ICLR 2026 Papers — Page 19
International Conference on Learning Representations · 5356 papers
From Broad Exploration to Stable Synthesis: Entropy-Guided Optimization for Autoregressive Image Generation
Han Song (Chinese University of Hong Kong), Yu Cheng (Chinese University of Hong Kong)
GenerationOptimizationTransformerReinforcement LearningImageTextBenchmarkChain-of-Thought
🎯 What it does: Proposed an entropy-based guidance strategy to improve the text-to-image generation process of GRPO;
From Cheap Geometry to Expensive Physics: A Physics-agnostic Pretraining Framework for Neural Operators
Zhizhou Zhang (Bosch (China) Investment Co., Ltd.), Yanjia Wang (Bosch (China) Investment Co., Ltd.)
Representation LearningData-Centric LearningTransformerAuto EncoderPoint CloudPhysics Related
🎯 What it does: Proposed a physics-agnostic pre-training framework that self-supervisedly reconstructs the occupancy field on point clouds with only geometric data to pre-train a VAE, providing better geometric embeddings for training neural operators with limited PDE annotations.
From Collapse to Control: Understanding and Extending Context Length in Emerging Hybrid Models via Universal Position Interpolation
Haochen Shen (University of Illinois Urbana-Champaign), Minjia Zhang (University of Illinois Urbana-Champaign)
GenerationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper proposes a training-free context length extension method called Universal Position Interpolation (UPI) for long-context reasoning in hybrid Mamba-Transformer language models;
From Concepts to Components: Concept-Agnostic Attention Module Discovery in Transformers
Jingtong Su (New York University), Karen Ullrich (Meta AI)
Explainability and InterpretabilityComputational EfficiencyTransformerImageTextBenchmark
🎯 What it does: Proposes a concept-agnostic attention module discovery (SAMD) and scalar intervention (SAMI) method to locate and control attention heads in Transformers corresponding to arbitrary concepts.
From Conversation to Query Execution: Benchmarking User and Tool Interactions for EHR Database Agents
Gyubok Lee (KAIST), Edward Choi (KAIST)
TransformerLarge Language ModelAgentic AITabularBiomedical DataElectronic Health RecordsBenchmark
🎯 What it does: Proposes EHR-ChatQA, an interactive question-answering benchmark for electronic health record (EHR) databases, designed to evaluate the complete workflow of database agents in clarifying user intent, resolving terminology mismatches, invoking tools, and generating accurate SQL queries.
From Curiosity to Caution: Mitigating Reward Hacking for Best-of-$N$ with Pessimism
Zhuohao Yu (Carnegie Mellon University), Adam Block (Columbia University)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmark
🎯 What it does: Propose an uncertainty estimation method based on a lazy reward model, called 'caution,' to suppress vulnerabilities in reward models (reward hacking) during Best-of-N sampling and enhance computational scalability during inference.
From Data Statistics to Feature Geometry: How Correlations Shape Superposition
Lucas Prieto (Imperial College London), Pedro A. M. Mediano
Explainability and InterpretabilityRepresentation LearningAuto EncoderText
🎯 What it does: Propose the BOWS framework and train autoencoders within it to investigate the geometric structure of feature superposition on real data.
From EduVisBench to EduVisAgent: A Benchmark and Multi-Agent Framework for Reasoning-Driven Pedagogical Visualization
Haonian Ji (University Of Chicago), Huaxiu Yao (Rutgers University)
Large Language ModelAgentic AIVision Language ModelDiffusion modelImageTextBenchmark
🎯 What it does: This paper proposes the EduVisBench benchmark and the EduVisAgent multi-agent framework for evaluating and generating visualization explanations with pedagogical effectiveness;
From Embedding to Control: Representations for Stochastic Multi-Object Systems
Xiaoyuan Cheng (University College London), Yukun Hu (University College London)
Representation LearningGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: This study investigates how to learn controllable embeddings in random multi-object systems to achieve precise modeling and effective control.
