NeurIPS 2025 Papers — Page 20
Conference on Neural Information Processing Systems · 5275 papers
GenPO: Generative Diffusion Models Meet On-Policy Reinforcement Learning
Shutong Ding (ShanghaiTech University), Ye Shi (ShanghaiTech University)
Robotic IntelligenceReinforcement LearningDiffusion model
🎯 What it does: Proposes the GenPO method, integrating generative diffusion models into an on-policy reinforcement learning framework to achieve reversible diffusion policies and computable log likelihood;
Geo-Sign: Hyperbolic Contrastive Regularisation for Geometrically Aware Sign Language Translation
Edward Fish (University of Surrey), Richard Bowden (University of Surrey)
Image TranslationPose EstimationGraph Neural NetworkLarge Language ModelContrastive LearningVideoText
🎯 What it does: Proposes the Geo-Sign framework, using hyperbolic geometry regularization for skeletal representation to enhance sign language translation quality.
GeoAda: Efficiently Finetune Geometric Diffusion Models with Equivariant Adapters
Wanjia Zhao (Stanford University), Stefano Ermon (Stanford University)
GenerationPose EstimationDrug DiscoveryGraph Neural NetworkDiffusion modelGraphTabular
🎯 What it does: This paper proposes a SE(3)-equivariant adapter (GeoAda) that allows for lightweight fine-tuning of downstream tasks while maintaining geometric consistency, without altering the pre-trained geometric diffusion model's front structure.
GeoCAD: Local Geometry-Controllable CAD Generation with Large Language Models
Zhanwei Zhang (Zhejiang University), Deng Cai (Zhejiang University)
GenerationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelText
🎯 What it does: A GeoCAD system based on large language models is proposed, achieving controllable generation of local geometry in CAD models.
GeoClip: Geometry-Aware Clipping for Differentially Private SGD
Atefeh Gilani (Arizona State University), Oliver Kosut
OptimizationSafty and PrivacyGaussian SplattingTabular
🎯 What it does: The GeoClip method is proposed, which utilizes the geometric structure of gradient distribution in different private stochastic gradient descent (DP-SGD) approaches. Gradients are projected onto an adaptive basis, clipped, and then noise is added to improve the privacy-utility balance.
GeoComplete: Geometry-Aware Diffusion for Reference-Driven Image Completion
Beibei Lin (National University of Singapore), Robby T. Tan (National University of Singapore)
RestorationGenerationTransformerDiffusion modelImagePoint Cloud
🎯 What it does: A geometry-aware diffusion framework named GeoComplete is proposed for reference-driven image completion.
GeoLink: Empowering Remote Sensing Foundation Model with OpenStreetMap Data
Lubin Bai (Peking University), Shihong Du (Peking University)
ClassificationSegmentationGraph Neural NetworkTransformerContrastive LearningImageMultimodality
🎯 What it does: A remote sensing foundational model named GeoLink has been constructed, which directly enhances the image encoder using OpenStreetMap (OSM) vector data, achieving multimodal fusion in pre-training and downstream tasks.
GeoLLaVA-8K: Scaling Remote-Sensing Multimodal Large Language Models to 8K Resolution
Fengxiang Wang (National University of Defense Technology), Jing Zhang (Wuhan University)
Data SynthesisSuper ResolutionRetrievalTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes a super high-resolution (8K×8K) multimodal large language model for remote sensing, named GeoLLaVA-8K, and constructs two large-scale UHR remote sensing image-text datasets, SuperRS-VQA and HighRS-VQA.
Geometric Algebra-Enhanced Bayesian Flow Network for RNA Inverse Design
Rubo Wang (Institute of Microelectronics, Chinese Academy of Sciences), Peilin Zhao (School of Artificial Intelligence, Shanghai Jiao Tong University)
GenerationDrug DiscoveryProtein Structure PredictionGraph Neural NetworkFlow-based ModelBiomedical Data
🎯 What it does: A Bayesian Flow Network (RBFN) enhanced by geometric algebra is proposed for the reverse generation of corresponding nucleotide sequences given the 3D backbone of RNA.
Geometric Algorithms for Neural Combinatorial Optimization with Constraints
Nikolaos Karalias, Stefanie Jegelka
OptimizationGraph Neural NetworkGraph
🎯 What it does: An end-to-end differentiable geometric decomposition framework is designed for constrained combinatorial optimization under self-supervised learning.
Geometric Imbalance in Semi-Supervised Node Classification
Liang Yan (Fudan University), Zengfeng Huang (Logs AI)
ClassificationOptimizationGraph Neural NetworkGraph
🎯 What it does: This study investigates the geometric imbalance problem in semi-supervised node classification, proposing a theoretical analysis and designing the UNREAL framework to improve pseudo-label generation and fuzzy sample filtering.
Geometric Logit Decoupling for Energy-Based Graph Out-of-distribution Detection
Min Wang (National University of Defense Technology), Jincai Huang (National University of Defense Technology)
Anomaly DetectionGraph Neural NetworkGraphTime Series
🎯 What it does: This paper proposes GeoEnergy, a method for improving energy-based graph OOD detection geometrically without modifying the GNN architecture or training process.
Geometric Mixture Models for Electrolyte Conductivity Prediction
Anyi Li (Renmin University of China), Wenbing Huang (Renmin University of China)
Graph Neural NetworkGraphTabular
🎯 What it does: The GeoMix framework is proposed, utilizing Set-SE(3) equivalence and geometric graph representation to predict the conductivity of electrolyte systems, and achieving fine-grained message passing of inter-molecular geometric information through the GIN module.
