihood of observed data using a policy gradient with dynamic reward strategy, thereby inferring discrete latent structures.
Generative Modeling of Irregular Time Series via SDE-Induced Continuous-Discrete Variational Inference
Zexin Yuan (Sun Yat-sen University), Junxi Xiao (Sun Yat-sen University)
GenerationData SynthesisAnomaly DetectionComputational EfficiencyTransformerFlow-based ModelTabularTime SeriesBiomedical DataElectronic Health RecordsBenchmarkStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Proposes the SDE-VI framework, which uses SDE-induced continuous-discrete variational inference for generative modeling of irregular time series.
Generative Neural Operators through Diffusion Last Layer
Sungwon Park (Korea University), Amir Barati Farimani (Carnegie Mellon University)
GenerationData SynthesisComputational EfficiencyConvolutional Neural NetworkGraph Neural NetworkTransformerDiffusion modelScore-based ModelFlow-based ModelNeural Radiance FieldAuto EncoderTabularTime SeriesBenchmarkPhysics RelatedStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Propose a lightweight probabilistic output head called Diffusion Last Layer (DLL), which can convert any neural operator (such as FNO, DeepONet) into a conditional generative model, enabling sampling from the function space distribution and uncertainty quantification.
Generative Online Reinforcement Learning
Chubin Zhang (Nanyang Technological University), Bo An (Nanyang Technological University)
TransformerSupervised Fine-TuningReinforcement LearningDiffusion modelFlow-based ModelMultimodality
🎯 What it does: Propose the GORL framework, which decouples policy optimization from generation in online reinforcement learning, using an optimizable latent policy and a powerful conditional generation decoder to train a multi-modal action distribution from scratch.
Generative Representation Learning on Hyper-relational Knowledge Graphs via Masked Discrete Diffusion
Jaejun Lee (KAIST), Joyce Jiyoung Whang (KAIST)
GenerationData SynthesisRepresentation LearningGraph Neural NetworkTransformerDiffusion modelGenerative Adversarial NetworkContrastive LearningGraph
🎯 What it does: A unified framework for fact generation and link prediction on hyper-relational knowledge graphs;
Generative Visual Code Mobile World Models
Woosung Koh (Trillion Labs), Jamin Shin (Trillion Labs)
GenerationData SynthesisAI Code AssistantTransformerVision Language ModelWorld ModelImageTextMultimodalityChain-of-Thought
🎯 What it does: This paper proposes a visual mobile interface world model called gWorld, which generates renderable web code and directly predicts the next frame of GUI state;
GenExam: A Multidisciplinary Text-to-Image Exam
Zhaokai Wang (Shanghai Jiao Tong University), Gen Luo (Shanghai AI Laboratory)
TransformerLarge Language ModelPrompt EngineeringVision Language ModelDiffusion modelImageTextMultimodalityBenchmark
🎯 What it does: Propose the GenExam benchmark, using an exam-style multi-disciplinary text-to-image task to evaluate models' comprehensive understanding, reasoning, and generation capabilities.
Genome-Factory: A Library for Tuning, Deploying, and Interpreting Genomic Foundation Models
Weimin Wu (Northwestern University), Han Liu (Northwestern University)
Explainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningAuto EncoderBiomedical DataBenchmark
🎯 What it does: Propose GENOME-FACTORY, a unified Python library that integrates six modules: data collection, model loading, tuning, inference, benchmark evaluation, and sparse autoencoder interpretation, supporting end-to-end usage of various genomic foundational models.
GenShield: Unified Detection and Artifact Correction for AI-Generated Images
Zhipei Xu (Peking University), Jian Zhang (Peking University)
RestorationAnomaly DetectionExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsDiffusion modelFlow-based ModelImageMultimodalityChain-of-Thought
🎯 What it does: Propose GenShield, a unified autoregressive framework that can perform interpretable detection on AI-generated images and controllably repair detected defects, forming a closed-loop diagnostic-repair loop.
