NeurIPS 2025 Papers — Page 19
Conference on Neural Information Processing Systems · 5275 papers
From stability of Langevin diffusion to convergence of proximal MCMC for non-log-concave sampling
Marien Renaud (University of Bordeaux), Nicolas Papadakis (University of Bordeaux)
RestorationOptimizationDiffusion modelImageStochastic Differential Equation
🎯 What it does: A Proximal Stochastic Gradient Langevin Algorithm (PSGLA) for non-logarithmic concave sampling is proposed, and its convergence proof under non-convex potential is provided. Additionally, model-free prior image posterior sampling is achieved through Plug-and-Play.
From Style to Facts: Mapping the Boundaries of Knowledge Injection with Finetuning
Eric Zhao (Google Research University of California Berkeley), Nika Haghtalab (University of California Berkeley)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper investigates the performance differences of language model fine-tuning in knowledge injection and task customization through large-scale experiments, revealing the impact of factors such as information type, quantity, training data format, and evaluation tasks on fine-tuning effectiveness.
From Synapses to Dynamics: Obtaining Function from Structure in a Connectome Constrained Model of the Head Direction Circuit
Sunny Duan (Massachusetts Institute of Technology), Ila R Fiete
Supervised Fine-TuningReinforcement LearningGraphBiomedical Data
🎯 What it does: Using cell-type-level parameterization and self-supervised learning, a rate-based model of the Drosophila head circuit is constructed based on FlyEM hemibrain connectome data, recovering continuous attractor dynamics.
FSEO: Few-Shot Evolutionary Optimization via Meta-Learning for Expensive Multi-Objective Optimization
Xunzhao Yu (University of Warwick)
OptimizationMeta LearningNeural Architecture SearchTransformerReinforcement LearningAuto EncoderTabular
🎯 What it does: A few-shot evolutionary optimization framework (FSEO) is proposed, combining meta-learning deep kernel Gaussian processes (MDKL) as a regression surrogate to address expensive multi-objective optimization problems with a limited number of samples.
FSI-Edit: Frequency and Stochasticity Injection for Flexible Diffusion-Based Image Editing
Kaixiang Yang (Wuhan National Laboratory for Optoelectronics Huazhong University of Science and Technology), Zhiwei Wang (Wuhan National Laboratory for Optoelectronics Huazhong University of Science and Technology)
Image TranslationGenerationDiffusion modelRectified FlowImageBenchmark
🎯 What it does: A non-tuning diffusion model image editing framework FSI-Edit is proposed, utilizing frequency residual fusion and controlled noise injection to achieve non-rigid (structural or pose changes) editing.
FSNet: Feasibility-Seeking Neural Network for Constrained Optimization with Guarantees
Hoang T. Nguyen (Massachusetts Institute of Technology), Priya L. Donti (Massachusetts Institute of Technology)
Optimization
🎯 What it does: The FSNet framework is proposed, which directly embeds the feasibility finding step into the training and inference process of neural networks to solve constrained parametric optimization problems and ensure that the output meets the constraints.
FUDOKI: Discrete Flow-based Unified Understanding and Generation via Kinetic-Optimal Velocities
Jin Wang (University of Hong Kong), Ping Luo (University of Hong Kong)
GenerationData SynthesisTransformerFlow-based ModelImageTextMultimodality
🎯 What it does: A unified multimodal model FUDOKI based on discrete flow matching is proposed, balancing visual understanding and image generation;
Fully Autonomous Neuromorphic Navigation and Dynamic Obstacle Avoidance
Xiaochen Shang (Dalian University of Technology), Xin Yang (Dalian University of Technology)
Autonomous DrivingOptimizationRobotic IntelligenceSpiking Neural NetworkSimultaneous Localization and MappingImageVideo
🎯 What it does: A fully neuromorphic technology-based autonomous navigation and dynamic obstacle avoidance system for drones has been developed, achieving real-time obstacle avoidance with a low latency of 2.3 ms and a 21% reduction in energy consumption using a monocular event camera and IMU.
Fully Dynamic Algorithms for Chamfer Distance
Gramoz Goranci (University of Vienna), Qiaoyuan Yang (Peking University)
Point CloudTabular
🎯 What it does: This paper proposes an algorithm for maintaining the approximate Chamfer distance between two sets of points in a fully dynamic environment.
Fully Spiking Neural Networks for Unified Frame-Event Object Tracking
Jingjun Yang (National University of Defense Technology), Dewen Hu (National University of Defense Technology)
Object TrackingSpiking Neural NetworkTransformerImageVideoMultimodality
🎯 What it does: A fully spiking neural network called SpikeFET is proposed, specifically designed for unified frame-event visual object tracking, integrating convolutional local feature extraction with Transformer global modeling under the spiking paradigm. It also designs a Random Patch Module (RPM) and Spatiotemporal Regularization (STR) to enhance localization accuracy and energy efficiency.
FuncGenFoil: Airfoil Generation and Editing Model in Function Space
Jinouwen Zhang (Shanghai Artificial Intelligence Laboratory), SHIXIANG TANG
GenerationOrdinary Differential Equation
🎯 What it does: A model for generating and editing airfoil leading edges in function space, called FuncGenFoil, is proposed for high-fidelity airfoil design.
Functional Complexity-adaptive Temporal Tensor Decomposition
Panqi Chen (Zhejiang University), Shikai Fang (Microsoft Research Asia)
Time SeriesOrdinary Differential Equation
🎯 What it does: A functionally adaptive complexity tensor decomposition model CATTE is proposed, which utilizes Fourier feature encoding for continuous indexing and learns factor trajectories through neural ODE, while applying a sparse prior on functional factor trajectories to achieve automatic tensor rank inference.
