ICLR 2026 Papers — Page 20
International Conference on Learning Representations · 5356 papers
GenDR: Lighten Generative Detail Restoration
Yan Wang (ByteDance), Li zhang
Super ResolutionComputational EfficiencyDiffusion modelAuto EncoderImage
🎯 What it does: Proposed a single-step detail restoration model called GenDR for high-quality image super-resolution, improving the VAE structure of Stable Diffusion and designing a novel one-step diffusion inference pipeline.
GeneBreaker: Jailbreak Attacks against DNA Language Models with Pathogenicity Guidance
ZAIXI ZHANG, Mengdi Wang (Princeton University)
Adversarial AttackTransformerAgentic AIPrompt EngineeringBiomedical DataBenchmark
🎯 What it does: Studied the safety risks of DNA language models in generating human viruses, and proposed a Jailbreak-based attack framework called GeneBreaker along with the corresponding evaluation benchmark JailbreakDNABench.
General Exploratory Bonus for Optimistic Exploration in RLHF
Wendi Li (University of Wisconsin-Madison), Sharon Li (University of Wisconsin-Madison)
Reinforcement Learning from Human FeedbackReinforcement LearningText
🎯 What it does: This paper proposes a general exploratory bonus framework, General Exploratory Bonus (GEB), aiming to address the theoretical failure of exploration reward mechanisms based on KL and α-divergence regularization in RLHF.
General search techniques without common knowledge for imperfect-information games, and application to superhuman Fog of War chess
Brian Hu Zhang (Massachusetts Institute of Technology), Tuomas Sandholm (Carnegie Mellon University)
Reinforcement Learning
🎯 What it does: Developed Obscuro AI and achieved superhuman performance in Fog of War chess (a variant of Chinese chess with hidden pieces).
Generalised Flow Maps for Few-Step Generative Modelling on Riemannian Manifolds
Oscar Davis (University of Oxford), Joey Bose
GenerationData SynthesisKnowledge DistillationFlow-based ModelBiomedical Data
🎯 What it does: Proposed a Generalised Flow Maps (GFM) framework for few-step generation on arbitrary Riemannian manifolds.
Generalizable Coarse-to-Fine Robot Manipulation via Language-Aligned 3D Keypoints
Jianshu Hu (Shanghai Jiao Tong University), Yutong Ban (Shanghai Jiao Tong University)
Robotic IntelligenceTransformerSupervised Fine-TuningVision Language ModelImageTextChain-of-Thought
🎯 What it does: Proposed a novel coarse-to-fine 3D manipulation strategy called CLAP, which decomposes tasks using language instructions and combines pre-trained Vision-Language Models (VLM) with 3D perception to achieve stronger generalization capabilities.
Generalizable End-to-End Tool-Use RL with Synthetic CodeGym
Weihua Du (Carnegie Mellon University), Jiecao Chen (ByteDance)
AI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This paper proposes the CodeGym framework, which converts programming problems into interactive Gym environments to train LLMs for multi-round tool calls, enhancing their robustness in real-world workflows.
Generalizable Heuristic Generation Through LLMs with Meta-Optimization
Yiding Shi (Nanyang Technological University), Jie Zhang (Eindhoven University Of Technology)
OptimizationMeta LearningTransformerLarge Language ModelGraph
🎯 What it does: A meta-optimization framework is constructed by automatically generating and iteratively improving heuristics for combinatorial optimization problems using large language models (LLMs).
Generalization Below the Edge of Stability: The Role of Data Geometry
Tongtong Liang (University of California San Diego), Yu-Xiang Wang (University of California San Diego)
OptimizationRepresentation LearningImage
🎯 What it does: This paper theoretically studies the generalization behavior of overparameterized two-layer ReLU networks trained under edge stability (BEoS), proposes and proves a geometric measure called 'data shatterability,' provides upper and lower bounds for generalization under different data distributions (such as concentric Beta-radial distributions and low-dimensional spherical mixtures), and demonstrates how the model adaptively adjusts its intrinsic dimensionality based on data geometry.
Generalization in LLM Problem Solving: The Case of the Shortest Path
Yao Tong (National University of Singapore), Reza Shokri (National University of Singapore)
Data SynthesisTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningGraph
🎯 What it does: Built a controlled synthetic environment based on shortest path planning to evaluate the system generalization capability of language models in two dimensions: spatial transfer and length extension.
Generalization of Diffusion Models Arises with a Balanced Representation Space
Zekai Zhang (University of Michigan), Qing Qu (University of Michigan)
GenerationData SynthesisRepresentation LearningDiffusion modelAuto EncoderImage
🎯 What it does: Investigate the structural properties of the internal representation space in diffusion models, theoretically analyze how two-layer ReLU denoising autoencoders (DAE) lead to memorization and generalization under overparameterized and underparameterized conditions, respectively, and propose a memory detection method and a training-free editing approach based on representations.
Generalization of RLVR Using Causal Reasoning as a Testbed
Brian Lu (Johns Hopkins University), Hongyuan Mei (Toyota Technological Institute at Chicago)
Explainability and InterpretabilityData-Centric LearningReinforcement Learning from Human FeedbackLarge Language ModelSupervised Fine-TuningReinforcement LearningGraphBenchmark
🎯 What it does: This paper constructs the RLCausal dataset under the verifiable reward (RLVR) framework for causal inference tasks, and conducts fine-grained comparative experiments between RLVR and SFT on different scales of Qwen2.5-Instruct models, exploring the impact of model scale, query hierarchy, and query complexity on generalization ability.
Generalized Parallel Scaling with Interdependent Generations
Harry Dong (Meta), Karthik Abinav Sankararaman (Meta)
GenerationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This paper proposes Bridge, a lightweight architectural modification that enables multiple parallel generations from the same input to interact with each other during the decoding process, thereby improving the quality of both individual samples and the generated ensemble.