GenUnfold: Rapidly Predict Protein Mechanical Unfolding Trajectory via a Physics-Guided Diffusion Model
Yiyuan Zhang (University of Minnesota), Murti Salapaka
Protein Structure PredictionTransformerDiffusion modelBiomedical Data
🎯 What it does: Propose GenUnfold, a physics-guided diffusion model for rapidly predicting protein mechanical unfolding (F-x) trajectories;
GeoAlign: Geometric Rollout Curation for Robust LLM Reinforcement Learning
Ting Zhou (Sun Yat Sen University), Daoyuan Chen (Alibaba Group)
Representation LearningData-Centric LearningTransformerLarge Language ModelReinforcement LearningPrompt EngineeringContrastive LearningText
🎯 What it does: Proposes GEOALIGN, a lightweight episode generation curve screening plugin used to detect and correct directionally inconsistent high-reward episodes in LLM reinforcement learning.
Geodesic Calculus on Implicitly Defined Latent Manifolds
Florine Hartwig (University of Bonn), Benedikt Wirth (University of Münster)
OptimizationRepresentation LearningDiffusion modelScore-based ModelFlow-based ModelRectified FlowAuto EncoderContrastive LearningImagePoint CloudMeshGraphTime SeriesStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This work learns an implicit representation of the potential manifold through approximate projection, constructing a time-discrete Riemannian geometry computational framework, which enables numerical solutions of geometric operations such as geodesics and exponential maps in the latent space of an autoencoder.
Geodesic Flow Matching for Denoising High-Dimensional Structured Representations
Karim Habashy (University of Waterloo), Chris Eliasmith (University of Waterloo)
OptimizationRepresentation LearningSpiking Neural NetworkScore-based ModelFlow-based ModelSimultaneous Localization and MappingOptical FlowPoint CloudTabularTime Series
🎯 What it does: The paper proposes a SSP cleaning method based on geodesic flow matching, used for denoising and structure preservation in high-dimensional structured vector representations.
GeoDM: Geometry-aware Distribution Matching for Dataset Distillation
Xuhui Li (Mohamed bin Zayed University of Artificial Intelligence), zhiqiang xu
Knowledge DistillationRepresentation LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: Propose GeoDM, a data distillation framework that performs distribution matching on the product of three geometric spaces: Euclidean, hyperbolic, and spherical.
GeoEvo: Identity-Aware Potential Game with Geometric Evolution for Personalized Multimodal Federated Learning
Chen Wang (Beijing University of Technology), Baocai Yin (Beijing University of Technology)
OptimizationFederated LearningRepresentation LearningReinforcement LearningDiffusion modelScore-based ModelFlow-based ModelAuto EncoderGenerative Adversarial NetworkContrastive LearningImageTextMultimodalityAudio
🎯 What it does: Propose GeoEvo, a personalized multi-modal federated learning framework that treats missing modalities as client structural identity, based on identity-aware potential games and geometric evolution.
GeoFlow: Geo-Aware Modeling of Inter-Area Relationships in Origin-Destination Flow Prediction and Generation
Zherui Huang (Shanghai Jiao Tong University), Linghe Kong (Shanghai Jiao Tong University)
GenerationData SynthesisGraph Neural NetworkTransformerDiffusion modelScore-based ModelFlow-based ModelGraphTime Series
🎯 What it does: This paper proposes the GeoFlow framework for predicting and generating origin-destination (OD) flows, combining geographic attributes and graph neural networks to achieve more accurate modeling.
GeoLoom: High-quality Geometric Diagram Generation from Textual Input
Xiaojing Wei (Beijing Normal University), Hua Huang (Beijing Normal University)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningImageText
🎯 What it does: This study proposes and implements the GeoLoom framework, which can automatically convert natural language descriptions into high-quality geometric figures;
Geometric and Stochastic Analysis of Discontinuities in Sparse Mixture-of-Experts
Tho Tran Huu (National University of Singapore), Tan Minh Nguyen (National University of Singapore)
Domain AdaptationOptimizationComputational EfficiencyRepresentation LearningTransformerMixture of ExpertsDiffusion modelImageTextMultimodalityStochastic Differential Equation
🎯 What it does: Study the input space discontinuity caused by Top-k routing in Sparse Mixture-of-Experts, and provide geometric and stochastic analysis.