Geometry Aware Operator Transformer as an efficient and accurate neural surrogate for PDEs on arbitrary domains
Shizheng Wen (ETH Zurich), Siddhartha Mishra (ETH Zurich)
OptimizationComputational EfficiencyTransformerPoint CloudPhysics Related
🎯 What it does: A geometric-aware operator transformer named GAOT is designed and implemented to efficiently and accurately learn PDE operators over arbitrary domains.
Geometry Meets Incentives: Sample-Efficient Incentivized Exploration with Linear Contexts
Benjamin Schiffer (Harvard University), Mark Sellke (Harvard University)
OptimizationReinforcement Learning
🎯 What it does: This paper proposes a Bayesian Incentive Compatible (BIC) exploration algorithm in a linear context, capable of achieving λ-spectral exploration over a multi-dimensional unit ball action set, thereby ensuring that subsequent Thompson Sampling also remains BIC. The algorithm overcomes the exponential sample complexity bottleneck caused by high dimensions through an initial exploration phase and an exponential growth phase.
Geometry of Decision Making in Language Models
Abhinav Joshi (Indian Institute of Technology Kanpur), Ashutosh Modi (Indian Institute of Technology Kanpur)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: A large-scale quantitative analysis of the intrinsic geometry (intrinsic dimension, ID) of hidden layer representations in 28 open-source transformer models for the multiple-choice question answering (MCQA) task, exploring the relationship between ID variation across layers and model decisions;
Geometry-Aware Collaborative Multi-Solutions Optimizer for Model Fine-Tuning with Parameter Efficiency
Van-Anh Nguyen (Monash University), Dinh Phung (Monash University)
Domain AdaptationOptimizationSupervised Fine-TuningImage
🎯 What it does: A lightweight multi-solution optimization framework GAC‑MSO based on gradient flow and geometric structure is proposed for efficiently fine-tuning large-scale foundational models with parameter efficiency, generating diverse and collaborative solution sets.
Geometry-Aware Edge Pooling for Graph Neural Networks
Katharina Limbeck (Helmholtz Munich), Bastian Rieck (Helmholtz Munich)
Graph Neural NetworkGraph
🎯 What it does: Two edge contraction graph pooling layers based on graph size (Magnitude) or spread (Spread) are proposed (MagEdgePool and SpreadEdgePool).
GeoRanker: Distance-Aware Ranking for Worldwide Image Geolocalization
Pengyue Jia (City University of Hong Kong), Sharon Li (University of Wisconsin-Madison)
RetrievalTransformerPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes GeoRanker, a distance-aware ranking framework based on a large visual-language model, designed to select locations from a candidate set that are closest to the geographic location of a query image.
GeoRemover: Removing Objects and Their Causal Visual Artifacts
Zixin Zhu (University at Buffalo), Junsong Yuan (University at Buffalo)
Image TranslationRestorationSegmentationDepth EstimationOptimizationSupervised Fine-TuningDiffusion modelImageBenchmark
🎯 What it does: A two-stage geometry-aware framework is proposed to clearly remove target objects and their causal visual artifacts (shadows, reflections); the first stage completes geometric removal in the depth map, and the second stage renders realistic RGB images based on the updated geometry.
GeoSVR: Taming Sparse Voxels for Geometrically Accurate Surface Reconstruction
Jiahe Li (Beihang University), Lin Gu (RIKEN AIP)
RestorationSegmentationDepth EstimationPoint Cloud
🎯 What it does: This paper proposes GeoSVR, an explicit surface reconstruction framework based on sparse voxels, which achieves high-precision, complete, and efficient geometric reconstruction by utilizing voxel uncertainty depth constraints and voxel dropout regularization.
GeoVideo: Introducing Geometric Regularization into Video Generation Model
Yunpeng Bai (University of Texas at Austin), Qixing Huang (DAMO Academy, Alibaba Group)
GenerationData SynthesisDepth EstimationDiffusion modelVideoPoint Cloud
🎯 What it does: In the pre-trained latent diffusion video generation model, depth prediction for each frame is incorporated, and cross-frame geometric regularization (utilizing camera pose and depth reprojection) is employed to enhance the spatiotemporal consistency and 3D structure of the video.
GeRaF: Neural Geometry Reconstruction from Radio Frequency Signals
Jiachen Lu (Ecole Polytechnique Federale de Lausanne), Haitham Al Hassanieh (Ecole Polytechnique Federale de Lausanne)
Object DetectionRobotic IntelligenceNeural Radiance FieldPoint CloudMesh
🎯 What it does: The GeRaF method is proposed, which utilizes neural implicit representations to reconstruct millimeter-level 3D geometries from millimeter-wave radar signals.
GFM-RAG: Graph Foundation Model for Retrieval Augmented Generation
Linhao Luo (Monash University), Shirui Pan (Griffith University)
GenerationRetrievalGraph Neural NetworkLarge Language ModelTextGraphRetrieval-Augmented Generation
🎯 What it does: A graph-based model GFM-RAG was constructed based on a knowledge graph index to achieve one-time multi-hop retrieval and enhance the reasoning ability of large language models.
GIST: Greedy Independent Set Thresholding for Max-Min Diversification with Submodular Utility
Matthew Fahrbach (Google), Giulia DeSalvo (Google)
OptimizationImage
🎯 What it does: This paper proposes a subset selection problem (MDMS) for the joint optimization of maximum-minimum diversity and monotone submodular functions, and presents an efficient approximation algorithm GIST.
Glance2Gaze: Efficient Vision-Language Models from Glance Fusion to Gaze Compression
Juan Chen (South China University of Technology), Jintao Fang (Meituan Inc.)