Functional data analysis for multivariate distributions through Wasserstein slicing
Han Chen (University of California), Hans-Georg Müller (University of California)
TabularTime Series
🎯 What it does: By mapping multidimensional probability distributions to Hilbert space through Radon projection and log quantile density transformation, functional data analysis of multivariate distributions is achieved.
Functional Matching of Logic Subgraphs: Beyond Structural Isomorphism
Ziyang Zheng (Chinese University of Hong Kong), Qiang Xu (Chinese University of Hong Kong)
RecognitionSegmentationGraph Neural NetworkContrastive LearningGraph
🎯 What it does: This paper proposes a new method for 'functional subgraph matching' to identify subgraphs in logic circuits that maintain the same functionality even when there are significant structural changes after synthesis or technology mapping. The method includes a two-stage multimodal framework: the first stage trains functional invariant embeddings to achieve functional subgraph detection, and the second stage transforms fuzzy boundary recognition into a graph segmentation task for fine localization.
Functional Scaling Laws in Kernel Regression: Loss Dynamics and Learning Rate Schedules
Binghui Li (Peking University), Lei Wu (Peking University)
Large Language ModelTime SeriesStochastic Differential Equation
🎯 What it does: This paper studies the impact of learning rate scheduling on the dynamics of SGD loss in kernel regression and proposes the Functional Scaling Law (FSL) to describe the entire training process.
Functional Virtual Adversarial Training for Semi-Supervised Time Series Classification
Qingyi Pan (Tsinghua University), Yicheng Li (Tsinghua University)
ClassificationOptimizationAdversarial AttackRecurrent Neural NetworkGenerative Adversarial NetworkTime SeriesFinance Related
🎯 What it does: This paper proposes a Functional Virtual Adversarial Training (f-VAT) framework for semi-supervised time series classification, which generates perturbations using the structure of the functional space to enhance the model's smoothness.
Fundamental Limitations in Pointwise Defences of LLM Finetuning APIs
Xander Davies (University of Oxford), Yarin Gal (University of Oxford)
Adversarial AttackTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed and validated a type of attack method that is feasible on LLM fine-tuning APIs and cannot be detected by pointwise detection systems, using natural benign inputs and systematic variations in model outputs to convey harmful information;
Fuse2Match: Training-Free Fusion of Flow, Diffusion, and Contrastive Models for Zero-Shot Semantic Matching
Jing Zuo (Beijing University of Posts and Telecommunications), Yi-Zhe Song (University of Surrey)
RetrievalDiffusion modelContrastive LearningImage
🎯 What it does: A training-free multi-model feature fusion method called Fuse2Match is designed, utilizing Stable Diffusion 3 (SD3), the older version of Stable Diffusion (SD), and the contrastive learning model DINO to improve zero-shot semantic correspondence tasks.
Fused View-Time Attention and Feedforward Reconstruction for 4D Scene Generation
Chaoyang Wang (Snap Inc.), Peter Wonka (KAUST)
GenerationData SynthesisTransformerDiffusion modelVideo
🎯 What it does: A two-stage 4D video generation framework is proposed, utilizing a diffusion model to synchronously generate multi-view videos, which are then transformed into dynamic Gaussian particles through a feedforward reconstruction network, achieving the generation of complete 4D scenes from text prompts.
Future Link Prediction Without Memory or Aggregation
Lu Yi (Renmin University of China), Yuhang Ye (Huawei Technology Ltd)
Graph Neural NetworkGraph
🎯 What it does: A new temporal graph learning framework called CRAFT is proposed, which removes the memory and aggregation modules and replaces them with learnable node embeddings and cross-attention mechanisms.
Future-Aware End-to-End Driving: Bidirectional Modeling of Trajectory Planning and Scene Evolution
Bozhou Zhang (Fudan University), Li Zhang (Fudan University)
Autonomous DrivingOptimizationTransformerReinforcement LearningMultimodality
🎯 What it does: The SeerDrive framework is proposed, which achieves end-to-end adaptive driving decision-making by predicting future Bird's Eye View (BEV) scenarios and interacting in a closed-loop with trajectory planning.
FutureSightDrive: Thinking Visually with Spatio-Temporal CoT for Autonomous Driving
Shuang Zeng (Xi'an Jiaotong University), Xing Wei (Xi'an Jiaotong University)
Autonomous DrivingTransformerLarge Language ModelVision Language ModelImageMultimodalityChain-of-Thought
🎯 What it does: The FSDrive framework is proposed, allowing VLA models to think about the future and plan trajectories at the image level through visualized spatiotemporal Chain-of-Thought.
FuXi-Ocean: A Global Ocean Forecasting System with Sub-Daily Resolution
Qiusheng Huang (Fudan University), Hao Li (Fudan University)
TransformerTime Series
🎯 What it does: FuXi-Ocean is proposed, an end-to-end deep learning model that can predict sea temperature, salinity, currents, and sea surface height in the depth range of 0–1500 m at a global scale (1/12° grid) with a 6-hour sub-daily resolution.
Fuz-RL: A Fuzzy-Guided Robust Framework for Safe Reinforcement Learning under Uncertainty
Xu Wan (Zhejiang University), Mingyang Sun (Peking University)
Safty and PrivacyReinforcement Learning
🎯 What it does: This paper proposes Fuz-RL, a robust safe reinforcement learning framework that integrates fuzzy logic, achieving robust value estimation of multi-source uncertainty through fuzzy Bellman operators and Choquet integrals.