From Evaluation to Defense: Advancing Safety in Video Large Language Models
Yiwei Sun (University of Science and Technology of China), Hongtao Xie (University of Science and Technology of China)
Safty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringVideoTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposed the Video Safety Evaluation (VSE) benchmark and the post-training defense framework VideoSafety-R1, systematically evaluating and enhancing the safety of video large models
From f(x) and g(x) to f(g(x)): LLMs Learn New Skills in RL by Composing Old Ones
Lifan Yuan (University of Illinois Urbana Champaign), Hao Peng (Tsinghua University)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This paper constructs clean string transformation composite tasks to demonstrate that RL can learn new composite skills based on existing atomic skills and transfer them to different tasks;
From Fields to Random Trees
Yaomin Wang (Chinese University of Hong Kong Shenzhen), Tianshu Yu (Chinese University of Hong Kong Shenzhen)
OptimizationComputational EfficiencyGraph
🎯 What it does: This paper proposes a method for MAP approximate inference in Markov Random Fields based on random spanning tree sampling, performing exact inference on each tree, and then merging results through edge occurrence probability weighted combination.
From Five Dimensions to Many: Large Language Models as Precise and Interpretable Psychological Profilers
Yi-Fei Liu (Peking University), Hang Zhang (Peking University)
Explainability and InterpretabilityTransformerPrompt EngineeringTextTabularChain-of-Thought
🎯 What it does: Investigating how large language models (LLMs) can zero-shot predict and reconstruct the correlation structures of nine psychological scales based on only a few individual inputs from the Big Five personality scale;
From Gradient Volume to Shapley Fairness: Towards Fair Multi-Task Learning
Xiao Wang (Beijing University of Chemical Technology), Min Tang (National Innovation Institute of Defense Technology)
OptimizationExplainability and InterpretabilityReinforcement LearningImageBenchmark
🎯 What it does: This paper proposes SVFair, a fair gradient aggregation framework based on Shapley values, designed to address gradient conflicts in multi-task learning.
From Language to Locomotion: Retargeting-free Humanoid Control via Motion Latent Guidance
Zhe Li (BAAI), Shanghang Zhang (BAAI)
Robotic IntelligenceTransformerMixture of ExpertsDiffusion modelSequential
🎯 What it does: Propose a robot control framework called RoboGhost that completely eliminates action redirection, enabling language instructions to directly drive full-body motion;
From Large to Small: Transferring CUDA Optimization Expertise via Reasoning Graph
Junfeng Gong (Institute of Computing Technology Chinese Academy of Sciences), Huawei Li (Institute of Computing Technology Chinese Academy of Sciences)
OptimizationComputational EfficiencyAI Code AssistantGraph Neural NetworkTransformerLarge Language ModelTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes a training-free framework called ReGraphT, which constructs a CUDA-optimized inference graph using the multi-step reasoning trajectories of large language models (LLMs), and enables small language models to generate efficient CUDA code through Monte Carlo graph search.
From Markov to Laplace: How Mamba In-Context Learns Markov Chains
Marco Bondaschi (École Polytechnique Fédérale de Lausanne), Ashok Vardhan Makkuva (Télécom Paris)
Explainability and InterpretabilityRepresentation LearningLarge Language ModelTextSequential
🎯 What it does: Studied Mamba's in-context learning (ICL) capability on random Markov chains, and proved that a single-layer Mamba can learn the Laplacian smoothing estimator, demonstrating its ability to reach optimal statistical inference;
From Medical Records to Diagnostic Dialogues: A Clinical-Grounded Approach and Dataset for Psychiatric Comorbidity
Tianxi Wan (Shanghai Jiao Tong University), Mengyue Wu (Shanghai Jiao Tong University)
TransformerLarge Language ModelSupervised Fine-TuningAgentic AIPrompt EngineeringTextTabularElectronic Health Records
🎯 What it does: Constructed the first large-scale merged psychiatric diagnosis dialogue dataset, PsyCoTalk. Initially, a modular workflow was used to convert Reddit posts self-reporting multiple mental health conditions into 502 structured electronic medical records (PsyCoProfile). Subsequently, a hierarchical diagnostic state machine (HDSM) and diagnostic context tree (DCT) based on DSM-5 were employed to generate 3,000 multiround, verifiable Chinese diagnostic dialogues among three agents (doctor, patient, tool).