CompressionComputational EfficiencyTransformerLarge Language ModelVision Language ModelImageVideo
🎯 What it does: The Glance2Gaze framework is proposed, simulating the human visual 'glance-to-gaze' two-stage attention mechanism. It collects global semantics through Glance Fusion and iteratively compresses visual tokens within the LLM using Gaze Compression, enhancing the efficiency and performance of Vision-Language models.
GLID$^2$E: A Gradient-Free Lightweight Fine-tune Approach for Discrete Biological Sequence Design
Hanqun Cao (Chinese University of Hong Kong), Pheng-Ann Heng (Chinese University of Hong Kong)
GenerationOptimizationReinforcement LearningDiffusion modelBiomedical Data
🎯 What it does: A lightweight reinforcement learning framework GLID E is proposed for fine-tuning pre-trained discrete diffusion models to generate DNA and protein sequences with target functions.
GLNCD: Graph-Level Novel Category Discovery
Bowen Deng (Sun Yat-sen University), Chuan Chen (Sun Yat-sen University)
Graph Neural NetworkGraphBenchmark
🎯 What it does: A new task of Graph-Level New Category Discovery (GLNCD) is proposed, and four cross-domain benchmark datasets are constructed.
Global Convergence for Average Reward Constrained MDPs with Primal-Dual Actor Critic Algorithm
Yang Xu (Purdue University), Vaneet Aggarwal (Purdue University)
OptimizationReinforcement Learning
🎯 What it does: A globally convergent Primal-Dual Natural Actor-Critic algorithm for infinite average reward constrained Markov Decision Processes (CMDP) is proposed, which achieves constraint satisfaction and reward maximization without requiring model knowledge.
Global Minimizers of $\ell^p$-Regularized Objectives Yield the Sparsest ReLU Neural Networks
Julia B Nakhleh, Robert D Nowak
OptimizationTabular
🎯 What it does: This paper proposes and proves a method for ℓp path norm regularization based on 0 < p < 1, which can directly obtain the sparsest interpolation solution for single hidden layer ReLU networks through gradient descent.
Global Minimizers of Sigmoid Contrastive Loss
Kiril Bangachev (Massachusetts Institute of Technology), Yury Polyanskiy (Massachusetts Institute of Technology)
RetrievalOptimizationContrastive LearningImageMultimodality
🎯 What it does: This paper analyzes the Sigmoid contrastive loss (trainable temperature and bias) from a theoretical perspective, characterizing its global optimal solution for the first time in practical dimensions and sample sizes (d ≪ N ≪ 2d). It introduces the (m_b, rel)-Constellation geometric object to describe all zero-loss configurations; subsequently, it proves that zero-loss configurations necessarily exhibit linear separability between modalities (modality gap) and provides corresponding boundaries and construction methods. Additionally, it proposes a reparameterization of the Sigmoid loss with relative bias as a parameter to enhance training stability and implicitly implement a linear adapter. Finally, experiments are conducted on the SigLIP model and the ImageNet validation set.
Global Prompt Refinement with Non-Interfering Attention Masking for One-Shot Federated Learning
Zhuang Qi (Shandong University), Xiangxu Meng (Shandong University)
Federated LearningPrompt EngineeringImage
🎯 What it does: This study investigates a round of federated prompt learning and proposes the GPR‑NIAM method, which enhances the model's cross-task generalization ability through a two-stage strategy of attention isolation masks and cross-machine collaborative refinement based on a frozen CLIP model.
Globally Optimal Policy Gradient Algorithms for Reinforcement Learning with PID Control Policies
Vipul Kumar Sharma, S Sivaranjani
OptimizationReinforcement LearningTime Series
🎯 What it does: This paper proposes a global optimal optimization framework that combines the policy gradient method in reinforcement learning with the parameterization of PID controllers, providing the gradient expression for the PID control problem and designing both model-based and model-free policy gradient algorithms based on this.
Glocal Information Bottleneck for Time Series Imputation
Jie Yang (University of Illinois Chicago), Kaize Ding (Northwestern University)
TransformerTime Series
🎯 What it does: A new training paradigm for missing value imputation in time series, Glocal-IB, is proposed, which incorporates global alignment loss into the standard information bottleneck framework to address the issues of model overfitting to local noise and inability to capture global structure under high missing rates.
GLSim: Detecting Object Hallucinations in LVLMs via Global-Local Similarity
Seongheon Park (University of Wisconsin Madison), Sharon Li (University of Wisconsin Madison)
RecognitionObject DetectionTransformerVision Language ModelImageMultimodality
🎯 What it does: The GLSIM framework is proposed, which uses global and local embedding similarity within the model to detect object hallucinations in large visual-language models.
GLVD: Guided Learned Vertex Descent
Pol Caselles, Francesc Moreno-Noguer
RestorationPose EstimationOptimizationConvolutional Neural NetworkImageMesh
🎯 What it does: A hybrid 3D face reconstruction framework called GLVD is proposed, which utilizes Learned Vertex Descent and dynamically predicted 3D key points for progressive mesh optimization.
GMM-based VAE model with Normalising Flow for effective stochastic segmentation
Conghui Li (Monash University), Xin Wang (Monash University)
SegmentationFlow-based ModelAuto EncoderImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A conditional variational autoencoder framework based on Gaussian mixture models and regularized flows is proposed to better model uncertainty in random segmentation tasks.