Generalized Spherical Neural Operators: Green’s Function Formulation
Hao Tang (University of Dundee), Chao Li (University of Dundee)
Convolutional Neural NetworkBiomedical DataFibre Orientation Distribution
🎯 What it does: This paper proposes a theoretical framework based on a designable spherical Green's function and develops the Green's-Function Spherical Neural Operator (GSNO) and multi-scale network SHNet to efficiently learn mappings for spherical PDEs.
Generalizing Linear Autoencoder Recommenders with Decoupled Expected Quadratic Loss
Ruixin Guo (Kent State University), Ruoming Jin (Kent State University)
Recommendation SystemAuto Encoder
🎯 What it does: This paper proposes extending the EDLAE objective to decoupled expected quadratic loss DEQL, providing a complete closed-form solution and expanding the hyperparameter range to b≥0, while achieving efficient O(n³) computation through Miller's matrix inverse theorem;
Generate Any Scene: Scene Graph Driven Data Synthesis for Visual Generation Training
Ziqi Gao (University of Washington), Ranjay Krishna (Allen Institute for Artificial Intelligence)
GenerationData SynthesisReinforcement Learning from Human FeedbackLarge Language ModelSupervised Fine-TuningVision Language ModelDiffusion modelImageText
🎯 What it does: Designed the GENERATE ANY SCENE data engine, which systematically enumerates scene graphs to generate diverse synthetic descriptions for four tasks: self-improvement of text-to-visual models, proprietary model distillation, low-cost RLHF reward models, and content moderation.
Generating Directed Graphs with Dual Attention and Asymmetric Encoding
Alba Carballo-Castro (EPFL), Pascal Frossard (EPFL)
GenerationGraph Neural NetworkTransformerDiffusion modelFlow-based ModelGraphBenchmark
🎯 What it does: Proposed and implemented DIRECTO, a directed graph generation framework based on discrete flow matching, combining bidirectional attention mechanisms, direction-aware position encoding, and discriminator-free conditional generation, while simultaneously releasing a standardized evaluation benchmark covering synthetic and real-world data.
Generating metamers of human scene understanding
Ritik Raina (Stony Brook University), Greg Zelinsky
GenerationData SynthesisTransformerDiffusion modelImage
🎯 What it does: Generate scene metamers based on human eye gaze information to probe the intrinsic representations of humans for complex scenes;
Generation then Reconstruction: Accelerating Masked Autoregressive Models via Two-Stage Sampling
Feihong Yan (Beijing Institute of Technology), Linfeng Zhang (Shanghai Jiao Tong University)
GenerationComputational EfficiencyTransformerDiffusion modelImageText
🎯 What it does: Propose a training-agnostic hierarchical sampling strategy called Generation then Reconstruction (GtR), which significantly accelerates Masked Autoregressive (MAR) image generation by first slowly generating dispersed global tokens and then rapidly reconstructing the remaining tokens.
Generative Adversarial Post-Training Mitigates Reward Hacking in Live Human-AI Music Interaction
Yusong Wu (Université de Montréal), Cheng-Zhi Anna Huang (Massachusetts Institute of Technology)
GenerationReinforcement Learning from Human FeedbackTransformerReinforcement LearningGenerative Adversarial NetworkContrastive LearningSequentialAudio
🎯 What it does: Propose Generative Adversarial Post-Training (GAPT), which suppresses reward hacking by incorporating a discriminator's adversarial reward in post-training of reinforcement learning, thereby improving harmony coherence and diversity in real-time music accompaniment.
Generative Adversarial Reasoner: Enhancing LLM Reasoning with Adversarial Reinforcement Learning
Qihao Liu (Johns Hopkins University), Alan Yuille (Johns Hopkins University)
Explainability and InterpretabilityTransformerLarge Language ModelReinforcement LearningGenerative Adversarial NetworkText
🎯 What it does: By jointly training the LLM reasoner and LLM discriminator, using slice-level alignment rewards to enhance mathematical reasoning performance.
Generative Bayesian Optimization: Generative Models as Acquisition Functions
Rafael Oliveira (CSIRO's Data61), Edwin V. Bonilla (CSIRO's Data61)
OptimizationTransformerReinforcement LearningGenerative Adversarial NetworkTextBiomedical Data
🎯 What it does: Proposes Generative Bayesian Optimization (GenBO), which directly trains a generative model to sample candidate points that satisfy the Bayesian optimization acquisition function, eliminating the need for intermediate regression/classification surrogates, and supports large-scale, high-dimensional, and combinatorial design.
Generative Blocks World: Moving Things Around in Pictures
Vaibhav Vavilala (University of Illinois Urbana Champaign), Anand Bhattad (Johns Hopkins University)
GenerationData SynthesisDepth EstimationConvolutional Neural NetworkDiffusion modelFlow-based ModelRectified FlowImagePoint CloudMesh
🎯 What it does: Propose a scene decomposition framework based on convex polyhedron primitives (Generative Blocks World), enabling 3D semantic image editing through primitive editing, and synthesizing new images using deep conditional generative models.
Generative Diffusion Prior Distillation for Long-Context Knowledge Transfer
Hewa Dehigahawattage Nilushika Udayangani (University of Melbourne), Marimuthu Palaniswami (University of Melbourne)
Knowledge DistillationDiffusion modelTime Series
🎯 What it does: Propose a method called GDPD that utilizes teacher knowledge priors generated by diffusion models for knowledge distillation, aiming to bridge the generalization gap from teacher models trained on full sequences to student models that only observe prefixes.