From movement to cognitive maps: recurrent neural networks reveal how locomotor development shapes hippocampal spatial coding
Marco P Abrate (University College London), Caswell Barry (University College London)
Representation LearningRecurrent Neural NetworkSequentialBiomedical Data
🎯 What it does: Integrate experimental gait development data with recurrent neural networks (RNNs) to explore how gait changes drive the formation and maturation of hippocampal spatial cells.
From Narrow to Panoramic Vision: Attention-Guided Cold-Start Reshapes Multimodal Reasoning
Ruilin Luo (Tsinghua University), Zhibo Yang (Qwen Team, Alibaba Group)
Data SynthesisReinforcement LearningVision Language ModelMultimodality
🎯 What it does: During the cold start phase of multimodal large-scale reasoning models, the Visual Attention Score (VAS) was introduced to measure the model's attention to visual tokens, revealing a strong correlation between this score and reasoning performance; subsequently, the 'lazy attention localization' phenomenon was proposed, and the causal impact of attention allocation on performance was validated through untrained attention regulation methods; based on this, the AVAR framework was constructed, including vision-anchored reflective data synthesis, attention-guided training objectives, and vision-anchored reward shaping to significantly enhance multimodal reasoning capabilities.
From Natural Alignment to Conditional Controllability in Multimodal Dialogue
Zeyu Jin (Tsinghua University), Jia Jia (Tsinghua University)
GenerationLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodalityBenchmark
🎯 What it does: Constructed the MM-DIA multimodal dialogue dataset and proposed the MM-DIA-BENCH benchmark based on this dataset, introducing the task of multimodal dialogue generation and its evaluation methods.
From Neural Networks to Logical Theories: The Correspondence between Fibring Modal Logics and Fibring Neural Networks
Ouns El Harzli (University of Oxford), Tarek R. Besold (Sony AI Barcelona)
Representation LearningGraph Neural NetworkTransformerGraph
🎯 What it does: This paper formally transplants the concept of fibrations from modal logic into neural networks, establishing a correspondence between fibrated neural networks and fibrated modal logic. It leverages this correspondence to derive logical expressiveness results for graph neural networks (GNN), graph attention networks (GAT), and Transformer encoders on non-uniform expressions.
From Observations to Events: Event-Aware World Models for Reinforcement Learning
Zhao-Han Peng (Tsinghua University), You He (Tsinghua University)
Reinforcement LearningWorld ModelMultimodalityBenchmark
🎯 What it does: Proposed the Event-Aware World Model (EAWM) framework, which learns critical event representations in multi-modal environments using unsupervised event generation and event segmentors, thereby improving sample efficiency and generalization ability of model-based reinforcement learning.
From Parameters to Behaviors: Unsupervised Compression of the Policy Space
Davide Tenedini (Politecnico di Milano), Marcello Restelli (Politecnico di Milano)
Representation LearningReinforcement LearningAuto Encoder
🎯 What it does: Proposed an unsupervised compression strategy that maps the high-dimensional policy parameter space to a low-dimensional latent space, using the latent space for task-specific gradient optimization;
From Pixels to Semantics: Unified Facial Action Representation Learning for Micro-Expression Analysis
Yicheng Deng (University of Osaka), Hajime Nagahara (University of Osaka)
RecognitionRepresentation LearningTransformerVision Language ModelAuto EncoderContrastive LearningVideo
🎯 What it does: Proposed the D-FACE framework, shifting micro-expression recognition from pixel-level motion to discrete semantic facial action coding, and combining Transformer with CLIP for emotion alignment.