Geometric Coherence Learning for Structuring Value Functions in Plain MDPs
Zuyuan Zhang (George Washington University), Tian Lan (George Washington University)
Reinforcement LearningContrastive LearningImageTabularTime Series
🎯 What it does: Propose GCR-RL, which leverages poset theory to view the learning of the value function as constructing and continuously refining a super poset set, inferring the comparison relationships between state-action pairs through temporal difference (TD) signals in reinforcement learning, and combining symmetry (equivariance) to constrain and accelerate the value function;
Geometric Collapse: When Vision Models Fail to Verify Physical Causality
Wentao Zhang (Macao Polytechnic University), Irwin King (Chinese University of Hong Kong)
Depth EstimationAnomaly DetectionExplainability and InterpretabilityConvolutional Neural NetworkTransformerDiffusion modelFlow-based ModelContrastive LearningOptical FlowImagePoint CloudBenchmark
🎯 What it does: This paper proposes a controlled contrast experiment called Scrambled Edges to test whether dense visual models can verify physical causality during inference, and through this experiment reveals a global failure mode called 'Geometric Collapse'; meanwhile, it analyzes the impact of disrupting different physical priors (continuity, illumination consistency, occlusion causality) on the model.
Geometric Conformal Prediction with Spatial Ranks and Multivariate Quantiles
Anton Conrad (Epita), julien perez
ClassificationOptimizationExplainability and InterpretabilityComputational EfficiencyConvolutional Neural NetworkImageTabularBenchmark
🎯 What it does: Propose a method for constructing multidimensional prediction sets based on geometric quantiles and spatial ranks, providing distribution-free coverage guarantees for split conformal prediction.
Geometric Control of Out-of-Distribution Shift in Safe Offline RL
Zhiqi Zhuang (McGill University), Benoit Boulet (McGill University)
Reinforcement LearningFlow-based ModelTabularSequentialBenchmark
🎯 What it does: Propose the GSCO framework, which enforces geometric constraints in the latent space through conditional invertible normalizing flows in offline safe reinforcement learning, thereby controlling out-of-distribution (OOD) behavior shifts.
Geometric Convergence of Gauss–Newton for Neural Networks: Riemannian Geometry and Adaptive Damping
Semih Cayci (RWTH Aachen University)
OptimizationPoint CloudTabular
🎯 What it does: Studied the convergence behavior of high-order Gauss-Newton (GN) methods in training shallow and deep neural networks, provided non-asymptotic, geometric-level convergence upper bounds, and proved that GN can overcome the convergence bottleneck caused by the condition number of the kernel matrix in both under-parameterized and over-parameterized scenarios.
Geometric Decoupling: Diagnosing the Structural Instability of Latent
Yuanbang Liang (Cardiff University), Yu-Kun Lai (Cardiff University)
GenerationAnomaly DetectionDiffusion modelImage
🎯 What it does: Propose a diagnostic framework based on Riemannian geometry to quantify and locate instabilities in the latent space.
Geometric Embedding Alignment via Curvature Matching in Transfer Learning
Sung Moon Ko (LG AI Research), Sehui Han (LG AI Research)
Drug DiscoveryGraph Neural NetworkDiffusion modelScore-based ModelFlow-based ModelRectified FlowAuto EncoderContrastive LearningGraphTabular
🎯 What it does: Propose a transfer learning framework called GEAR based on Riemannian geometry, which aligns embeddings by matching the Ricci curvature of the potential spaces of the source task and the target task, thereby improving the performance of molecular property regression.
Geometric Entropy and Retrieval Phase Transitions in Continuous Thermal Dense Associative Memory
Tatiana Petrova (University of Luxembourg), Radu State (University of Luxembourg)
RetrievalOptimizationPhysics Related
🎯 What it does: This paper studies the thermodynamic capacity and retrieval phase transition of continuous dense associative memory networks under N-sphere geometric constraints.