Generative Human Geometry Distribution
Xiangjun Tang (King Abdullah University of Science and Technology), Peter Wonka (King Abdullah University of Science and Technology)
GenerationData SynthesisFlow-based ModelAuto EncoderImageMesh
🎯 What it does: Propose a generative human geometry distribution model based on flow matching, using SMPL as the source distribution and encoding geometric information into 2D texture features, supporting random generation under pose conditions and pose variation generation.
Generative Modeling from Black-Box Corruptions via Self-Consistent Stochastic Interpolants
Chirag Modi (New York University), Joan Bruna (New York University)
GenerationData SynthesisDiffusion modelScore-based ModelImageTabularStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: A self-consistent stochastic interpolant (SCSI) framework is constructed to learn the original data distribution solely from corrupted observations and a black-box forward model.
Generative Universal Verifier as Multimodal Meta-Reasoner
Xinchen Zhang (Tsinghua University), Guang Shi (ByteDance Seed)
GenerationData SynthesisLarge Language ModelReinforcement LearningPrompt EngineeringVision Language ModelDiffusion modelImageTextMultimodalityBenchmark
🎯 What it does: Proposed the ViVerBench benchmark for evaluating visual result verification, constructed a scalable automated visual verification dataset, and trained a general-purpose generative visual verifier OmniVerifier-7B; subsequently, designed the sequential test-time scaling framework OmniVerifier-TTS using OmniVerifier to enhance image generation and editing quality in unified multi-modal models (UMM).
Generative Value Conflicts Reveal LLM Priorities
Andy Liu (Carnegie Mellon University), Max Kleiman-Weiner (University of Washington)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Proposes CONFLICTSCOPE, an automated pipeline that can automatically generate value conflict scenarios based on user-specified value sets, and extracts the value priorities of LLMs through open-ended user–model interactive evaluation.
Generative View Stitching
Chonghyuk Song (MIT CSAIL), Vincent Sitzmann (MIT CSAIL)
GenerationData SynthesisTransformerDiffusion modelVideo
🎯 What it does: Proposed a training-free, compatible Generative View Stitching (GVS) sampling method with existing video diffusion models to generate long videos satisfying predefined camera trajectories, avoiding collision and inconsistency issues from autoregressive sampling.
GenFusion: Feed-forward Human Performance Capture via Progressive Canonical Space Updates
YoungJoong Kwon, Ehsan Adeli (Stanford University)
GenerationPose EstimationConvolutional Neural NetworkDiffusion modelAuto EncoderImageVideo
🎯 What it does: Propose a feed-forward method to real-time capture human actions from monocular RGB streams, achieving novel view synthesis through a progressively updatable canonical space and probabilistic regression.
Genie Envisioner: A Unified World Foundation Platform for Robotic Manipulation
Yue Liao (Beihang University), Guanghui Ren (Beihang University)
Robotic IntelligenceTransformerVision Language ModelVision-Language-Action ModelDiffusion modelFlow-based ModelVideoTextMultimodality
🎯 What it does: Propose Genie Envisioner, a unified world foundation model that integrates multi-perspective video diffusion and parallel action decoding to achieve closed-loop generation and control for robotic manipulation.
GenSR: Symbolic regression based on equation generative space
Qian Li (Shanghai Jiao Tong University), Yuntian Chen (Imperial College London)
GenerationOptimizationTransformerAuto EncoderBenchmarkPhysics Related
🎯 What it does: Propose the GenSR framework, which maps the discrete symbolic equation space into a continuous generative latent space via a dual-branch Conditional Variational Autoencoder (CVAE), and performs symbolic regression by first conducting coarse localization and then fine search in this space.
GeoBench: Rethinking Multimodal Geometric Problem-Solving via Hierarchical Evaluation
Yuan Feng (Shanghai Jiao Tong University), Junchi Yan (Shanghai Artificial Intelligence Laboratory)
Large Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Constructed GeoBench, a hierarchical geometric reasoning evaluation benchmark generated by TrustGeoGen, covering six tasks ranging from visual perception to self-correction;
GeoDiv: Framework for Measuring Geographical Diversity in Text-to-Image Models
Abhipsa Basu (Indian Institute of Science), Venkatesh Babu Radhakrishnan
GenerationData SynthesisLarge Language ModelVision Language ModelImageText
🎯 What it does: Propose the GeoDiv framework to quantitatively evaluate the geographic diversity of images generated by text-to-image models
GeoFAR: Geography-Informed Frequency-Aware Super-Resolution for Climate Data
Chang Xu (EPFL), Devis Tuia (EPFL)
Super ResolutionConvolutional Neural NetworkImagePhysics Related
🎯 What it does: Proposes a climate super-resolution method called GeoFAR that simultaneously considers frequency features and geographic information.
GeoGramBench: Benchmarking the Geometric Program Reasoning in Modern LLMs
Shixian Luo (Li Auto Inc), Yong Wu (Li Auto Inc)
AI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Propose the GeoGramBench benchmark, using 500 programmatic geometry problems to evaluate LLMs' performance in the Program-to-Geometry task.
Geometric Autoencoder Priors for Bayesian Inversion: Learn First Observe Later
Arnaud Vadeboncoeur (University of Cambridge), Eleni Chatzi (ETH Zürich)
Representation LearningGraph Neural NetworkAuto EncoderGraphPhysics Related
🎯 What it does: Learn a prior distribution across geometric spaces using a geometric autoencoder without relying on PDEs, boundary conditions, or observation models, enabling full-field Bayesian inversion under sparse observations.
Geometric Constraints for Small Language Models to Understand and Expand Scientific Taxonomies
Liri Fang (University of Illinois Urbana Champaign), Vetle I Torvik (Meta Ai)
Knowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Proposed the SS-MONO framework for automatic expansion in scientific taxonomy, integrating LLM local increment, SLM self-supervised fine-tuning (with hyperbolic geometry constraints), and LLM calibration to achieve insertion from root to leaf nodes.