From Pixels to Words -- Towards Native Vision-Language Primitives at Scale
Haiwen Diao (Nanyang Technological University), Ziwei Liu (Nanyang Technological University)
TransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: Designed and trained a unified vision-language model named NEO from scratch, which achieves encoding, alignment, and reasoning for vision and language through a unified primitive.
From Prediction to Perfection: Introducing Refinement to Autoregressive Image Generation
Cheng Cheng (Xi'an Jiaotong University), Ying Shan (Tencent PCG)
GenerationTransformerDiffusion modelImage
🎯 What it does: Propose the TensorAR framework, extending traditional AR models from token-by-token prediction to tensor-by-tensor prediction, enabling iterative refinement of generated content during the generation process;
From Predictors to Samplers via the Training Trajectory
Soumya Ram (Independent Researcher), Akhila Ram (Massachusetts Institute of Technology)
GenerationImageBiomedical DataPhysics Related
🎯 What it does: Propose a trajectory annealing method that samples on the training trajectory of a trained predictor, leveraging the roughness of early models to accelerate mixing and address high-frequency or spiky targets;
From Reproduction to Replication: Evaluating Research Agents with Progressive Code Masking
Gyeongwon James Kim (Carnegie Mellon University), Daniel Fried (Carnegie Mellon University)
AI Code AssistantTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes the AUTOEXPERIMENT benchmark to evaluate AI agents' ability to automatically generate missing functions, run experiments, and reproduce results given a research paper, restricted code, and experimental commands.
From Samples to Scenarios: A New Paradigm for Probabilistic Forecasting
Xilin Dai (Zhejiang University), Qiang Xu (Chinese University of Hong Kong)
Time Series
🎯 What it does: Propose the Probabilistic Scenarios paradigm, construct the TimePrism model composed of only three linear layers, and output a limited number of {scenario, probability} pairs through a single forward inference;
From Seeing to Doing: Bridging Reasoning and Decision for Robotic Manipulation
Yifu Yuan (Tianjin University), Jianye HAO
Robotic IntelligenceTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVision-Language-Action ModelImageTextMultimodalityChain-of-Thought
🎯 What it does: Propose the FSD framework, which enables robot manipulation from 'observation' to 'execution' by generating intermediate visual aids (such as spatial feasible regions, points, and trajectories) through spatial reasoning of vision-language models;
From Seeing to Experiencing: Scaling Navigation Foundation Models with Reinforcement Learning
Honglin He (University of California, Los Angeles), Bolei Zhou (University of California, Los Angeles)
Autonomous DrivingRobotic IntelligenceSupervised Fine-TuningReinforcement LearningVision-Language-Action ModelGaussian SplattingVideoMultimodalityBenchmark
🎯 What it does: Propose the Seeing-to-Experiencing (S2E) framework by pre-training on massive offline videos and combining with reinforcement learning, enabling navigation foundation models to transition from 'watching' to 'experiencing' and achieve zero-shot transfer to different robot platforms.
From Sequential to Parallel: Reformulating Dynamic Programming as GPU Kernels for Large-Scale Stochastic Combinatorial Optimization
Jingyi Zhao (Shenzhen Research Institute of Big Data), Tian Ding (Shenzhen Research Institute of Big Data)
OptimizationBenchmark
🎯 What it does: This paper proposes a GPU-based parallel framework that re-expresses the second-stage dynamic programming (DP) in stochastic programming as batchable and scalable matrix operations, enabling high-throughput integer DP solving under millions of scenarios.