Geometric Flow Grounding: A Unified Manifold Decoupling Framework for Dynamics Discovery and Verification
Chang Yu (Westlake University), Stan Z. Li (Westlake University)
Anomaly DetectionOptimizationRepresentation LearningTransformerFlow-based ModelAuto EncoderContrastive LearningOptical FlowImageVideoTabularBiomedical DataBenchmarkStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Propose the Geometric Flow Grounding (GFG) framework, which enforces continuous dynamics to be confined within the learned data manifold's tangent space through neural tangent projection and velocity primitive quantization, addressing the out-of-manifold drift and smoothing drift issues in traditional methods.
Geometric Pocket-Centric Protein Encoding for Polypharmacology-Guided Multi-Target Drug Design
Haoran liu, Xiaolong Zhang (Wuhan University of Science and Technology)
Drug DiscoveryConvolutional Neural NetworkRecurrent Neural NetworkGraph Neural NetworkTransformerPrompt EngineeringDiffusion modelContrastive LearningTextGraphTabularBiomedical DataRetrieval-Augmented Generation
🎯 What it does: A multi-objective drug generation framework based on the geometric topology of protein pockets is proposed, which can directly generate high-affinity molecules that兼顾 multiple targets under the constraints of binding to multiple targets and drug-like properties.
Geometric Rate–Distortion Invariance for Domain Generalization
Tong Liu (State Key Laboratory of Biopharmaceutical Preparation and Delivery, Institute of Process Engineering, Chinese Academy of Sciences), Shuo Bai (State Key Laboratory of Biopharmaceutical Preparation and Delivery, Institute of Process Engineering, Chinese Academy of Sciences)
Domain AdaptationRepresentation LearningContrastive LearningImageBenchmark
🎯 What it does: Studies the domain generalization (DG) problem, proposing the Geometric Rate-Distortion Invariance (RDI) framework, which learns representations robust to distribution drift by aligning class-conditional subspaces on the Grassmann manifold and regularizing spectral capacity.
Geometric Reciprocity: Unlocking Self-Supervision for Stereoscopic Video Generation
Jingyi Lu (University of Hong Kong), Kai Han (University of Hong Kong)
GenerationDepth EstimationConvolutional Neural NetworkTransformerDiffusion modelAuto EncoderContrastive LearningOptical FlowImageVideo
🎯 What it does: Propose a self-supervised stereo reconstruction framework that directly generates disparity masks from monocular videos using the geometric reciprocity theorem (GRT), without requiring stereo matching or synthetic data, achieving high-quality stereo disparity reconstruction.
Geometrically Constrained Outlier Synthesis
Daniil Karzanov (AXA AI Research), Marcin Detyniecki (AXA AI Research)
Data SynthesisAnomaly DetectionDiffusion modelContrastive LearningImageBenchmark
🎯 What it does: Improve the detection capability of deep networks for approximate OOD samples through geometric-constrained synthetic anomalies and contrastive regularization.
Geometrically Constrained Stenosis Editing in Coronary Angiography via Entropic Optimal Transport
Jialin Li (Tianjin University), Jiachen Yang (Tianjin University)
Image TranslationGenerationData SynthesisTransformerDiffusion modelScore-based ModelImageBiomedical DataMagnetic Resonance ImagingComputed TomographyUltrasoundStochastic Differential Equation
🎯 What it does: Propose OT-Bridge Editor, which achieves geometrically accurate stenosis editing in coronary angiography images based on constrained entropy optimal transport and Schrödinger bridge, and is used to generate synthetic training data.
Geometry of Reason: Spectral Signatures of Valid Mathematical Reasoning
Valentin NOËL (Devoteam)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelContrastive LearningTextGraph
🎯 What it does: This paper proposes a spectral graph analysis method based on Transformer attention for untrained validation of mathematical reasoning.
Geometry-Aware Contrastive Learning for Few-Shot Automatic Modulation Recognition
guanqun zhao, Hongwen Yang (Beijing University of Posts and Telecommunications)
RecognitionConvolutional Neural NetworkTransformerContrastive LearningPhysics RelatedAudio
🎯 What it does: This paper proposes a geometry-aware contrastive learning framework called DyCo-CL for automatic modulation recognition in few-shot scenarios;
Geometry-Aware Dataset Condensation for Diffusion Model Training
Xiao Cui (University Of Science And Technology Of China), Houqiang Li (University Of Science And Technology Of China)
GenerationData SynthesisOptimizationTransformerDiffusion modelContrastive LearningImage
🎯 What it does: A geometry-aware dataset compression method is proposed, which trains diffusion models by selecting real samples that are consistent with the geometric distribution of the original data.