G-Memory: Tracing Hierarchical Memory for Multi-Agent Systems
Guibin Zhang (National University of Singapore), Shuicheng YAN
Graph Neural NetworkLarge Language ModelAgentic AITextGraphBenchmark
🎯 What it does: This paper proposes G-Memory, a three-layer graph structure (insight graph, query graph, interaction graph) for recording, retrieving, and updating the long-term collaboration history of multi-agent systems (MAS), supporting the self-evolution of agent teams.
G-Net: A Provably Easy Construction of High-Accuracy Random Binary Neural Networks
Alireza Aghasi (Oregon State University), Wyatt D. Whiting
ClassificationRecognitionOptimizationImage
🎯 What it does: A G-Net framework based on random binary embedding is proposed, which can directly map the trained floating-point neural network to a high-dimensional binary network while maintaining close prediction accuracy; it also provides theoretical guarantees and large-scale experimental validation.
Gains: Fine-grained Federated Domain Adaptation in Open Set
Zhengyi Zhong (National University of Defense Technology), Ju Ren (Tsinghua University)
Domain AdaptationFederated LearningImage
🎯 What it does: Proposes the Gains framework to achieve fine-grained knowledge discovery and rapid balance of knowledge adaptation in open-set federated learning;
GAM-Agent: Game-Theoretic and Uncertainty-Aware Collaboration for Complex Visual Reasoning
Jusheng Zhang (Sun Yat-sen University), Keze Wang (Snap Inc.)
Agentic AIVision Language ModelMultimodalityBenchmark
🎯 What it does: This paper studies the multi-agent framework GAM-Agent based on game theory, which achieves collaborative and self-correcting visual language reasoning through uncertainty-driven debates between baseline agents (visual sub-task experts) and critical agents (logic/fact-checking experts).
GAMMA: Gated Multi-hop Message Passing for Homophily-Agnostic Node Representation in GNNs
Amir Ghazizadeh Ahsaei, Hao Zheng (University of Central Florida)
Computational EfficiencyRepresentation LearningGraph Neural NetworkGraph
🎯 What it does: Proposes GAMMA, a GNN for adaptive multi-hop information aggregation in heterophilic graphs;
GaRA-SAM: Robustifying Segment Anything Model with Gated-Rank Adaptation
Sohyun Lee (POSTECH), Suha Kwak (POSTECH)
RestorationObject DetectionSegmentationTransformerSupervised Fine-TuningImage
🎯 What it does: A technique is proposed to enhance the robustness of the Segment Anything Model (SAM) in various image degradation scenarios through a dynamically adjustable low-rank adapter (GaRA).
GASP: Efficient Black-Box Generation of Adversarial Suffixes for Jailbreaking LLMs
Advik Raj Basani (Birla Institute of Technology and Science), Xiao Zhang (CISPA Helmholtz Center for Information Security)
OptimizationAdversarial AttackTransformerLarge Language ModelText
🎯 What it does: Proposes the GASP framework, which efficiently generates readable malicious suffixes that can bypass security shields in a black-box environment, significantly improving the jailbreak success rate of LLMs.
Gate to the Vessel: Residual Experts Restore What SAM Overlooks
Weili Jiang (Southwest Jiaotong University), Chubin Ou (Guangdong Provincial People's Hospital)
SegmentationConvolutional Neural NetworkMixture of ExpertsBiomedical Data
🎯 What it does: The FineSAM++ framework is proposed, which refines local errors in SAM for medical vascular segmentation through sparse experts and soft routing modules.
Gated Attention for Large Language Models: Non-linearity, Sparsity, and Attention-Sink-Free
Zihan Qiu (Alibaba Group), Junyang Lin (Alibaba Group)
TransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: The system evaluates and conducts comprehensive experiments on the gating mechanism in softmax attention, comparing various gating positions, granularities, whether they are shared, multiplicative/additive methods, and activation functions. It trains a 15B MoE model and a 1.7B dense model on 3.5T of data, analyzing the impact of gating on non-linearity, sparsity, attention drop, training stability, and long sequence generalization.
Gated Integration of Low-Rank Adaptation for Continual Learning of Large Language Models
Yan-Shuo Liang (Nanjing University), Wu-Jun Li (Nanjing University)
TransformerLarge Language ModelText
🎯 What it does: The GainLoRA method is proposed, which integrates low-rank adaptation branches with a gating mechanism in continual learning to reduce forgetting.
Gatekeeper: Improving Model Cascades Through Confidence Tuning
Stephan Rabanser (Princeton University), Federico Tombari (Google)
ClassificationOptimizationVision Language ModelImageText
🎯 What it does: A hybrid loss function named GATEKEEPER is proposed for confidence tuning of small models in model cascades, ensuring high confidence during correct predictions and tending towards a uniform distribution during incorrect predictions, thereby achieving more reliable inference routing and deferral.
GauSAM: Contour‑Guided 2D Gaussian Fields for Multi‑Scale Medical Image Segmentation with Segment Anything
Jinxuan Wu (Tongji University), Dongdong Zhang (Tongji University)
SegmentationGaussian SplattingImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A Contour-Guided Continuous Gaussian Field framework (GauSAM) based on SAM has been developed, capable of generating high-quality medical image segmentation masks at any resolution.
Gaussian Approximation and Concentration of Constant Learning-Rate Stochastic Gradient Descent
Ziyang Wei (University of Chicago), Wei Biao Wu (University of Chicago)
OptimizationTabular
🎯 What it does: This paper addresses the distribution characteristics of stochastic gradient descent (SGD) with a fixed learning rate under finite sample conditions, proposing a new theoretical framework that proves convergence under arbitrary initialization, the central limit theorem, the Berry-Esseen bound, and Nagaev-type tail probability upper bounds.