GMV: A Unified and Efficient Graph Multi-View Learning Framework
Qipeng zhu, Junping Zhang (Fudan University)
ClassificationComputational EfficiencyGraph Neural NetworkGraph
🎯 What it does: A unified and efficient graph multi-view learning framework GMV is proposed, which enhances the generalization and robustness of GNN/GT in graph classification tasks by utilizing structure-enhanced subgraph sampling and mixing, multi-view decomposition, and dual-head prediction.
GnnXemplar: Exemplars to Explanations - Natural Language Rules for Global GNN Interpretability
Burouj Armgaan (Indian Institute of Technology Delhi), Sayan Ranu (Indian Institute of Technology Delhi)
Explainability and InterpretabilityGraph Neural NetworkLarge Language ModelGraph
🎯 What it does: This paper proposes GNNXEMPLAR, a global explanation framework based on exemplars, which utilizes natural language rules to explain the predictions of GNN in node classification tasks.
Go With the Flow: Fast Diffusion for Gaussian Mixture Models
George Rapakoulias (Georgia Institute of Technology), Panagiotis Tsiotras (Georgia Institute of Technology)
GenerationData SynthesisOptimizationDiffusion modelAuto EncoderImageBiomedical Data
🎯 What it does: This paper proposes a training-free, low-complexity analytical parameter method that decomposes the Schrödinger bridge problem into a series of Gaussian bridge subproblems and solves the mixed strategy using linear programming, thereby achieving distribution transfer from one Gaussian mixture model to another.
GoalLadder: Incremental Goal Discovery with Vision-Language Models
Alexey Zakharov (University of Oxford), Shimon Whiteson (University of Oxford)
Robotic IntelligenceReinforcement LearningVision Language ModelImageMultimodality
🎯 What it does: A method called GoalLadder is proposed, which utilizes a Visual-Language Model (VLM) to progressively discover and approach the task goal state under a single language instruction, thereby providing a reward signal for reinforcement learning (RL) agents.
GOATex: Geometry & Occlusion-Aware Texturing
Hyunjin Kim (KRAFTON AI), Wonkwang Lee (Seoul National University)
GenerationData SynthesisDiffusion modelMesh
🎯 What it does: This paper proposes the GOATEX method, which achieves seamless texture generation for both external and internal surfaces of 3D meshes using ray hierarchical visibility analysis and diffusion models.
Gompertz Linear Units: Leveraging Asymmetry for Enhanced Learning Dynamics
Indrashis Das (University of Freiburg), Frank Hutter (University of Freiburg)
ClassificationObject DetectionSegmentationConvolutional Neural NetworkTransformerDiffusion modelImageText
🎯 What it does: This paper proposes and experiments with a self-gated activation function called GoLU based on the Gompertz function, which can reduce feature variance and smooth the loss surface through right-skewed asymmetrical gating.
GOOD: Training-Free Guided Diffusion Sampling for Out-of-Distribution Detection
Xin Gao (Carnegie Mellon University), Chenyang Si (Carnegie Mellon University)
Anomaly DetectionConvolutional Neural NetworkDiffusion modelImage
🎯 What it does: A training-free guidance method based on diffusion models (GOOD) is proposed, which utilizes a pre-trained classifier to provide gradient guidance at both the pixel and feature levels, directly generating diverse and semantically meaningful OOD samples to enhance the model's OOD detection performance.
GoRA: Gradient-driven Adaptive Low Rank Adaptation
haonan he, lei chen
TransformerSupervised Fine-TuningText
🎯 What it does: Proposes the GoRA framework, which utilizes gradient information to dynamically allocate the rank of LoRA before training and provides non-zero initialization for low-rank adapters;
GoT: Unleashing Reasoning Capability of MLLM for Visual Generation and Editing
Rongyao Fang (Chinese University of Hong Kong), Hongsheng Li (Chinese University of Hong Kong)
GenerationTransformerLarge Language ModelDiffusion modelImageTextMultimodalityChain-of-Thought
🎯 What it does: The Generation Chain-of-Thought (GoT) framework is proposed, which first generates a reasoning chain containing semantic and spatial coordinates through a multimodal large language model, and then drives a diffusion model to complete image generation and editing.
GPAS: Accelerating Convergence of LLM Pretraining via Gradient-Preserving Activation Scaling
Tianhao Chen (Hong Kong University of Science and Technology), Can Yang (Hong Kong University of Science and Technology)
OptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposes Gradient-Preserving Activation Scaling (GPAS), which accelerates pre-training convergence by applying learnable scaling to intermediate activations in Pre-LN Transformers while maintaining gradient magnitude.
GPLQ: A General, Practical, and Lightning QAT Method for Vision Transformers
Guang Liang (Nanjing University), Jianxin Wu (Nanjing University)
OptimizationComputational EfficiencyTransformerImage
🎯 What it does: A two-stage low-bit quantization framework called GPLQ has been developed, which first performs a single round of Quantization-Aware Training (QAT) on the activations of the Vision Transformer and then applies Post-Training Quantization (PTQ) to the weights, achieving a high-precision model with 4-bit numerical accuracy.
GPO: Learning from Critical Steps to Improve LLM Reasoning
Jiahao Yu (Northwestern University), Xinyu Xing (Northwestern University)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes the GPO (Guided Pivotal Optimization) strategy, which enhances multi-step reasoning performance by identifying and focusing on critical steps in the reasoning trajectory generated by LLMs for fine-grained model tuning.
GPSToken: Gaussian Parameterized Spatially-adaptive Tokenization for Image Representation and Generation
Zhengqiang ZHANG, Lei Zhang (Hong Kong Polytechnic University)
GenerationData SynthesisTransformerGaussian SplattingImage
🎯 What it does: This paper proposes GPSToken, a method that utilizes two-dimensional Gaussian parameterization to achieve non-uniform, spatially adaptive image segmentation, which is used for image representation and generation.