Geometry-Aware Decoding with Wasserstein-Regularized Truncation and Mass Penalties for Large Language Models
Arash Gholami Davoodi (Carnegie Mellon University), Pouya Pezeshkpour (Megagon Labs)
GenerationOptimizationComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed a geometry-aware truncated decoding method called TopW based on Wasserstein distance and quality penalty, aiming to balance diversity and coherence while maintaining semantic continuity.
Geometry-Aware Image Flow Matching
Junho Lee (Seoul National University), Joonseok Lee (Seoul National University)
GenerationFlow-based ModelAuto EncoderOptical FlowImage
🎯 What it does: This paper reveals through directional and norm decomposition analysis that the semantic information of natural images is mainly concentrated in the direction, allowing images to be approximated as spheres, and based on this, two spherical flow matching methods, SOT-CFM and SFM, are proposed;
Geometry-Aware Neural Optimizer for Shape Optimization and Inversion
Sun Guoze (Renmin University of China), Hao Sun (Renmin University of China)
OptimizationTransformerDiffusion modelScore-based ModelAuto EncoderPoint CloudMeshPhysics Related
🎯 What it does: Propose a unified framework called GANO for end-to-end differentiable shape optimization and inverse problems, achieving a closed-loop system for geometric representation, field prediction, and automatic optimization;
Geometry-Aware Probabilistic Circuits via Voronoi Tessellations
Sahil Sidheekh (University of Texas at Dallas), Sriraam Natarajan (University of Texas at Dallas)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningDiffusion modelScore-based ModelAuto EncoderContrastive LearningImagePoint CloudMeshTabular
🎯 What it does: This paper embeds Voronoi segmentation into probabilistic circuits (PC), endowing sum nodes with geometry-aware hybrid weights, and proposes two solutions to address the traceability conflicts caused by Voronoi: ① a provably approximate inference framework, ② hierarchical factorized Voronoi (HFV) to restore exact inference, along with gradient learning of softened gates and proof of soft-to-hard convergence.
Geometry-Aware Tabular Diffusion
David Turtora Zagardo (Independent Researcher)
GenerationData SynthesisGraph Neural NetworkTransformerDiffusion modelScore-based ModelTabularBenchmark
🎯 What it does: Propose Geometry-Aware Tabular Diffusion (GATD), incorporating geometric features such as angles and lengths between column pairs as input and auxiliary prediction targets in diffusion-based table data generation.
Geometry-based Schrödinger Bridges for Trustworthy Multimodal Fusion
Jiayu Xiong (Huaqiao University), Jun Xue (Wuhan University)
Anomaly DetectionComputational EfficiencyRepresentation LearningData-Centric LearningConvolutional Neural NetworkTransformerDiffusion modelScore-based ModelFlow-based ModelRectified FlowContrastive LearningImageTextMultimodalityBiomedical DataBenchmark
🎯 What it does: Propose the Geometry-based Multimodal Fusion (GMF) framework, which evaluates the reliability of multimodal data and dynamically weights the fusion by measuring the transmission energy of inputs in the latent space.
Geometry-Correct Diffusion Posterior Sampling with Denoiser-Pullback Curvature Guidance and Manifold-Aligned Damping
Seunghyeok Shin (Inha University), Hongki Lim (Inha University)
RestorationDiffusion modelImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Proposes the CLAMP posterior sampling method, which utilizes denoiser pull-back curvature guidance and manifold-aligned damping to achieve more stable and efficient inverse problem solving in conditional sampling of diffusion models.
Geometry-Guided Generative Representation for Functional Brain Graphs
Subati Abulikemu (University of Cambridge), John Suckling (University of Cambridge)
Data SynthesisRepresentation LearningGraph Neural NetworkTransformerDiffusion modelAuto EncoderContrastive LearningGraphBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Propose a geometry-guided unsupervised graph transformer autoencoder for learning compact graph-level representations of dense weighted functional connectivity graphs and generating synthetic connectivity graphs.