Gaussian Herding across Pens: An Optimal Transport Perspective on Global Gaussian Reduction for 3DGS
Tao Wang (Renmin University of China), Qiong Zhang (Renmin University of China)
CompressionOptimizationGaussian SplattingPoint Cloud
🎯 What it does: This paper proposes a Gaussian Herder Across Pens (GHAP) framework based on optimal transport, which first compresses the geometric information of the 3D Gaussian Splatting (3DGS) model using Gaussian Mixture Reduction (GMR), and then fine-tunes the color and opacity, achieving nearly lossless rendering quality while retaining only 10% of the Gaussian primitives.
Gaussian Process Upper Confidence Bound Achieves Nearly-Optimal Regret in Noise-Free Gaussian Process Bandits
Shogo Iwazaki (LY Corporation)
Reinforcement Learning
🎯 What it does: This study investigates the noise-free Gaussian Process Bandit problem and proves that the GP-UCB algorithm can achieve nearly optimal cumulative and simple reward upper bounds under this setting.
Gaussian Processes for Shuffled Regression
Masahiro Kohjima (NTT Inc)
Tabular
🎯 What it does: This paper proposes Gaussian Process Shuffled Regression (GPSR) and its sparse approximate version SS-GPSR, aimed at learning regression functions and quantifying uncertainty from shuffled corresponding regression data.
Gaussian Regression-Driven Tensorized Incomplete Multi-View Clustering with Dual Manifold Regularization
Zhenhao Zhong (Hebei Normal University), Ruiqiang Guo (Hebei Normal University)
Gaussian SplattingMultimodality
🎯 What it does: A tensor-based incomplete multi-view clustering framework GUITAR is proposed, which is based on Gaussian regression norm, improved ℓδ norm, and double manifold regularization.
Gaussian-Augmented Physics Simulation and System Identification with Complex Colliders
Federico Vasile (Istituto Italiano di Tecnologia), Xiaolong Wang (UC San Diego)
Neural Radiance FieldGaussian SplattingVideoPhysics Related
🎯 What it does: A differentiable material point method (AS-DiffMPM) is proposed, supporting arbitrary shape rigid body collisions and enabling systematic identification of continuum physical properties through video observation.
GaussianFusion: Gaussian-Based Multi-Sensor Fusion for End-to-End Autonomous Driving
Shuai Liu (Sun Yat-sen University), Kai Huang (Sun Yat-sen University)
Autonomous DrivingOptimizationExplainability and InterpretabilityComputational EfficiencyGaussian SplattingPoint CloudBenchmark
🎯 What it does: A multi-sensor fusion framework called GaussianFusion based on 2D Gaussian distribution is proposed for perception and path planning in end-to-end autonomous driving.
Gaze Beyond the Frame: Forecasting Egocentric 3D Visual Span
Heeseung Yun (Seoul National University), Gunhee Kim (Seoul National University)
Object DetectionSegmentationTransformerSimultaneous Localization and MappingMultimodalityPoint Cloud
🎯 What it does: This paper proposes a 3D visual focus prediction method based on a first-person perspective, capable of inferring the user's visual attention area in three-dimensional space in real-time based on observations from the past few seconds.
Gaze-VLM: Bridging Gaze and VLMs through Attention Regularization for Egocentric Understanding
Anupam Pani (Hong Kong University), Yanchao Yang (Hong Kong University)
RecognitionGenerationTransformerVision Language ModelOptical FlowVideoMultimodality
🎯 What it does: This paper proposes using human gaze as an attention regularization signal during the training phase of VLM to enhance activity understanding and future action prediction in first-person videos.
GD$^2$: Robust Graph Learning under Label Noise via Dual-View Prediction Discrepancy
Kailai Li (Shanghai Jiao Tong University), Jie LI
Anomaly DetectionOptimizationGraph Neural NetworkSupervised Fine-TuningGraph
🎯 What it does: The GD2 framework is proposed, which detects label noise and achieves robust graph learning by utilizing the prediction differences between the node's own feature view and the neighborhood aggregation view.
GeGS-PCR: Fast and Robust Color 3D Point Cloud Registration with Two-Stage Geometric-3DGS Fusion
Jiayi Tian (Xi'an Jiaotong University), Pengju Ren (Xi'an Jiaotong University)
RecognitionOptimizationGaussian SplattingPoint Cloud
🎯 What it does: A two-stage color point cloud registration method, GeGS-PCR, is proposed, achieving high-precision registration by utilizing geometric, color, and Gaussian distribution information.
GEM: Empowering MLLM for Grounded ECG Understanding with Time Series and Images
Xiang Lan (National University of Singapore), Mengling Feng (National University of Singapore)
Anomaly DetectionTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodalityTime SeriesBiomedical DataElectrocardiogramBenchmark
🎯 What it does: Proposes GEM, the first multimodal large language model that combines ECG time series, 12-lead images, and text for evidence-based electrocardiogram interpretation.
Gemstones: A Model Suite for Multi-Faceted Scaling Laws
Sean Michael McLeish (University of Maryland), Tom Goldstein (Columbia University)
TransformerText
🎯 What it does: This paper constructs the Gemstones dataset, providing over 4,000 Transformer checkpoints with different hyperparameters such as width, depth, and learning rate, systematically studying the impact of model design on scaling laws.