From Single to Multi-Granularity: Toward Long-Term Memory Association and Selection of Conversational Agents
Derong Xu (University of Science and Technology of China), Tong Xu (University of Science and Technology of China)
RetrievalComputational EfficiencyLarge Language ModelTextBenchmark
🎯 What it does: Propose the MemGAS framework, constructing multi-granularity memory units and achieving adaptive memory retrieval;
From Sorting Algorithms to Scalable Kernels: Bayesian Optimization in High-Dimensional Permutation Spaces
Zikai Xie (University of Science and Technology of China), Linjiang Chen (University of Science and Technology of China)
OptimizationBenchmark
🎯 What it does: This paper proposes a kernel construction framework based on sorting algorithms, and designs the Merge kernel within this framework to achieve Bayesian optimization in high-dimensional permutation spaces; systematic experiments are conducted on multi-dimensional permutation optimization benchmarks comparing the Merge kernel with existing methods such as Mallows kernel, BOPS-H, and TuRBO.
From Sparse to Dense: Spatio-Temporal Fusion for Multi-View 3D Human Pose Estimation with DenseWarper
Ling Li (Tsinghua University), Zhidong Deng (Tsinghua University)
Pose EstimationVideo
🎯 What it does: Propose a sparse interleaved multi-view input paradigm, utilizing images captured by different cameras at different time points to achieve spatiotemporal information fusion, and design the DenseWarper framework to achieve high-frequency pose output.
From Spatial to Actions: Grounding Vision-Language-Action Model in Spatial Foundation Priors
Zhengshen Zhang (ByteDance Seed), Pan Zhou (Singapore Management University)
Robotic IntelligenceSupervised Fine-TuningVision Language ModelVision-Language-Action ModelMultimodality
🎯 What it does: Proposes FALCON, a framework that injects 3D spatial tokens into the action head of a vision-language-action (VLA) model, significantly enhancing a robot's perception and decision-making capabilities regarding spatial information.
From Static Benchmarks to Dynamic Protocol: Agent-Centric Text Anomaly Detection for Evaluating LLM Reasoning
Seungdong Yoa (LG AI Research), Woohyung Lim (Hankuk University of Foreign Studies)
Anomaly DetectionData-Centric LearningTransformerLarge Language ModelAgentic AITextBenchmark
🎯 What it does: Proposed a dynamic text anomaly detection evaluation protocol ATAD based on three agents (teacher, coordinator, student) for adaptive generation, verification, and solving of problems to assess the reasoning capabilities of large language models.
From Text to Talk: Audio-Language Model Needs Non-Autoregressive Joint Training
Tianqiao Liu (Jinan University), Zitao Liu (Peking University)
GenerationTransformerLarge Language ModelDiffusion modelTextMultimodalityAudio
🎯 What it does: Proposed the Text-to-Talk (TtT) framework, which unifies the autoregressive generation of text with the non-autoregressive diffusion model for audio within a single Transformer, achieving end-to-end speech-to-speech dialogue.
From Ticks to Flows: Dynamics of Neural Reinforcement Learning in Continuous Environments
Saket Tiwari (Brown University), George Konidaris (Brown University)
Reinforcement LearningStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper constructs a theoretical framework for continuous-time reinforcement learning and derives a non-parametric closed-form system for the time evolution of gradients in the Actor-Critic algorithm under infinite-width single-hidden-layer neural networks.
From Tokens to Nodes: Semantic-Guided Motion Control for Dynamic 3D Gaussian Splatting
Jianing Chen (University of Chinese Academy of Sciences), Yucheng Zhang (University of Chinese Academy of Sciences)
RestorationGaussian SplattingVideo
🎯 What it does: This paper proposes a dynamic 3D Gaussian splatting framework based on semantic and motion priors, achieving high-quality reconstruction of dynamic scenes.
From Tokens to Thoughts: How LLMs and Humans Trade Compression for Meaning
Chen Shani (Stanford University), Ravid Shwartz-Ziv (New York University)
Representation LearningTransformerLarge Language ModelTextBenchmark
🎯 What it does: Compare LLMs and humans in the compression-meaning trade-off of concept representation, quantitatively evaluating 40+ LLM embeddings and classical cognitive experiments (Rosch 1973/1975, McCloskey & Glucksberg 1978) using the information bottleneck and rate-distortion theory.