Geometry-Guided Modeling of Foundation Features Enables Generalizable Object Shape Deformation Learning
Yiyao Ma (Chinese University of Hong Kong), Qi Dou (Chinese University of Hong Kong)
GenerationPose EstimationDepth EstimationRepresentation LearningRobotic IntelligenceConvolutional Neural NetworkTransformerDiffusion modelFlow-based ModelContrastive LearningOptical FlowImagePoint CloudMeshOrdinary Differential Equation
🎯 What it does: A geometry-guided template deformation learning framework is proposed, which utilizes a single-view image to guide the continuous deformation of 3D shapes through features from a pre-trained 2D base model, achieving precise reconstruction of target objects at the category level while maintaining dense correspondence.
Geometry-Misalignment in Distributional Learning
Tao Wang (University of Victoria), Xiaoting Zhong
Domain AdaptationOptimizationDiffusion modelScore-based ModelFlow-based ModelRectified FlowContrastive LearningImageTabularTime Series
🎯 What it does: This paper studies the problem of mismatch between the parameter space and its intrinsic geometry under the distribution learning objective, proposing a geometric misalignment measure and proving that it leads to inevitable slow convergence of Euclidean first-order methods.
Geometry-Preserving Orthonormal Initialization for Low-Rank Adaptation in RLVR
Ruijia Zhang (Johns Hopkins University), Laixi Shi (Johns Hopkins University)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringContrastive LearningText
🎯 What it does: Under the RLVR (Reinforcement Learning Verifiable Reward) framework, a LoRA low-rank adaptation method with geometric-preserving orthogonal initialization is proposed to stabilize training and improve performance on downstream tasks.
Geometry-Preserving Unsupervised Alignment for Heterogeneous Foundation Models
Shuwen Yu (Yunnan Normal University), Huafeng Li (Kunming University of Science and Technology)
ClassificationSegmentationDomain AdaptationRepresentation LearningTransformerVision Language ModelDiffusion modelScore-based ModelRectified FlowAuto EncoderGenerative Adversarial NetworkContrastive LearningOptical FlowImageTextMultimodality
🎯 What it does: Proposed the GPUA framework, which learns an orthogonal transformation to map the features of vision-specific models (VFM) into the semantic space of vision-language models (VLM), achieving unsupervised cross-model alignment.
GeoMoLa: Geometry-Aware Motion Latents for Learning Robust Manipulation Policies
Yunchao Zhang (University of Hong Kong), Yanchao Yang (University of Hong Kong)
Robotic IntelligenceTransformerReinforcement LearningVision Language ModelDiffusion modelAuto EncoderContrastive LearningImageVideoPoint Cloud
🎯 What it does: Proposes the Geometry-Aware Motion Latents (GeoMoLa) framework, which learns transferable discrete motion latent variables by self-supervised prediction of point cloud evolution over time from single-view RGB-D observations, and generates 6-DoF trajectories in this latent space using diffusion models.
GeoPT: Scaling Physics Simulation via Lifted Geometric Pre-Training
Haixu Wu (MIT), Wojciech Matusik (MIT)
TransformerGenerative Adversarial NetworkContrastive LearningOptical FlowPoint CloudMeshTabularPhysics Related
🎯 What it does: Propose GeoPT, which utilizes geometric data combined with randomly synthesized velocity fields for lifted geometric self-supervised pre-training, thereby providing dynamic-aware priors for neural physics simulators;
GeoReward: Mitigating Contextual Variable Overestimation in Vision-Language Models for Cross-Market Preference Prediction
Shuo Liu (Alibaba International Digital Commerce Group), Xiaoyi Zeng (Alibaba International Digital Commerce Group)
Recommendation SystemFederated LearningExplainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelDiffusion modelImageTextMultimodalityFinance RelatedRetrieval-Augmented Generation
🎯 What it does: Propose the GeoReward framework to address erroneous decision-making in Vision-Language Models for cross-national advertising image preference prediction caused by overestimation of sparse context variables (CVE).