GenColor: Generative and Expressive Color Enhancement with Pixel-Perfect Texture Preservation
Yi Dong (Nanyang Technological University), Rynson W. H. Lau (City University of Hong Kong)
Image TranslationGenerationConvolutional Neural NetworkTransformerDiffusion modelImage
🎯 What it does: The paper proposes a three-stage color enhancement framework named GenColor, which views color enhancement as a texture-preserving conditional generation task, utilizing diffusion models and texture-preserving networks to achieve professional-grade fine-grained color adjustments.
Gene Regulatory Network Inference in the Presence of Selection Bias and Latent Confounders
Gongxu Luo (Mohamed bin Zayed University of Artificial Intelligence), Kun Zhang (Mohamed bin Zayed University of Artificial Intelligence)
Biomedical Data
🎯 What it does: A gene regulatory network inference method that considers selection bias and potential confounding is proposed;
GeneFlow: Translation of Single-cell Gene Expression to Histopathological Images via Rectified Flow
Mengbo Wang (Purdue University), Nadia Atallah Lanman (Purdue University)
Image TranslationGenerationData SynthesisDiffusion modelRectified FlowImageBiomedical Data
🎯 What it does: Generating high-resolution tissue pathology images from spatial transcriptomics data
GeneMAN: Generalizable Single-Image 3D Human Reconstruction from Multi-Source Human Data
Wentao Wang (Shanghai AI Laboratory), Liang Pan (Shanghai AI Laboratory)
RestorationGenerationPose EstimationDiffusion modelNeural Radiance FieldImageVideo
🎯 What it does: The GeneMAN framework is proposed, utilizing a dedicated 2D/3D diffusion prior trained on multi-source human data to achieve high-quality 3D human reconstruction from a single natural image, supporting multiple poses, clothing styles, and different body types without relying on parameterized templates like SMPL.
General-Reasoner: Advancing LLM Reasoning Across All Domains
Xueguang Ma (University of Waterloo), Wenhu Chen (University of Waterloo)
OptimizationTransformerLarge Language ModelReinforcement LearningTextBenchmarkChain-of-Thought
🎯 What it does: A General-Reasoner framework is proposed, which directly trains large language models on verifiable reasoning data across multiple domains using Zero-RL, enhancing their cross-domain reasoning capabilities.
Generalizable Domain Adaptation for Sim-and-Real Policy Co-Training
Shuo Cheng (Georgia Institute of Technology), Danfei Xu (Georgia Institute of Technology)
Domain AdaptationRobotic IntelligencePoint Cloud
🎯 What it does: A unified simulation-real co-training framework is proposed, which learns generalizable robotic manipulation strategies using a large number of simulated demonstrations and a small number of real demonstrations.
Generalizable Hand-Object Modeling from Monocular RGB Images via 3D Gaussians
Xingyu Liu (Beijing University of Posts and Telecommunications), Jingyu Wang (Beijing University of Posts and Telecommunications)
GenerationPose EstimationTransformerGaussian SplattingImagePoint Cloud
🎯 What it does: The HOGS framework is proposed, which generates a general hand-object interaction model from monocular RGB images using 3D Gaussian Splatting.
Generalizable Insights for Graph Transformers in Theory and Practice
Timo Stoll (RWTH Aachen University), Christopher Morris (RWTH Aachen University)
Object DetectionOptimizationGraph Neural NetworkTransformerGraph
🎯 What it does: This paper proposes the Generalized-Distance Transformer (GDT), a universal graph transformer system based on the standard Transformer, and conducts a systematic evaluation on large-scale datasets across multiple domains.
Generalizable Reasoning through Compositional Energy Minimization
Alexandru Oarga (University of Barcelona), Yilun Du (Harvard University)
OptimizationGraph Neural NetworkContrastive LearningGraphBenchmark
🎯 What it does: A general reasoning framework is proposed through energy minimization, which constructs the energy function of a larger problem using the energy landscape of subproblems, and achieves generalization and optimization of reasoning tasks through parallel energy minimization (PEM) sampling.
Generalizable, real-time neural decoding with hybrid state-space models
Avery Hee-Woon Ryoo (Mila Quebec AI Institute), Guillaume Lajoie (Mila Quebec AI Institute)
Recurrent Neural NetworkTransformerTime SeriesSequential
🎯 What it does: A new hybrid structured neural decoder, POSSM, is proposed, which combines single-cell synaptic labeling with a cross-attention module integrated into a state space model (SSM) backbone, achieving millisecond-level real-time predictions.
Generalization Bound of Gradient Flow through Training Trajectory and Data-dependent Kernel
Yilan Chen (University of California San Diego), Arya Mazumdar (University of California San Diego)
OptimizationConvolutional Neural NetworkImageOrdinary Differential Equation
🎯 What it does: This paper presents an upper bound on the generalization error of gradient flow by combining the dynamics of gradient flow with the loss path kernel (LPK).
Generalization Bounds for Kolmogorov-Arnold Networks (KANs) and Enhanced KANs with Lower Lipschitz Complexity
Pengqi Li (Beijing Institute of Technology), Ye Yuan (Beijing Institute of Technology)
ClassificationOptimizationImageTextMultimodality
🎯 What it does: This paper studies the generalization mechanism of Kolmogorov-Arnold Networks (KAN), defining Lipschitz complexity for the first time as a measure of structural complexity of KAN, and deriving a generalization upper bound based on this; subsequently, it proposes the LipKAN architecture, which inserts Lip layers between each activation layer and employs L1^5 regularization, significantly reducing Lipschitz complexity and thereby enhancing the model's generalization performance.
Generalization Bounds for Model-based Algorithm Configuration
Zhiyang Chen (Tsinghua University), Xia Yin (Tsinghua University)
OptimizationHyperparameter Search
🎯 What it does: This paper derives the generalization error upper bound for model-based algorithm configurators (especially those using random forests as surrogate models) through the framework of algorithm stability in statistical learning theory.