Geometric Graph Neural Diffusion for Stable Molecular Dynamics Simulations
Haokai Hong (Hong Kong Polytechnic University), KC Tan
Drug DiscoveryGraph Neural NetworkDiffusion modelGraph
🎯 What it does: Propose the Geometric Graph Neural Diffusion (GGND) framework, which can be seamlessly integrated as a plugin into existing local equivariant message passing networks, significantly enhancing the energy and force prediction accuracy and trajectory stability of molecular dynamics (MD) in unseen conformations.
Geometric Image Editing via Effects-Sensitive In-Context Inpainting with Diffusion Transformers
Shuo Zhang (School of Artificial Intelligence Beijing University of Posts and Telecommunications), Zhanyu Ma (School of Artificial Intelligence Beijing University of Posts and Telecommunications)
GenerationTransformerDiffusion modelImage
🎯 What it does: Propose GeoEdit, a geometry image editing framework based on diffusion transformers, capable of precisely performing object translation, rotation, scaling, and generating realistic lighting and shadow effects.
Geometric-Mean Policy Optimization
Yuzhong Zhao (University of Chinese Academy of Sciences), Furu Wei (Microsoft Research)
OptimizationLarge Language ModelReinforcement LearningMixture of ExpertsTextMultimodalityBenchmark
🎯 What it does: Propose the GMPO (Geometric Mean Policy Optimization) algorithm, replacing the arithmetic mean reward in GRPO with geometric mean to enhance the training stability and performance of large language models in reasoning tasks.
Geometry Forcing: Marrying Video Diffusion and 3D Representation for Consistent World Modeling
Haoyu Wu (Microsoft Research), Jiang Bian (Tsinghua University)
GenerationData SynthesisRepresentation LearningTransformerDiffusion modelFlow-based ModelWorld ModelVideo
🎯 What it does: Propose the Geometry Forcing method, which enhances the 3D consistency and visual quality of video generation by aligning the intermediate features of video diffusion models with a 3D base model.
Geometry of Uncertainty: Learning Metric Spaces for Multimodal State Estimation in RL
Alfredo Reichlin (KTH Royal Institute of Technology), Miguel Vasco (KTH Royal Institute of Technology)
Representation LearningReinforcement LearningContrastive LearningImageMultimodalityPoint CloudAudio
🎯 What it does: Propose and train a multi-modal state estimation model (METRICMM), which achieves robust state estimation against observation noise by learning a geometric metric space where the Euclidean distance between latent representations matches the minimal number of actions required for the environment state, and uses this representation for reinforcement learning strategies.
Geometry-aware 4D Video Generation for Robot Manipulation
Zeyi Liu (Stanford University), Shuran Song (Stanford University)
GenerationRobotic IntelligenceTransformerDiffusion modelAuto EncoderVideo
🎯 What it does: This paper proposes a geometry-aware 4D video generation model for robot manipulation tasks;
Geometry-aware Policy Imitation
Yiming Li (Idiap Research Institute), Sylvain Calinon (Idiap Research Institute)
Robotic IntelligenceReinforcement Learning from Human FeedbackConvolutional Neural NetworkTransformerDiffusion modelFlow-based ModelAuto EncoderImageMultimodality
🎯 What it does: Propose Geometry-aware Policy Imitation (GPI), treating demonstrations as geometric curves to generate distance fields and flow fields, achieving compositional multi-modal control in a non-parametric manner;
GeomMotif: A Benchmark for Arbitrary Geometric Preservation in Protein Generation
Pavel Strashnov (AXXX, Center for AI-driven Drug Design), Dmitry Vetrov (Constructor University)
Protein Structure PredictionBiomedical DataBenchmark
🎯 What it does: Introduces the GeomMotif benchmark to systematically evaluate the ability of protein generation models to maintain arbitrary geometry while completely removing functional constraints.
GeoPurify: A Data-Efficient Geometric Distillation Framework for Open-Vocabulary 3D Segmentation
Weijia Dou (Tongji University), Heng Tao Shen (Tongji University)
SegmentationKnowledge DistillationConvolutional Neural NetworkVision Language ModelContrastive LearningPoint Cloud
🎯 What it does: This paper proposes GeoPurify, a data-efficient geometric distillation framework for open-vocabulary 3D semantic segmentation.
GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning
Lakshya A Agrawal (UC Berkeley), Omar Khattab (MIT)
OptimizationTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Propose GEPA, a genetic prompt optimizer that integrates natural language reflection and Pareto front sampling, significantly improving the performance of LLMs in multi-task scenarios.
GEPO: Group Expectation Policy Optimization for Stable Heterogeneous Reinforcement Learning
Han Zhang (Pengcheng Laboratory), Yue Yu (Pengcheng Laboratory)
Large Language ModelReinforcement LearningText
🎯 What it does: Propose the HeteroRL framework and GEPO algorithm to achieve reinforcement learning for large-scale LLMs in high-latency, heterogeneous networks by decoupling sampling and learning;
Get RICH or Die Scaling: Profitably Trading Inference Compute for Robustness
Tavish Malcolm McDonald (Lawrence Livermore National Laboratory), Brian R. Bartoldson (Lawrence Livermore National Laboratory)
Computational EfficiencyAdversarial AttackPrompt EngineeringVision Language ModelImageMultimodalityChain-of-Thought
🎯 What it does: This paper proposes the 'Robustness from Inference Compute Hypothesis (RICH)', systematically evaluating how combining inference-time computation (e.g., reasoning chains, prompt iteration) with a model's inherent robustness enhances resistance to adversarial attacks across vision-language models (VLMs) of varying scales and training intensities.