From Utterance to Vividity: Training Expressive Subtitle Translation LLM via Adaptive Local Preference Optimization
Chaoqun Cui (Chinese Academy of Sciences), Wenji Mao (Chinese Academy of Sciences)
GenerationTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Built and trained an expressive translation large language model (LLM) for visual media captioning, and proposed a local preference optimization method called ALPO;
From Verifiable Dot to Reward Chain: Harnessing Verifiable Reference-based Rewards for Reinforcement Learning of Open-ended Generation
Yuxin Jiang (Huawei Technologies Co Ltd), Lifeng Shang (Huawei Technologies Co Ltd)
GenerationData-Centric LearningLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: This paper proposes the RLVRR framework, utilizing a verifiable reference reward chain to achieve open-ended generation in reinforcement learning.
From Vicious to Virtuous Cycles: Synergistic Representation Learning for Unsupervised Video Object-Centric Learning
Hyun Seok Seong (Sungkyunkwan University), Jae-Pil Heo (Sungkyunkwan University)
Object TrackingRepresentation LearningTransformerContrastive LearningVideo
🎯 What it does: By aligning the encoder's sharp attention maps with the decoder's blurry reconstruction maps, complementary mutual refinement is achieved, constructing Synergistic Representation Learning (SRL), which resolves the vicious cycle in traditional video object-centric learning.
From What to Why: A Multi-Agent System for Evidence-based Chemical Reaction Condition Reasoning
Cheng Yang (Hangzhou Dianzi University), Feiwei Qin (Hangzhou Dianzi University)
Recommendation SystemDrug DiscoveryTransformerLarge Language ModelAgentic AITabularBiomedical DataRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the ChemMAS multi-agent system, redefining chemical reaction condition recommendation as an evidence-based reasoning task;
Front-Loading Reasoning: The Synergy between Pretraining and Post-Training Data
Syeda Nahida Akter (Carnegie Mellon University), Bryan Catanzaro (NVIDIA)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmark
🎯 What it does: Systematically investigated the injection of reasoning data into different stages of large language model training (pre-training and post-training), and evaluated its impact on multiple reasoning tasks through large-scale experiments, providing principles for allocating reasoning data throughout the entire training process.
FrontierCO: Real-World and Large-Scale Evaluation of Machine Learning Solvers for Combinatorial Optimization
Shengyu Feng (Carnegie Mellon University), Yiming Yang (Carnegie Mellon University)
OptimizationGraph Neural NetworkLarge Language ModelReinforcement LearningAgentic AIDiffusion modelGraphTabularBenchmark
🎯 What it does: Proposed the FRONTIERCO benchmark, collecting real and large-scale instances of eight categories of combinatorial optimization problems to uniformly evaluate the performance of neural networks, LLMs, and classical solvers.
FROST: Filtering Reasoning Outliers with Attention for Efficient Reasoning
Haozheng Luo (Northwestern University), Soumalya Sarkar (RTX Technology Research Center)
Computational EfficiencyTransformerSupervised Fine-TuningText
🎯 What it does: Propose the FROST method, which identifies and removes low-importance reasoning outliers during the reasoning process by using Softmax 1 in the attention mechanism, thereby achieving shorter and more reliable reasoning paths.
Frozen Policy Iteration: Computationally Efficient RL under Linear $Q^{\pi}$ Realizability for Deterministic Dynamics
Yijing Ke (Peking University), Ruosong Wang (Peking University)
Reinforcement Learning
🎯 What it does: Proposed an online reinforcement learning algorithm called Frozen Policy Iteration (FPI), which applies to Markov Decision Processes (MDP) with random initial states, random rewards, and deterministic transitions under the assumption of linear Qπ realizability.
Frozen Priors, Fluid Forecasts: Prequential Uncertainty for Low-Data Deployment with Pretrained Generative Models
Fernando Ruiz-Mazo (Aalto University), Vikas K Garg (Aalto University)
GenerationData SynthesisTransformerImageText
🎯 What it does: Propose a prequential quantification framework that, when only a small number of real samples are available, mixes a pre-trained generator with an empirical distribution via Dirichlet scheduling and estimates long-term metrics using a Markov posterior.