GeoSense: Internalizing Geometric Necessity Perception for Multimodal Reasoning
Ruiheng Liu (University of Science and Technology of China), Xiaojun Chang (University of Science and Technology of China)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerSupervised Fine-TuningPrompt EngineeringVision Language ModelImageTextMultimodality
🎯 What it does: Propose the GeoSense framework, enabling multimodal large language models to adaptively determine when 3D geometric information is needed and request it on demand.
GePBench: Evaluating Fundamental Geometric Perception for Multimodal Large Language Models
Shangyu Xing (Nanjing University), Xinyu Dai (Nanjing University)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Constructed GePBench, a large-scale geometry-aware benchmark dataset, systematically evaluating the perception capabilities of multi-modal large language models (MLLMs) regarding geometric shapes and their spatial relationships, and enhancing model performance on downstream tasks through retraining on this dataset.
GEPC: Group-Equivariant Posterior Consistency for Out-of-Distribution Detection in Diffusion Models
Rouzoumka Yadang Alexis (ONERA), Chengfang Ren (CentraleSupelEC)
GenerationAnomaly DetectionDiffusion modelScore-based ModelContrastive LearningImage
🎯 What it does: Proposes a training-agnostic group equivariant posterior consistency detection method, GEPC, for OOD detection in diffusion models.
GFD-EMVC: Evolutionary Multi-View Classification with Label Noise via Gradient and Feature Dual-Perception
Shuai Li (Shanxi University), Li Lv (Shanxi University)
ClassificationAnomaly DetectionData-Centric LearningSupervised Fine-TuningContrastive LearningImageTextMultimodality
🎯 What it does: Studied the fitness evaluation bias caused by label noise in evolutionary multi-view classification, and proposed a 'detect-then-calibrate' data purification framework, improving the quality of training data through anomaly detection in the gradient space and feature space prototype calibration.
GFedCL: Graph-Based Federated Continual Learning with Spatial and Temporal Awareness
Qingyang Yu (Stevens Institute of Technology), Hao Wang (Stevens Institute of Technology)
Data SynthesisFederated LearningSafty and PrivacyGraph Neural NetworkGenerative Adversarial NetworkContrastive LearningImageTextBiomedical Data
🎯 What it does: Propose GFedCL, a method that utilizes spatial and temporal attention to generate relational graphs and employs adversarial generative models in federated continual learning to produce high-quality synthetic data, aiming to alleviate catastrophic forgetting and statistical heterogeneity.
GFFMERGE: Efficient Merging of Graph Neural Force Fields and Beyond
Parth Verma (Indian Institute of Technology Delhi), Sayan Ranu (Indian Institute of Technology Delhi)
OptimizationComputational EfficiencyKnowledge DistillationDrug DiscoveryGraph Neural NetworkSupervised Fine-TuningContrastive LearningGraphTabularTime SeriesPhysics Related
🎯 What it does: Propose the GFFMERGE framework, achieving closed-form linear merging of graph neural network force field models, and quickly restoring joint training performance through minimal subsequent fine-tuning; simultaneously extended to general GNNs as GNNMERGE.
GFMate: Empowering Graph Foundation Models with Test-time Prompt Tuning
Yan Jiang (University of Queensland), Zi Huang (University of Queensland)
ClassificationDomain AdaptationRepresentation LearningGraph Neural NetworkSupervised Fine-TuningPrompt EngineeringContrastive LearningGraph
🎯 What it does: Propose the GFMate framework, which achieves no pre-training coupling and fine-tuning with prompts at test time for graph foundation models, leveraging central point prompts, layer prompts, and test-time complementary learning to enhance cross-domain node and graph classification performance.
GHOST: Geometry-Guided Hallucination of Opaque Surface Textures
Langxu Zhao (Northeastern University), Tianhan Gao (Northeastern University)
RestorationDepth EstimationConvolutional Neural NetworkTransformerDiffusion modelAuto EncoderGenerative Adversarial NetworkContrastive LearningOptical FlowImageMultimodalityPoint CloudMeshBenchmark
🎯 What it does: Proposes a general preprocessing framework called GHOST, which leverages visual foundation models to convert the pixels of transparent objects into structurally consistent and texture-rich opaque representations, significantly improving performance without modifying downstream depth estimation and 3D reconstruction models.