Gradient Alignment in Physics-informed Neural Networks: A Second-Order Optimization Perspective
Sifan Wang (Yale University), Paris Perdikaris (University of Pennsylvania)
OptimizationComputational EfficiencyPhysics RelatedOrdinary Differential Equation
🎯 What it does: This study investigates the gradient direction conflict problem in the training of Physics-Informed Neural Networks (PINN), proposing a gradient alignment metric and demonstrating that second-order optimization methods can alleviate the conflict.
Gradient Descent as Loss Landscape Navigation: a Normative Framework for Deriving Learning Rules
John Vastola, Kanaka Rajan (Harvard University)
OptimizationConvolutional Neural NetworkImage
🎯 What it does: Treating learning rules as a navigation problem in partially observable loss landscapes, this paper proposes deriving gradient descent, momentum, natural gradient, Adam, and continual learning strategies from a single objective function through optimal control and Bayesian inference.
Gradient Multi-Normalization for Efficient LLM Training
Meyer Scetbon (Microsoft Research), Edward Meeds
OptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: A Gradient Multi-Normalization framework is proposed, and based on this, a lightweight optimizer SinkGD is designed without internal states, implementing matrix gradient preprocessing through row/column Euclidean normalization.
Gradient Variance Reveals Failure Modes in Flow-Based Generative Models
Teodora Reu (University of Oxford), Francisco Vargas (Xaira Therapeutics)
GenerationData SynthesisFlow-based ModelRectified FlowImage
🎯 What it does: This paper reveals that the ODE-based Rectified Flow model experiences issues of memorization and convergence stagnation when using deterministic interpolation by analyzing gradient variance, and proposes that introducing a small amount of noise during training can alleviate this problem.
Gradient-Guided Epsilon Constraint Method for Online Continual Learning
Song Lai (City University of Hong Kong), Qingfu Zhang (City University of Hong Kong)
OptimizationConvolutional Neural NetworkReinforcement LearningSequential
🎯 What it does: Reinterpret the experience replay method in online continual learning from the perspective of ε-constraint optimization, and propose a dynamic gradient-guided ε-constraint method (GEC) to balance learning new tasks and retaining past tasks.
Gradient-Variation Online Adaptivity for Accelerated Optimization with Hölder Smoothness
Yuheng Zhao (Nanjing University), Peng Zhao
Optimization
🎯 What it does: This paper studies gradient variation in online learning under Hölder smooth functions, deriving the optimal regret upper bound for convex and strongly convex online functions, and transforming it into a universal offline optimization algorithm through online-to-batch conversion, achieving accelerated convergence rates whether the smoothness parameter is known or unknown.
Gradient-Weight Alignment as a Train-Time Proxy for Generalization in Classification Tasks
Florian A. Hölzl (Institute for Artificial Intelligence in Medicine Technical University of Munich), Georgios Kaissis (Institute for Artificial Intelligence in Medicine Technical University of Munich)
ClassificationExplainability and InterpretabilityTransformerImage
🎯 What it does: A gradient-weight alignment (GWA) metric is proposed to evaluate model generalization during training and identify important training samples.
GradMetaNet: An Equivariant Architecture for Learning on Gradients
Yoav Gelberg (University of Oxford), Haggai Maron (Technion)
OptimizationMeta LearningTransformerImageText
🎯 What it does: This paper proposes GradMetaNet, a equivariant architecture for gradient sets, aimed at learning gradient processing tasks.
GraLoRA: Granular Low-Rank Adaptation for Parameter-Efficient Fine-Tuning
Yeonjoon Jung (SqueezeBits), Eunhyeok Park (POSTECH)
OptimizationComputational EfficiencyAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A new parameter-efficient fine-tuning method called GraLoRA is proposed, which splits the low-rank adapters of LoRA into sub-blocks, allowing each sub-block to learn independently, thereby improving the model's performance across different tasks and model sizes.
Grammars of Formal Uncertainty: When to Trust LLMs in Automated Reasoning Tasks
Debargha Ganguly (Case Western Reserve University), Vipin Chaudhary (Case Western Reserve University)
TransformerLarge Language ModelText
🎯 What it does: This study investigates the implicit uncertainty of large language models in generating formal reasoning (SMT-LIB), constructs a uncertainty quantification framework based on Probabilistic Context-Free Grammar (PCFG), and achieves selective verification through lightweight signal fusion, significantly reducing the error rate.
GRAPE: Optimize Data Mixture for Group Robust Multi-target Adaptive Pretraining
Simin Fan (École Polytechnique Fédérale de Lausanne), Martin Jaggi (École Polytechnique Fédérale de Lausanne)
Domain AdaptationOptimizationLarge Language ModelText
🎯 What it does: This paper proposes GRAPE, an adaptive data mixing reweighting method for multi-source-multi-target LLM pre-training tasks.
Graph Alignment via Birkhoff Relaxation
Sushil Mahavir Varma (University of Michigan), Laurent Massoulié (INRIA)
OptimizationGraph Neural NetworkGraphPhysics Related
🎯 What it does: This paper studies the theoretical performance of Birkhoff relaxation under the Gaussian Wigner model, proving that it approximates the optimal permutation when the noise level σ=o(n^{-1}), and is far from the optimal permutation when σ=Ω(n^{-1/2}), providing corresponding error bounds and phase transition thresholds.