FrugalRAG: Less is More in RL Finetuning for Multi-hop Question Answering
Abhinav Java (Microsoft Research India), Amit Sharma (Microsoft Research India)
RetrievalOptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose a two-stage RL fine-tuning framework FRUGALRAG to achieve adaptive optimization in the retrieval step for multi-hop question answering
Frustratingly Simple Retrieval Improves Challenging, Reasoning-Intensive Benchmarks
Xinxi Lyu (University of Illinois Urbana Champaign), Sewon Min (University of California Berkeley)
RetrievalTransformerTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This work proposes a high-quality, deployable web-scale retrieval database named COMPACTDS, achieving significant performance improvements using minimal retrieval-generation (RAG) on multiple reasoning-intensive benchmarks (MMLU, MMLU Pro, AGI Eval, GPQA, MATH).
FS-DFM: Fast and Accurate Long Text Generation with Few-Step Diffusion Language Models
Amin Karimi Monsefi (Ohio State University), Irina Belousova (Apple)
GenerationComputational EfficiencyLarge Language ModelDiffusion modelText
🎯 What it does: Designed and implemented FS-DFM, a discrete flow matching language model supporting few-step sampling, capable of generating 1024-token text in just 8 steps while achieving up to 128 times speed improvement.
FS-KAN: Permutation Equivariant Kolmogorov-Arnold Networks via Function Sharing
Ran Elbaz (Technion Israel Institute of Technology), Haggai Maron (Technion Israel Institute of Technology)
ClassificationPoint CloudTabular
🎯 What it does: Proposed a generic framework named FS-KAN, constructing equivariant and invariant Kolmogorov-Arnold networks for arbitrary permutation symmetric groups using function sharing;
FSA: An Alternative Efficient Implementation of Native Sparse Attention Kernel
Ran Yan (Hong Kong University of Science and Technology), Binhang Yuan (Carnegie Mellon University)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Propose Flash Sparse Attention (FSA), an improved implementation of Native Sparse Attention (NSA) kernel, achieving efficient sparse attention computation for small GQA group LLMs through reversed loop order, index tensor usage, online softmax, and separated aggregation kernel.
FSD-CAP: Fractional Subgraph Diffusion with Class-Aware Propagation for Graph Feature Imputation
Xin Qiao (Xidian University), zhang liang
Representation LearningGraph Neural NetworkDiffusion modelGraph
🎯 What it does: Proposed the FSD-CAP two-stage framework for feature missing completion, first using fractional subgraph diffusion for preliminary completion, then refining features through class-level propagation to address graph feature completion under high missing rates.
FSOD-VFM: Few-Shot Object Detection with Vision Foundation Models and Graph Diffusion
Chen-Bin Feng (University of Macau), Xi SHEN
Object DetectionMeta LearningGraph Neural NetworkTransformerDiffusion modelImage
🎯 What it does: Propose FSOD-VFM, which utilizes Vision Foundation Models to achieve zero-shot object detection without training
FSPO: Few-Shot Optimization of Synthetic Preferences Effectively Personalizes to Real Users
Anikait Singh (Stanford University), Chelsea Finn (Stanford University)
Data SynthesisOptimizationMeta LearningReinforcement Learning from Human FeedbackLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Proposed the FSPO framework to enable rapid personalization generation of large language models with few user preferences, and constructed three synthetic preference datasets: Reviews, ELIX, and Roleplay.
Full-Graph vs. Mini-Batch Training: Comprehensive Analysis from a Batch Size and Fan-Out Size Perspective
Mengfan Liu (University of Hong Kong), Chuan Wu (University of Hong Kong)
Computational EfficiencyRepresentation LearningGraph Neural NetworkGraph
🎯 What it does: Compare full-graph training with mini-batch training, systematically analyze the impact of batch size (batch size) and fan-out size on the convergence speed, generalization capability, and computational efficiency of graph neural networks (GNN), and provide theoretical proofs and empirical validation.