Getting Your LLMs Ready for Reinforcement Learning with Lightweight SFT
Xinran Li (Hong Kong University of Science and Technology), Jun Zhang (Hong Kong University of Science and Technology)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Study how to use lightweight self-supervised fine-tuning during the cold start phase before reinforcement learning to optimize post-training of large language models, propose Adaptive Early-Stop Loss (AESL) and prove its effectiveness.
GGBall: Graph Generative Model on Poincaré Ball
Tianci Bu (Westlake University), Tailin Wu (Westlake University)
GenerationData SynthesisGraph Neural NetworkTransformerDiffusion modelFlow-based ModelAuto EncoderGraph
🎯 What it does: Proposed a fully Poincaré ball (negative curvature space)-based graph generation framework called GGBall, which utilizes hypergraph space to encode and decode hierarchical structures of graphs;
Ghost in the Cloud: Your Geo-Distributed Large Language Models Training is Easily Manipulated
Zichen TANG, Bo Li (Hong Kong University of Science and Technology)
Federated LearningSafty and PrivacyAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
🎯 What it does: Under geographically distributed training and federated learning scenarios, two attack methods were designed and implemented: Trigger-based Pseudo-Contrastive Safety Alignment (TPCSA) and Downstream-Preserved Malicious Training (DPT), enabling malicious clients to inject induced triggers into global large language models without being detected by server-side security detection (MOS and TPC), leading to the generation of harmful content when triggered.
GHOST: Hallucination-Inducing Image Generation for Multimodal LLMs
Aryan Yazdan Parast (University of Melbourne), NAVEED AKHTAR
GenerationData SynthesisAdversarial AttackTransformerLarge Language ModelDiffusion modelGenerative Adversarial NetworkContrastive LearningImageMultimodality
🎯 What it does: Propose the GHOST method, which automatically generates images to induce object hallucinations in multimodal large language models (MLLMs).
GhostEI-Bench: Do Mobile Agent Resilience to Environmental Injection in Dynamic On-Device Environments?
Chiyu Chen (Shanghai Jiao Tong University), Yingchun Wang (Shanghai Artificial Intelligence Laboratory)
Safty and PrivacyLarge Language ModelAgentic AIVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Constructed GhostEI-Bench to systematically evaluate the robustness of mobile vision-language models (VLMs) against environment injection attacks in dynamic device environments;
GIQ: Benchmarking 3D Geometric Reasoning of Vision Foundation Models with Simulated and Real Polyhedra
Mateusz Michalkiewicz (Rice University), Guha Balakrishnan (Rice University)
TransformerVision Language ModelImageBenchmark
🎯 What it does: This paper creates and uses the GIQ benchmark to systematically evaluate the 3D geometric reasoning capabilities of visual and vision-language foundation models.
GIR-Bench: Versatile Benchmark for Generating Images with Reasoning
Hongxiang Li (Hong Kong University of Science and Technology), Long Chen (Hong Kong University of Science and Technology)
GenerationLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Designed and released GIR-Bench, a multimodal benchmark focused on reasoning, to evaluate the reasoning alignment capabilities of unified models in understanding, generation, and editing tasks.
Gistify: Codebase-Level Understanding via Runtime Execution
Hyunji Lee (University of North Carolina at Chapel Hill), Lucas Caccia (Cornell University)
AI Code AssistantTransformerLarge Language ModelAgentic AITextBenchmark
🎯 What it does: Proposed the GISTIFY task, which requires the code LLM to generate a single-file self-contained code based on a given codebase and entry command, capable of fully reproducing the execution results of the command in the original codebase;
GIT-BO: High-Dimensional Bayesian Optimization with Tabular Foundation Models
Rosen Ting-Ying Yu (Massachusetts Institute of Technology), Faez Ahmed (Massachusetts Institute of Technology)
OptimizationTransformerTabular
🎯 What it does: Designed and implemented a gradient information guided Bayesian optimization framework (GIT-BO) based on TabPFN v2, utilizing a frozen Tabular Foundation Model for zero-shot inference, combined with gradient activation subspace and UCB acquisition function to achieve high-dimensional (up to 500 dimensions) black-box optimization tasks without online retraining.
Glance and Focus Reinforcement for Pan-cancer Screening
Linshan Wu (Hong Kong University of Science and Technology), Hao Chen (Hong Kong University of Science and Technology)
Object DetectionSegmentationAnomaly DetectionConvolutional Neural NetworkTransformerReinforcement LearningImageBiomedical DataComputed Tomography
🎯 What it does: Proposed GF-Screen, a reinforcement learning framework combining Glance and Focus, achieving multi-cancer detection and segmentation in large-scale CT scans through a lightweight coarse localization model and a fine segmentation model.
Glance for Context: Learning When to Leverage LLMs for Node-Aware GNN-LLM Fusion
Donald Loveland (University of Michigan), Danai Koutra (University of Michigan)
ClassificationRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelReinforcement LearningTextGraph
🎯 What it does: This paper proposes the GLANCE framework, which leverages node-level routers to invoke large language models (LLMs) only when graph neural networks (GNNs) underperform, enhancing node classification performance on text-attribute graphs (TAGs).