Graph Data Selection for Domain Adaptation: A Model-Free Approach
Ting-Wei Li, Hanghang Tong
Domain AdaptationGraph Neural NetworkGraph
🎯 What it does: A model-free graph data selection framework called GRADATE is proposed, which selects the most beneficial training graph samples from the source domain for the target domain using Graph Data Distribution Distance (GDD);
Graph Diffusion that can Insert and Delete
Matteo Ninniri (University of Pisa), Davide Bacciu (University of Pisa)
GenerationData SynthesisDrug DiscoveryGraph Neural NetworkTransformerDiffusion modelGraph
🎯 What it does: This paper proposes GRIDDD, a discrete graph diffusion probability model that supports dynamic insertion and deletion of nodes during the diffusion process for variable-sized molecular generation.
Graph Few-Shot Learning via Adaptive Spectrum Experts and Cross-Set Distribution Calibration
Yonghao Liu (Jilin University), Renchu Guan (Jilin University)
Representation LearningMeta LearningGraph Neural NetworkMixture of ExpertsGraph
🎯 What it does: A new framework for few-shot learning in graphs, called GRACE, is proposed, which addresses local structural heterogeneity and support/query distribution shifts through adaptive spectral experts and cross-set distribution calibration.
Graph Neural Network Based Action Ranking for Planning
Rajesh Devaraddi Mangannavar, Prasad Tadepalli (Oregon State University)
Graph Neural NetworkGraph
🎯 What it does: A motion ranking learning method based on graph neural networks, GABAR, is proposed to generate general strategies in large classical planning problems.
Graph Persistence goes Spectral
Mattie Ji (University of Pennsylvania), Vikas K Garg
Representation LearningGraph Neural NetworkGraph
🎯 What it does: The researchers proposed a new topological descriptor called SpectRe, which combines spectral information with RePHINE for graph representation learning.
Graph Your Own Prompt
Xi Ding (Griffith University), Yongsheng Gao (Griffith University)
ClassificationConvolutional Neural NetworkTransformerImage
🎯 What it does: A Graph Consistency Regularization (GCR) framework is proposed, which inserts a parameter-free Graph Consistency Layer (GCL) within the network. By constructing a feature similarity graph and aligning it with the mask similarity graph predicted by the model itself, self-prompting structural regularization is achieved.
Graph-based Symbolic Regression with Invariance and Constraint Encoding
Ziyu Xiang (Texas A&M University), Xiaoning Qian (Brookhaven National Laboratory)
Graph Neural NetworkReinforcement LearningGraphTabularPhysics Related
🎯 What it does: A symbolic regression framework GSR based on expression graphs (EG) is proposed, utilizing mixed neural-guided MCTS for efficient search and encoding equivalence and constraints into the sampling process.
Graph-KV: Breaking Sequence via Injecting Structural Biases into Large Language Models
Haoyu Peter Wang (Georgia Institute of Technology), Pan Li (Georgia Institute of Technology)
GenerationRetrievalComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningTextGraphBiomedical DataRetrieval-Augmented Generation
🎯 What it does: To address the biases and inefficiencies caused by serialization in large language models when handling structured information (such as retrieval-augmented generation, citation networks, graph node classification, etc.), the Graph-KV method is proposed, which injects structure-induced biases into autoregressive LLMs using KV caching and graph structural edge relationships.
Graph-Theoretic Insights into Bayesian Personalized Ranking for Recommendation
Kai Zheng (Central South University), Jinhui Xu (University of Science and Technology of China)
Recommendation SystemDrug DiscoveryGraph Neural NetworkContrastive LearningMultimodalityGraph
🎯 What it does: This paper analyzes the essence of the Bayesian Personalized Ranking (BPR) loss from the perspective of graph theory and proposes a BPR+ loss that better captures global topology.
Graph–Smoothed Bayesian Black-Box Shift Estimator and Its Information Geometry
Masanari Kimura (University of Melbourne)
Domain AdaptationGraph Neural NetworkImage
🎯 What it does: This paper proposes the Graph-Smoothed Bayesian Black-Box Shift Estimator (GS-BSE) to recover the target prior probabilities under label distribution shift and provides Bayesian uncertainty estimation.
GraphChain: Large Language Models for Large-scale Graph Analysis via Tool Chaining
Chunyu Wei (Renmin University of China), Yueguo Chen (Beijing Jiaotong University)
Graph Neural NetworkLarge Language ModelReinforcement LearningPrompt EngineeringGraphFinance Related
🎯 What it does: Proposes the GraphChain framework, which utilizes LLM and toolchains for step-by-step analysis of large-scale graph data.
GraphKeeper: Graph Domain-Incremental Learning via Knowledge Disentanglement and Preservation
Zihao Guo (Beihang University), Jianxin Li (Guangxi Normal University)
Domain AdaptationGraph Neural NetworkSupervised Fine-TuningContrastive LearningGraph
🎯 What it does: The GraphKeeper framework is proposed to address the problem of catastrophic forgetting in incremental learning for graphs (Domain-IL), maintaining performance on previous domains while continuously adding new graph domains.
GraphMaster: Automated Graph Synthesis via LLM Agents in Data-Limited Environments
Enjun Du (Beijing Institute of Technology), Guoren Wang (Beijing Institute of Technology)
GenerationData SynthesisGraph Neural NetworkLarge Language ModelAgentic AITextGraphRetrieval-Augmented Generation
🎯 What it does: A multi-agent framework called GraphMaster is proposed, which utilizes LLM to generate semantically rich and structurally consistent text attribute graphs.