Generalization Bounds for Rank-sparse Neural Networks
Antoine Ledent (Singapore Management University), Yunwen Lei (University of Hong Kong)
Convolutional Neural NetworkImage
🎯 What it does: This paper derives the generalization error upper bounds for neural networks (linear networks, fully connected networks, and convolutional networks) with approximately low-rank weight matrices.
Generalization Error Analysis for Selective State-Space Models Through the Lens of Attention
Arya Honarpisheh (Northeastern University), Mario Sznaier (Northeastern University)
Recurrent Neural NetworkTransformerTextSequential
🎯 What it does: This study investigates the generalization error of Selective State Space Models (Selective SSM) in sequence modeling and provides a generalization bound based on covering numbers.
Generalization Guarantees for Learning Score-Based Branch-and-Cut Policies in Integer Programming
Hongyu Cheng (Johns Hopkins University), Amitabh Basu (Johns Hopkins University)
OptimizationScore-based Model
🎯 What it does: This paper constructs a theoretical framework that proves when the scoring function of branch-and-cut (B&C) decisions has a piecewise polynomial structure, the overall performance metrics (such as tree size) are piecewise constant with respect to the parameters, and provides upper bounds for pseudo-dimension and sample complexity.
Generalization or Hallucination? Understanding Out-of-Context Reasoning in Transformers
Yixiao Huang (University of California Berkeley), Song Mei (University of California Berkeley)
TransformerLarge Language ModelText
🎯 What it does: This study investigates the reasons behind the ability to generalize while also producing hallucinations when injecting new factual knowledge into large language models (LLMs), proposing that both phenomena stem from the same mechanism—cross-context reasoning (OCR).
Generalization vs Specialization under Concept Shift
Alex Nguyen (Princeton University), Vudtiwat Ngampruetikorn (University of Sydney)
TransformerImageTabular
🎯 What it does: This study investigates ridge regression under conditions of concept shift, deriving an analytical expression for predictive risk in the high-dimensional limit and revealing the phase transition and non-monotonic behavior of risk between weak and strong concept shifts.
Generalized and Invariant Single-Neuron In-Vivo Activity Representation Learning
Wei Wu (Peking University), Jinzhuo Wang (Peking University)
Representation LearningAdversarial AttackTransformerAuto EncoderContrastive LearningBiomedical Data
🎯 What it does: This paper proposes a general adversarial training framework to eliminate batch effects in single-cell activity characterization, thereby enhancing the generalization ability of single-cell characterization models under different animals and stimulation conditions.
Generalized Category Discovery under Domain Shift: A Frequency Domain Perspective
Wei Feng (Monash University), Zongyuan Ge (Monash University)
ClassificationDomain AdaptationContrastive LearningImage
🎯 What it does: A method for General Category Discovery under Domain Shift (DS_GCD) called FREE is proposed, which enhances the model's clustering ability on unknown domains and unknown categories through frequency domain information.
Generalized Contrastive Learning for Universal Multimodal Retrieval
Jungsoo Lee (Qualcomm AI Research), Sungha Choi (Qualcomm AI Research)
RetrievalContrastive LearningImageTextMultimodality
🎯 What it does: A General Contrastive Learning (GCL) loss is proposed, utilizing existing image-text paired data to perform contrastive learning on image, text, and fused image-text embeddings within the same mini-batch, thereby enhancing the performance of multimodal retrieval models across different modality combinations.
Generalized Gradient Norm Clipping & Non-Euclidean $(L_0,L_1)$-Smoothness
Thomas Pethick (École Polytechnique Fédérale de Lausanne), Volkan Cevher (École Polytechnique Fédérale de Lausanne)
OptimizationImageText
🎯 What it does: This paper proposes a non-Euclidean gradient norm clipping method that combines conditional gradient and steepest descent, and proves its descent property under (L0, L1)-smoothness.
Generalized Linear Bandits: Almost Optimal Regret with One-Pass Update
Yu-Jie Zhang (RIKEN AIP), Masashi Sugiyama (University of Tokyo)
OptimizationReinforcement LearningTabular
🎯 What it does: This paper studies the Generalized Linear Bandit (GLB) problem and proposes an efficient algorithm that can achieve nearly optimal regret bounds with O(1) time and space complexity per round.
Generalized Linear Mode Connectivity for Transformers
Alexander Theus (ETH Zurich), Valentina Boeva (ETH Zurich)
TransformerImageText
🎯 What it does: A unified framework for network symmetry is proposed, and this framework is used to achieve linear mode connections between Transformer models.
Generalized Top-k Mallows Model for Ranked Choices
Shahrzad Haddadan (Rutgers Business School), Sara Ahmadian (Google)
Recommendation SystemOptimizationTabular
🎯 What it does: A weighted TopKGMM (Generalized Top-k Mallows Model) is proposed, along with three efficient algorithms: Profile-Based Repeated Insertion Sampling (PRIM), a dynamic programming method for calculating selection probabilities called DYPCHIP, and an active learning center approach named BUCCHOI.
Generalizing Experience for Language Agents with Hierarchical MetaFlows
Shengda Fan (Renmin University of China), Yankai Lin (Renmin University of China)
Computational EfficiencyMeta LearningTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This paper proposes the MetaFlowLLM framework, which constructs an experience hierarchy tree to enable large language model agents to reuse experiences in multi-step tasks through MetaFlow (including static steps and dynamic subtasks), significantly improving task success rates and execution efficiency.