GLASS Flows: Efficient Inference for Reward Alignment of Flow and Diffusion Models
Peter Holderrieth (MIT), Brian Karrer (Meta)
GenerationReinforcement Learning from Human FeedbackDiffusion modelFlow-based ModelImageTextStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Propose GLASS Flows, a method that constructs an internal flow matching model within pre-trained flow matching/diffusion models, efficiently sampling Markov transitions via ODEs to achieve untrained random transition sampling;
Global and Local Topology-Aware Graph Generation via Dual Conditioning Diffusion
Yuhang Xie (Chinese University of Hong Kong), Sinno Jialin Pan (Chinese University of Hong Kong)
GenerationGraph Neural NetworkDiffusion modelScore-based ModelAuto EncoderGraphBiomedical Data
🎯 What it does: Propose DualDiff, a dual-branch latent diffusion model with a dual conditional mechanism that can simultaneously model the global and local topological information of graphs, achieving high-quality graph generation.
Global Resolution: Optimal Multi-Draft Speculative Sampling via Convex Optimization
Rahul Krishna Thomas (Ritual), Arka Pal (Ritual)
OptimizationComputational EfficiencyTransformerText
🎯 What it does: Propose a Global Resolution algorithm that can precisely and efficiently solve the Optimal Transport problem in single-step multi-draft reasoning, significantly improving the acceptance rate of speculative sampling and reducing inference latency.
Global-Recent Semantic Reasoning on Dynamic Text-Attributed Graphs with Large Language Models
Yunan Wang (Beihang University), Ziwei Zhang (Beihang University)
Recurrent Neural NetworkGraph Neural NetworkTransformerLarge Language ModelTextGraph
🎯 What it does: Proposed and implemented DyGRASP, a model specifically designed for dynamic text-attribute graphs (DyTAG), which captures recent and global temporal semantics through implicit and explicit reasoning of LLMs, and fuses graph structural information via temporal graph neural networks;
Globally aware optimization with resurgence
Wei Bu (Harvard University)
OptimizationTransformerImageText
🎯 What it does: Propose a global optimization framework called SURGE based on the theory of reproduction in complex analysis, which acquires the target value by analyzing the singularities of the Borel transform of the partition function of neural networks and dynamically adjusts the learning rate during gradient descent.
GlobeDiff: State Diffusion Process for Partial Observability in Multi-Agent System
Yiqin Yang (Key Laboratory of Cognition and Decision Intelligence for Complex Systems, Institute of Automation, Chinese Academy of Sciences), Bo XU
Convolutional Neural NetworkReinforcement LearningDiffusion model
🎯 What it does: Propose a global state inference framework (GlobeDiff) based on conditional diffusion models, which transforms the many-to-one mapping of multi-modal observations into a reversible denoising process through latent variables, to address the local observability problem in multi-agent systems.
GlowQ: Group-Shared LOw-Rank Approximation for Quantized LLMs
Selim An (DGIST), Yeseong Kim (POSTECH)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Propose GlowQ, a group-sharing low-rank compensation method for quantized large language models (LLMs), reducing computational and memory overhead during inference.
gLSTM: Mitigating Over-Squashing by Increasing Storage Capacity
Hugh Blayney (University of Oxford), Michael M. Bronstein (University of Oxford)
Graph Neural NetworkGraph
🎯 What it does: This paper reinterprets the over-compression problem in GNNs, distinguishing between capacity over-compression and sensitivity over-compression, and proposes a novel neighbor-associated memory task (NAR) for single-layer testing of capacity bottlenecks. Based on this task, the authors designed gLSTM (a MPNN combining graph networks with associative memory) and achieved significant improvements on various graph long-distance tasks.
GmNet: Revisiting Gating Mechanisms From A Frequency View
Yifan Wang (Northeastern University), Yun Fu (Northeastern University)
ClassificationComputational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: Analyze the low-frequency bias in lightweight networks from a frequency perspective, and design GmNet by introducing a simple GLU to enhance high-frequency information.
GneissWeb: Preparing High Quality Data for LLMs at Scale
Hajar Emami Gohari (IBM Research), Bishwaranjan Bhattacharjee (IBM Research)
Data-Centric LearningLarge Language ModelText
🎯 What it does: Proposed the GneissWeb dataset, with a scale of approximately 10 trillion words, aiming to meet the dual demands of scale and quality for LLM Stage-1 pre-training;
GNN Explanations that do not Explain and How to find Them
Steve Azzolin (University of Trento), Sagar Malhotra (TU Wien)
Explainability and InterpretabilityGraph Neural NetworkImageTextGraphBenchmark
🎯 What it does: This paper studies self-explaining graph neural networks (SE-GNN), revealing that they may generate 'degenerate' explanations unrelated to actual predictions, and proposes attack methods, evaluation benchmarks, and more reliable credibility metrics.
GNN-as-Judge: Unleashing the Power of LLMs for Graph Learning with GNN Feedback
Ruiyao Xu (Northwestern University), Kaize Ding (Northwestern University)
ClassificationRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningTextGraph
🎯 What it does: Propose a framework that uses graph neural networks (GNNs) as judges for pseudo-label learning and fine-tuning of large language models (LLMs), specifically designed for semi-supervised node classification under text attribute graphs (TAG) with scarce labels.
Go Beyond Earth: Understanding Human Actions and Scenes in Microgravity Environments
Di Wen (Karlsruhe Institute of Technology), Rainer Stiefelhagen (Karlsruhe Institute of Technology)
RecognitionGenerationRetrievalConvolutional Neural NetworkTransformerSupervised Fine-TuningVideoTextMultimodalityBenchmarkPhysics Related
🎯 What it does: This paper constructs and releases the MicroG-4M dataset and MicroG-Bench benchmark, providing 4,759 segments of 3-second microgravity videos with fine-grained action annotations, 1,280 high-quality captions, and 7,428 visual question-answer pairs, supporting action recognition, video caption generation, and visual question answering tasks in microgravity environments.