Graphs Help Graphs: Multi-Agent Graph Socialized Learning
Jialu Li (Tianjin University), Qinghua Hu (Tianjin University)
Graph Neural NetworkPrompt EngineeringGraph
🎯 What it does: This paper proposes the Graph Socialized Learning (GSL) framework and its implementation method, Graphs Help Graphs (GHG), to achieve efficient collaborative learning among multiple agents in heterogeneous dynamic environments.
GraphTOP: Graph Topology-Oriented Prompting for Graph Neural Networks
Xingbo Fu (University of Virginia), Jundong Li (University of Virginia)
ClassificationGraph Neural NetworkPrompt EngineeringGraph
🎯 What it does: Proposes the GraphTOP framework, which utilizes graph topology hints to adapt pre-trained GNNs for node classification.
Grasp2Grasp: Vision-Based Dexterous Grasp Translation via Schrödinger Bridges
Tao Zhong (Princeton University), Christine Allen-Blanchette (Lockheed Martin Corporation)
Robotic IntelligenceTransformerAuto EncoderImageStochastic Differential Equation
🎯 What it does: This paper proposes a visual-driven grasp translation framework based on the Schrödinger bridge, which can convert the grasping intention of one robotic hand under visual perception into an executable grasp for another robotic hand of a different form.
GraSS: Scalable Data Attribution with Gradient Sparsification and Sparse Projection
Pingbang Hu (University of Illinois Urbana-Champaign), Jiaqi W. Ma (University of Illinois Urbana-Champaign)
OptimizationComputational EfficiencyText
🎯 What it does: This paper proposes two gradient compression algorithms, GRASS and FACTGRASS, which significantly reduce the memory and computational costs of large-scale model data attribution by leveraging the natural sparsity of gradients and parameters.
GRAVER: Generative Graph Vocabularies for Robust Graph Foundation Models Fine-tuning
Haonan Yuan (Beihang University), Philip S. Yu (University of Illinois)
ClassificationDomain AdaptationRepresentation LearningGraph Neural NetworkLarge Language ModelMixture of ExpertsContrastive LearningGraph
🎯 What it does: Proposes the GRAVER framework, which enhances the support set using a generative graph dictionary to achieve robust and efficient fine-tuning of graph-based models under multi-domain pre-training.
GRE Suite: Geo-localization Inference via Fine-Tuned Vision-Language Models and Enhanced Reasoning Chains
Chun Wang (Zhejiang University), Yiren Song (LibLib AI)
TransformerSupervised Fine-TuningReinforcement LearningVision Language ModelImageChain-of-Thought
🎯 What it does: The GRE Suite framework is proposed, combining visual language models with multi-stage reinforcement learning to enhance inference for global image geolocation.
Greed is Good: A Unifying Perspective on Guided Generation
Zander W. Blasingame (Clarkson University), Chen Liu (Clarkson University)
RestorationGenerationData SynthesisDiffusion modelImageMultimodalityOrdinary Differential Equation
🎯 What it does: This paper proposes a unified perspective, treating training-independent posterior guidance and end-to-end guidance as different implementations of a greedy strategy, and based on this, designs an intermediate method.
Greedy Algorithms for Structured Bandits: A Sharp Characterization of Asymptotic Success / Failure
Aleksandrs Slivkins (Microsoft Research), Shiliang Zuo (University of Illinois Urbana-Champaign)
OptimizationReinforcement Learning from Human Feedback
🎯 What it does: This paper provides a theoretical analysis of the long-term performance of the greedy algorithm in structured multi-armed bandits (Structured Bandits) with known reward structures, presenting complete criteria for the success (sublinear tuning) and failure (linear tuning) of the greedy approach.
Greedy Sampling Is Provably Efficient For RLHF
Di Wu (University of Virginia), Cong Shen (University of Virginia)
Reinforcement Learning from Human FeedbackReinforcement LearningTabular
🎯 What it does: A greedy sampling algorithm based on empirical estimation is proposed and validated for human feedback reinforcement learning (RLHF) within the context of KL-regularized contextual bandits, applicable to both general preference models and the Bradley-Terry model;
Grids Often Outperform Implicit Neural Representation at Compressing Dense Signals
Namhoon Kim (Georgia Institute of Technology), Sara Fridovich-Keil (Georgia Institute of Technology)
RestorationSuper ResolutionCompressionGaussian SplattingImageBiomedical DataComputed TomographyBenchmark
🎯 What it does: Compared various implicit neural representations (INR) with traditional representations such as grid interpolation in terms of compression and reconstruction performance across different dimensions, signal types, and tasks (overfitting, CT reconstruction, super-resolution, denoising), and systematically evaluated their capacity and computational efficiency.
GRIFFIN: Effective Token Alignment for Faster Speculative Decoding
Shijing Hu (Fudan University), Pan Zhou (Singapore Management University)
GenerationComputational EfficiencyTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: This paper proposes a new speculative decoding framework called GRIFFIN, which explicitly addresses the token misalignment issue between the training and inference phases, significantly improving the generation speed of large language models.
GRIP: A Graph-Based Reasoning Instruction Producer
Jiankang Wang (University of Science and Technology of China), Hongtao Xie (University of Science and Technology of China)
Graph Neural NetworkLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: The GRIP framework is proposed, which generates 2.1M high-quality mathematical reasoning question-and-answer pairs by constructing a Key Concept Relationship Graph (KCRG) and utilizing explicit and implicit concept relationships based on a small amount of seed data; simultaneously, the GRIP-MATH dataset is constructed.