Generalizing Single-Frame Supervision to Event-Level Understanding for Video Anomaly Detection
Junxi Chen (University of Chinese Academy of Sciences), Qingming Huang (University of Chinese Academy of Sciences)
Anomaly DetectionTransformerVideo
🎯 What it does: This paper proposes a Single Frame Supervised Video Anomaly Detection (SF-VAD) paradigm and designs a Frame-guided Progressive Learning (FPL) framework, utilizing only one frame annotation per anomalous video to achieve event-level anomaly understanding.
Generalizing while preserving monotonicity in comparison-based preference learning models
Julien Fageot (Tournesol), Lê-Nguyên Hoang (Tournesol)
Recommendation SystemDiffusion modelVideo
🎯 What it does: A new class of linear Generalized Bradley-Terry models (Linear GBT with Diffusion Prior) is proposed, which retains comparative data while utilizing embedding information to generalize to uncomparable objects, and provides monotonicity guarantees under specific embeddings (such as diffusion embedding and one-hot encoding).
Generate, but Verify: Reducing Hallucination in Vision-Language Models with Retrospective Resampling
Tsung-Han Wu (University of California Berkeley), David M. Chan (University of California Berkeley)
RecognitionGenerationTransformerVision Language ModelMultimodality
🎯 What it does: This paper proposes the REVERSE framework, which achieves the detection and self-correction of visual hallucinations by combining self-supervised training and retrospective resampling in a vision-language model.
Generating and Checking DNN Verification Proofs
Hai Duong (George Mason University), Matthew B. Dwyer (University of Virginia)
Convolutional Neural NetworkReinforcement LearningBenchmark
🎯 What it does: This paper proposes a proof format APTP and a lightweight proof checker APTPchecker that are independent of existing DNN verification tools, capable of independently verifying UNSAT proofs provided by DNN verification tools and achieving scalable checking on large-scale models.
Generating Computational Cognitive models using Large Language Models
Milena Rmus (Helmholtz Munich), Eric Schulz (Helmholtz Munich)
OptimizationTransformerLarge Language ModelText
🎯 What it does: Developed the GeCCo pipeline, utilizing open-source LLMs to generate cognitive models and optimize them through iterative feedback, applied in four cognitive domains: decision-making, learning, planning, and working memory;
Generating Creative Chess Puzzles
Xidong Feng (Google DeepMind), Tom Zahavy (Google DeepMind)
GenerationTransformerReinforcement LearningDiffusion modelSequential
🎯 What it does: This paper studies the use of generative models and reinforcement learning to automatically create creative, counterintuitive, and aesthetically valuable chess puzzles.
Generating Full-field Evolution of Physical Dynamics from Irregular Sparse Observations
Panqi Chen (Zhejiang University), Shikai Fang (Microsoft Research Asia)
GenerationData SynthesisDiffusion modelTime SeriesSequentialPhysics Related
🎯 What it does: A generative model named SDIFT is proposed, capable of recovering the complete evolution of continuous spatiotemporal physical fields from discrete sparse observations.
Generating Informative Samples for Risk-Averse Fine-Tuning of Downstream Tasks
Heasung Kim (University of Texas at Austin), Gustavo De Veciana
OptimizationDiffusion modelScore-based ModelTabular
🎯 What it does: Proposes a method that uses loss information from a pre-trained model to guide a score-based generative model in generating high-loss samples, thereby performing risk-averse CVaR optimization in downstream tasks;
Generating Multi-Table Time Series EHR from Latent Space with Minimal Preprocessing
Eunbyeol Cho (Korea Advanced Institute of Science and Technology), Edward Choi (Korea Advanced Institute of Science and Technology)
GenerationData SynthesisAnomaly DetectionTransformerAuto EncoderTabularTime SeriesBiomedical DataElectronic Health Records
🎯 What it does: This paper proposes the RawMed framework, which achieves lossless synthesis of multi-table time series electronic health records.
Generating Physically Sound Designs from Text and a Set of Physical Constraints
Gregory Barber (DEVCOM Army Research Laboratory), Mulugeta A Haile
GenerationOptimizationVision Language ModelImageTextPhysics Related
🎯 What it does: The TIDES method is proposed, which combines the pre-trained text-image model CLIP with differentiable finite element simulation to jointly optimize structure (topology) and visual attributes, generating designs from text prompts that satisfy both physical constraints and visual features.
Generation as Search Operator for Test-Time Scaling of Diffusion-based Combinatorial Optimization
Yang Li (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
GenerationOptimizationTransformerDiffusion modelRectified FlowGraphOrdinary Differential Equation
🎯 What it does: This paper proposes GenSCO, a framework that views the generation process of diffusion models as a search step, achieving efficient solution through iterative perturbation-enhancement-post-processing for combinatorial optimization.
Generative Caching for Structurally Similar Prompts and Responses
Sarthak Chakraborty (University of Illinois at Urbana-Champaign), Indranil Gupta (University of Illinois at Urbana-Champaign)
GenerationData SynthesisOptimizationTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: A new caching technology called GenCache is proposed, which utilizes large language models to automatically identify patterns between structurally similar prompts and responses, generating executable programs that can dynamically produce differentiated responses based on new prompts when a cache hit occurs.
Generative Data Augmentation via Diffusion Distillation, Adversarial Alignment, and Importance Reweighting
Ruyi An (University of Texas at Austin), Mingyuan Zhou (University of Texas at Austin)
GenerationData SynthesisDiffusion modelScore-based ModelGenerative Adversarial NetworkImage
🎯 What it does: A three-stage generative data augmentation framework called DAR-GDA is proposed, which first compresses a multi-step diffusion model into a single-step generator using score distillation, then aligns the real distribution through adversarial training, and finally performs importance reweighting using the probability output of the discriminator.