Go-Browse: Training Web Agents with Structured Exploration
Apurva Gandhi (Carnegie Mellon University), Graham Neubig (Carnegie Mellon University)
TransformerLarge Language ModelSupervised Fine-TuningAgentic AITextSequential
🎯 What it does: Automatically collect high-quality web proxy data by structuring exploration of the web environment, and use this data to fine-tune a 7B LLM to improve task success rates on WebArena.
Goal Reaching with Eikonal-Constrained Hierarchical Quasimetric Reinforcement Learning
Vittorio Giammarino (Purdue University), Ahmed H Qureshi (Purdue University)
Reinforcement LearningBenchmark
🎯 What it does: Proposes a continuous-time, trajectory-free reinforcement learning framework (Eik-QRL) and its hierarchical variant (Eik-HiQRL) based on the Eikonal partial differential equation (PDE) with bounded value functions as quasi-metrics.
Goal-Aware Identification and Rectification of Misinformation in Multi-Agent Systems
Zherui Li (Beijing University of Posts and Telecommunications), Junfeng Fang (National University of Singapore)
Anomaly DetectionAdversarial AttackTransformerLarge Language ModelAgentic AITextChain-of-Thought
🎯 What it does: Investigate and evaluate misinformation injection attacks in multi-agent systems (MAS), propose a novel red-team dataset MISINFOTASK, and design a two-phase training-free defense framework ARGUS
GoalRank: Group-Relative Optimization for a Large Ranking Model
Kaike Zhang (Kuaishou Technology), Kun Gai (Kuaishou Technology)
Recommendation SystemReinforcement Learning from Human FeedbackTextSequential
🎯 What it does: Proposed a large-scale single-stage generative ranking model called GoalRank, which directly generates recommendation lists using a large generator, replacing the traditional multi-stage generate-evaluate architecture.
Goedel-Prover-V2: Scaling Formal Theorem Proving with Scaffolded Data Synthesis and Self-Correction
Yong Lin (Princeton University), Chi Jin (Princeton University)
AI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
🎯 What it does: This paper proposes and releases the open-source automated theorem proving model series Goedel-Prover-V2 (in two sizes: 8B and 32B), which can generate complete proofs using the Lean 4 compiler and supports self-correction;
Gogo: Group-wise granularity-ordered codec for stable and efficient speech generation
Weidong Chen (Chinese University of Hong Kong), Xixin Wu (Chinese University of Hong Kong)
GenerationTransformerReinforcement LearningFlow-based ModelAuto EncoderAudio
🎯 What it does: Proposed the Gogo encoder and GogoSpeech two-stage speech generation model, achieving hierarchical speech generation from coarse to fine granularity, supporting autoregressive modeling and improving generation efficiency;
GoldenStart: Q-Guided Priors and Entropy Control for Distilling Flow Policies
He Zhang (Hong Kong University of Science and Technology), Hui Xiong (Hong Kong University of Science and Technology)
OptimizationKnowledge DistillationRobotic IntelligenceReinforcement LearningFlow-based ModelTabular
🎯 What it does: Proposed a flow matching strategy distillation framework named GS-flow, which leverages Q-guided priors and entropy regularization to enhance the inference speed, accuracy, and explorability of sparse multi-modal strategies.
GOLDILOCS: GENERAL OBJECT-LEVEL DETECTION AND LABELING OF CHANGES IN SCENES
Almog Friedlander (Reichman University), Ohad Fried (Reichman University)
Object DetectionSegmentationNeural Radiance FieldGaussian SplattingImageVideo
🎯 What it does: This paper proposes GOLDILOCS, a zero-shot, pose-independent object-level change detection method based on 3D reconstruction.
Good Allocations from Bad Estimates
Sílvia Casacuberta (Stanford University), Moritz Hardt (Max Planck Institute for Intelligent Systems)
OptimizationTabularElectronic Health Records
🎯 What it does: Studied how to quickly obtain approximately optimal treatment allocation schemes with fewer samples under a limited budget, without requiring full CATE estimation.
GOOD: Geometry-guided Out-of-Distribution Modeling for Open-set Test-time Adaptation in Point Cloud Semantic Segmentation
Tianpei Zou (Tongji University), Guang Chen (Tongji University)
SegmentationDomain AdaptationPoint Cloud
🎯 What it does: Propose a framework named GOOD for open-set test-time adaptation (OSTTA) in 3D point cloud semantic segmentation, which clusters point clouds into superpoints, constructs superpoint confidence and prototypes, and generates pseudo-labels and updates the model by combining temporal consistency.
GoR: A Unified and Extensible Generative Framework for Ordinal Regression
Hongxu Ma (Fudan University), Shuigeng Zhou
GenerationRecurrent Neural NetworkTransformerImage
🎯 What it does: Proposed a generative sequence generation framework named GoR, transforming ordinal regression tasks into autoregressive token sequence prediction problems.
GOT-Edit: Geometry-Aware Generic Object Tracking via Online Model Editing
Shih-Fang Chen (National Yang Ming Chiao Tung University), Yen-Yu Lin (National Yang Ming Chiao Tung University)
Object TrackingTransformerContrastive LearningVideoBenchmark
🎯 What it does: Integrating 3D geometric information into 2D video streams, the paper proposes the GOT-Edit framework, which maintains semantic consistency through online model editing to achieve geometry-aware generic object tracking.
GoT-R1: Unleashing Reasoning Capability of Autoregressive Visual Generation with Reinforcement Learning
Chengqi Duan (HKU MMLAB), Xihui Liu (HKU MMLAB)
GenerationTransformerLarge Language ModelReinforcement LearningVision Language ModelImageTextMultimodalityChain-of-Thought
🎯 What it does: Propose the GoT-R1 framework, which enhances the semantic-space reasoning capabilities of autoregressive visual generation models through reinforcement learning, enabling the model to self-discover more efficient reasoning strategies.