GRIT: Teaching MLLMs to Think with Images
Yue Fan (University of California Santa Cruz), Xin Eric Wang (University of California Santa Barbara)
Object DetectionGenerationTransformerLarge Language ModelReinforcement LearningVision Language ModelImageTextMultimodality
🎯 What it does: Train multimodal large language models (MLLMs) to generate a visual reasoning chain that interleaves natural language and bounding box coordinates when answering image-related questions, referred to as 'thinking about images.'
Ground-Compose-Reinforce: Grounding Language in Agentic Behaviours using Limited Data
Andrew C Li (University of Toronto), Sheila A. McIlraith (University of Toronto)
Robotic IntelligenceReinforcement LearningAgentic AISequential
🎯 What it does: An end-to-end framework is proposed to directly train RL agents using a pre-trained symbolic labeler from a limited number of annotated trajectories, followed by high-level tasks described by Reward Machines (RM).
Grounded Reinforcement Learning for Visual Reasoning
Gabriel Herbert Sarch (Carnegie Mellon University), Katerina Fragkiadaki (Carnegie Mellon University)
TransformerReinforcement LearningVision Language ModelImageMultimodality
🎯 What it does: A visual language model visualization-based reinforcement learning framework called ViGoRL is proposed, which achieves interpretable guidance of visual information by explicitly binding image coordinates at each step of inference.
Grounding Language with Vision: A Conditional Mutual Information Calibrated Decoding Strategy for Reducing Hallucinations in LVLMs
Hao Fang (Tsinghua University), Shu-Tao Xia (Tsinghua University)
GenerationOptimizationTransformerLarge Language ModelVision Language ModelImageText
🎯 What it does: This paper proposes a decoding strategy based on Conditional Pointwise Mutual Information (C-PMI), utilizing a dual-layer optimization to simultaneously adjust the text sampling distribution and filter visual tokens, significantly reducing the hallucination phenomenon in large visual language models (LVLM) during the generation process.
Group-in-Group Policy Optimization for LLM Agent Training
Lang Feng (Nanyang Technological University), Bo An (Skywork AI)
OptimizationTransformerLarge Language ModelReinforcement LearningAgentic AIText
🎯 What it does: This paper proposes GiGPO, a group-based reinforcement learning algorithm that enables fine-grained credit allocation in multi-step LLM agent training.
Group-Level Data Selection for Efficient Pretraining
Zichun Yu (Carnegie Mellon University), Chenyan Xiong (Carnegie Mellon University)
OptimizationComputational EfficiencyData-Centric LearningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: An efficient group-level data selection framework called Group-MATES is proposed, which utilizes a relational data influence model to achieve speed-quality trade-off optimization during pre-training.
GSAlign: Geometric and Semantic Alignment Network for Aerial-Ground Person Re-Identification
Qiao Li (Wuhan University), Jiayi Ji
RetrievalTransformerImage
🎯 What it does: This paper proposes a Geometric and Semantic Alignment Network (GSAlign) specifically designed to address the issues of geometric distortion and semantic misalignment caused by extreme viewpoint differences in aerial-ground person retrieval (AG-ReID).
GSPN-2: Efficient Parallel Sequence Modeling
Hongjun Wang (NVIDIA), Sifei Liu (NVIDIA)
GenerationComputational EfficiencyTransformerDiffusion modelImageText
🎯 What it does: Proposes GSPN-2, which achieves efficient parallel two-dimensional sequence modeling through unified CUDA kernels, channel-shared propagation weights, and low-dimensional proxy compression;
GSRF: Complex-Valued 3D Gaussian Splatting for Efficient Radio-Frequency Data Synthesis
Kang Yang (University of California), Mani Srivastava (University of California)
Data SynthesisComputational EfficiencyGaussian SplattingPoint Cloud
🎯 What it does: A framework for RF data synthesis based on complex-valued 3D Gaussian rendering, GSRF, has been developed, which can efficiently generate RF signals such as RSSI, CSI, and spatial spectrum.
GST-UNet: A Neural Framework for Spatiotemporal Causal Inference with Time-Varying Confounding
Miruna Oprescu (Cornell University), Nathan Kallus (Cornell University)
Convolutional Neural NetworkRecurrent Neural NetworkTabularTime SeriesElectronic Health Records
🎯 What it does: The GST-UNet framework is proposed to achieve single-trajectory spatiotemporal causal inference, combining a U-Net encoder with iterative G-computation, capable of simultaneously handling spatial interference, temporal confounding, and spatiotemporal lag effects.
GTR-Loc: Geospatial Text Regularization Assisted Outdoor LiDAR Localization
Shangshu Yu (Northeastern University), Cheng Wang (Xiamen University)
Pose EstimationAutonomous DrivingKnowledge DistillationTransformerContrastive LearningPoint Cloud
🎯 What it does: This paper proposes GTR-Loc, a LiDAR positioning framework that utilizes geospatial text assistance to achieve accurate pose regression on single-frame point clouds.
Guarantees for Alternating Least Squares in Overparameterized Tensor Decompositions
Dionysis Arvanitakis (Northwestern University), Aravindan Vijayaraghavan (Northwestern University)
OptimizationImageText
🎯 What it does: This paper proves that under moderate over-parameterization (k = O(r²)), the randomly initialized Alternating Least Squares (ALS) algorithm can globally converge to the optimal decomposition of third-order tensors in polynomial time with high probability, and can obtain approximate solutions with polynomial relative error in the case of low-rank approximation.
Guard Me If You Know Me: Protecting Specific Face-Identity from Deepfakes
Kaiqing Lin (Shenzhen University), Bin Li (Shenzhen University)
RecognitionSafty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringImageMultimodalityBenchmark
🎯 What it does: The VIPGuard framework is proposed for personalized deepfake detection and interpretable reasoning for known identities.