Generative diffusion for perceptron problems: statistical physics analysis and efficient algorithms
Davide Straziota (Bocconi University), Carlo Lucibello (Bocconi University)
GenerationOptimizationDiffusion modelPhysics Related
🎯 What it does: The feasibility of using Approximate Message Passing-based Generative Diffusion (ASL) for sampling the random perceptron problem (spherical and binary) in the high-dimensional limit is studied, and thresholds are provided through replica theory.
Generative Distribution Embeddings
Nic Fishman (Harvard University), Omar Abudayyeh (Mass General Brigham)
GenerationData SynthesisRepresentation LearningDrug DiscoveryAuto EncoderBiomedical Data
🎯 What it does: A general framework called Generative Distribution Embeddings (GDE) is proposed, which elevates autoencoders to distribution space. It learns distribution-level representations through distribution-invariant encoders and conditional generators, and is capable of distribution reconstruction and interpolation on large-scale biological data.
Generative Graph Pattern Machine
Zehong Wang (University of Notre Dame), Yanfang Ye (University of Notre Dame)
GenerationRepresentation LearningGraph Neural NetworkTransformerGraph
🎯 What it does: Designed and evaluated a fully Transformer-based, message-passing-free graph pre-training framework G2PM, which utilizes substructure sequences generated by random walks for masked substructure reconstruction to learn graph representations.
Generative Model Inversion Through the Lens of the Manifold Hypothesis
Xiong Peng (Hong Kong Baptist University), Mingyuan Zhou (University of Texas at Austin)
GenerationData SynthesisConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: This paper studies generative model inversion from a geometric perspective, analyzes the mechanism of gradient projection onto the generator manifold, and proposes gradient-manifold alignment metrics, gradient alignment training objectives, and the non-training AlignMI method.
Generative Modeling of Full-Atom Protein Conformations using Latent Diffusion on Graph Embeddings
Aditya Sengar (École Polytechnique Fédérale de Lausanne), Pierre Vandergheynst (École Polytechnique Fédérale de Lausanne)
GenerationData SynthesisProtein Structure PredictionGraph Neural NetworkDiffusion modelGraphBiomedical Data
🎯 What it does: A latent diffusion model based on graph embedding (LD-FPG) has been constructed, capable of directly generating a full-atom protein conformation set from molecular dynamics trajectories, including side-chain atoms.
Generative Perception of Shape and Material from Differential Motion
Xinran Han, Todd Zickler
GenerationData SynthesisTransformerDiffusion modelVideoMultimodality
🎯 What it does: A conditional video diffusion model is proposed, which jointly infers the shape (surface normal vectors) and material (diffuse reflection, roughness, metallicity, specular reflection) of objects in short videos based on their subtle motion differences, and generates multimodal samples.
Generative Pre-trained Autoregressive Diffusion Transformer
Yuan Zhang (Peking University), Nan Duan (StepFun)
GenerationData SynthesisTransformerDiffusion modelImageVideo
🎯 What it does: Proposes GPDiT, a continuous space long video generation framework that combines autoregressive diffusion models with Transformers;
Generative RLHF-V: Learning Principles from Multi-modal Human Preference
Jiayi Zhou (Peking University), Yaodong Yang (Peking University)
GenerationReinforcement Learning from Human FeedbackReinforcement LearningMultimodality
🎯 What it does: A two-stage alignment framework called Generative RLHF-V is proposed by combining generative reward models with multimodal RLHF.
Generative Trajectory Stitching through Diffusion Composition
Yunhao Luo (Georgia Institute of Technology), Danfei Xu (Georgia Institute of Technology)
GenerationRobotic IntelligenceReinforcement LearningDiffusion modelTime SeriesSequential
🎯 What it does: A generative trajectory stitching framework named CompDiffuser is proposed, which segments long trajectories into overlapping short segments and uses a single bidirectional diffusion model for conditional generation of the short segments, achieving unsupervised long-range trajectory planning.
Generator-Mediated Bandits: Thompson Sampling for GenAI-Powered Adaptive Interventions
Marc Brooks (University of Michigan), Ambuj Tewari (University of Michigan)
OptimizationLarge Language ModelReinforcement LearningText
🎯 What it does: A generator-mediated bandwidth model based on generative artificial intelligence and the corresponding GAMBITTS algorithm is proposed to learn randomly generated treatment responses and maximize rewards within a multi-armed bandit framework.
Genesis: Multimodal Driving Scene Generation with Spatio-Temporal and Cross-Modal Consistency
Xiangyu Guo (Huazhong University of Science and Technology), Xinggang Wang (Huazhong University of Science and Technology)
GenerationData SynthesisAutonomous DrivingTransformerVision Language ModelDiffusion modelNeural Radiance FieldAuto EncoderVideoMultimodalityPoint Cloud
🎯 What it does: We propose Genesis, a unified world model for simultaneously generating multi-view RGB videos and LiDAR sequences, achieving spatial-temporal and cross-modal consistency.
GenIR: Generative Visual Feedback for Mental Image Retrieval
Diji Yang (University of California Santa Cruz), James Davis (University of California Santa Cruz)
GenerationRetrievalVision Language ModelDiffusion modelImage
🎯 What it does: Designed and implemented an interactive visual retrieval framework GenIR, which utilizes a text-to-image diffusion generator to create synthetic images as visual feedback for users, thereby helping them approach their target images in multi-round retrieval; simultaneously defined the Mental Image Retrieval (MIR) task and proposed an automated pipeline for constructing multi-round datasets.