GPG: A Simple and Strong Reinforcement Learning Baseline for Model Reasoning
Xiangxiang Chu (AMAP, Alibaba Group), Yong Wang (AMAP, Alibaba Group)
Large Language ModelReinforcement LearningTextMultimodalityBenchmark
🎯 What it does: Propose a simplified reinforcement learning (RL) method based on Group Policy Gradient (GPG) to enhance the reasoning capabilities of large language models and multimodal models.
GPS: Graph-guided Proactive Information Seeking in Large Language Models
Ruiqing Li (Peking University), Xu Chu (Peking University)
RetrievalComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextRetrieval-Augmented Generation
🎯 What it does: Propose the GPS framework, enabling LLMs to actively ask clarifying questions in RAG scenarios. First, construct a conditional reasoning DAG and dynamically prune the graph based on user feedback during traversal, ultimately achieving precise answers.
GPT4Scene: Understand 3D Scenes from Videos with Vision-Language Models
Zhangyang Qi (University of Hong Kong), Hengshuang Zhao (University of Hong Kong)
SegmentationGenerationSupervised Fine-TuningVision Language ModelVideoTextMultimodalityPoint Cloud
🎯 What it does: Propose the GPT4Scene framework, which utilizes pure visual input (video) to generate bird's-eye-view (BEV) and spatiotemporal object markers (STO-markers) through 3D reconstruction, achieving global-local understanding of indoor 3D scenes and performing inference on multiple tasks (3D QA, dense caption, visual grounding).
GPTailor: Large Language Model Pruning Through Layer Cutting and Stitching
Guinan Su (Max Planck Institute for Intelligent Systems), Jonas Geiping (Max Planck Institute for Intelligent Systems)
CompressionComputational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose a zeroth-order optimization method (GPTAILOR) that achieves efficient compression of large language models by pruning, selecting, and fusing layers across multiple fine-tuned models.
GRACE: A Language Model Framework for Explainable Inverse Reinforcement Learning
Silvia Sapora, Bogdan Mazoure
Explainability and InterpretabilityTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Propose GRACE, an IRL framework that leverages large language models and evolutionary search to reverse-engineer interpretable code-based reward functions;
GRACE: Generative Representation Learning via Contrastive Policy Optimization
Jiashuo Sun (University of Illinois Urbana Champaign), Jiawei Han (University of Illinois Urbana Champaign)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelReinforcement LearningContrastive LearningText
🎯 What it does: Propose the GRACE framework, which rewrites contrastive learning signals as rewards to guide LLMs in generating interpretable reasoning texts (rationales) and encodes them into high-quality embeddings, thereby transforming LLMs from 'black-box encoders' into interpretable representation learners.
GRADIEND: Feature Learning within Neural Networks Exemplified through Biases
Jonathan Drechsel (University of Passau), Steffen Herbold (University of Passau)
Explainability and InterpretabilityRepresentation LearningLarge Language ModelText
🎯 What it does: Introduces a gradient-based encoder-decoder architecture called GRADIEND, designed to learn interpretable feature neurons and directly modify existing model weights, achieving debiasing and bias amplification for social biases (e.g., gender, race, religion).
Gradient Descent Dynamics of Rank-One Matrix Denoising
Zeyan Zhuang (Hong Kong University of Science and Technology), Shenghui Song (Hong Kong University of Science and Technology)
OptimizationOrdinary Differential Equation
🎯 What it does: This paper studies the learning dynamics of gradient descent in rank-one matrix denoising, providing closed-form deterministic approximations for the inner product and loss function, and proving the almost sure convergence of the stochastic process and its long-term behavior in the high-dimensional limit.
Gradient Descent with Large Step Sizes: Chaos and Fractal Convergence Region
Shuang Liang (UCLA), Guido Montufar
Optimization
🎯 What it does: Studied the performance of gradient descent with large step sizes in matrix factorization, revealing fractal structures in the parameter space and deriving the convergence critical step size for scalar-vector decomposition.
Gradient Intrinsic Dimensionality Alignment：Narrowing The Gap Between Low-Rank Adaptation and Full Fine-Tuning
Jingqi Ye (Shanghai Artificial Intelligence Laboratory), Peng Ye (Shanghai Artificial Intelligence Laboratory)
Computational EfficiencyTransformerSupervised Fine-TuningImageText
🎯 What it does: Investigate the performance bottlenecks of LoRA, propose the concept of gradient intrinsic dimension (GID), and enhance parameter-efficient fine-tuning effectiveness through entropy-based estimators and intra- and inter-layer alignment mechanisms.
Gradient-Aligned Calibration for Post-Training Quantization of Diffusion Models
Dung Anh Hoang (Monash University), Thanh-Toan Do (Monash University)
GenerationDiffusion modelImage
🎯 What it does: Proposed a meta-learning weighted calibration method based on gradient alignment for post-training quantization of diffusion models.
Gradient-Based Diversity Optimization with Differentiable Top-$k$ Objective
Tianyi Zhou (KTH Royal Institute of Technology), Aristides Gionis (KTH Royal Institute of Technology)
Recommendation SystemMeta LearningTabular
🎯 What it does: This paper proposes a differentiable top-k diversity objective, integrating it with the relevance objective into gradient optimization to achieve model-agnostic collaborative learning of diversity and relevance; meanwhile, it provides two strategies, direct optimization (DDT) and indirect reweighting (MDR), applicable to both training from scratch and fine-tuning.
Gradient-Based Program Synthesis with Neurally Interpreted Languages
Matthew Macfarlane, Levi Lelis
AI Code AssistantTransformerTextBenchmark
🎯 What it does: Achieve program inference through end-to-end learning of a discrete programming language and its differentiable interpreter.