FullPart: Generating each 3D Part at Full Resolution
Lihe Ding (CUHK MMLab), Tianfan Xue (CUHK MMLab)
GenerationData SynthesisDiffusion modelTextMultimodalityMesh
🎯 What it does: Proposed the FullPart framework to hierarchically generate 3D objects by parts, first generating layout boxes via implicit vector set diffusion, then generating details for each part in independent full-resolution voxel grids, ultimately refining them into textured meshes;
Function Induction and Task Generalization: An Interpretability Study with Off-by-One Addition
Qinyuan Ye (University of Southern California), Xiang Ren (University of Southern California)
Explainability and InterpretabilityTransformerText
🎯 What it does: Investigate how language models solve discrete tasks (such as offset addition) through function induction mechanisms in context learning
Functional MRI Time Series Generation via Wavelet-Based Image Transform and Spectral Flow Matching for Brain Disorder Identification
Hwa Hui Tew (Monash University Malaysia), Chee-Ming Ting (King Abdullah University of Science and Technology)
ClassificationGenerationData SynthesisTransformerOptical FlowTime SeriesBiomedical DataMagnetic Resonance ImagingOrdinary Differential Equation
🎯 What it does: Proposed a Dual-Spectral Flow Matching (DSFM) framework to generate high-fidelity fMRI BOLD time series and applied it to classify brain disorders (MDD, ASD);
Fused-Planes: Why Train a Thousand Tri-Planes When You Can Share?
Karim Kassab (Criteo AI Lab), Valerie Gouet-Brunet
GenerationData SynthesisComputational EfficiencyNeural Radiance FieldAuto EncoderPoint CloudMesh
🎯 What it does: Propose the Fused-Planes structure, decomposing Tri-Planes into object-specific micro-planes and shared macro base planes, jointly training them in a 3D-aware latent space to achieve efficient reconstruction of large-scale 3D object categories.
Fusing Pixels and Genes: Spatially-Aware Learning in Computational Pathology
Minghao Han (Fudan University), Lihua Zhang (Fudan University)
Representation LearningAuto EncoderContrastive LearningMultimodalityBiomedical Data
🎯 What it does: Proposed the STAMP framework, which performs multimodal representation learning by integrating histopathological images and spatial transcriptomics.
FutureFill: Fast Generation from Convolutional Sequence Models
Naman Agarwal (Google DeepMind), Elad Hazan (Google DeepMind)
GenerationComputational EfficiencyConvolutional Neural NetworkTransformerText
🎯 What it does: Designed a new subroutine called FutureFill, achieving efficient autoregressive generation based on convolutional sequence models. By introducing two variants, Epoched-FutureFill and Continuous-FutureFill, the generation time was reduced from O(L²) to O(L log²L) or O(L logL), while significantly decreasing cache size.
FutureMind: Equipping Small Language Models with Strategic Thinking-Pattern Priors via Adaptive Knowledge Distillation
Shaoxiong Yang, Jian Luan (Xiaomi Inc)
RetrievalKnowledge DistillationTransformerTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes FutureMind, a training-agnostic modular reasoning framework that leverages adaptive knowledge distillation from large language models to small language models, enabling small models to perform efficient and structured retrieval and reasoning in multi-hop question answering.
FutureX: An Advanced Live Benchmark for LLM Agents in Future Prediction
Zhiyuan Zeng (Fudan University), Wenhao Huang (ByteDance Seed)
Large Language ModelAgentic AITextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes FutureX, a real-time updated, cross-domain covered, data-pollution-free benchmark for future prediction evaluation;
FZOO: Fast Zeroth-Order Optimizer for Fine‑Tuning Large Language Models towards Adam‑Scale Speed
Sizhe Dang (Xi'an Jiaotong University), Haishan Ye (Xi'an Jiaotong University)
OptimizationComputational EfficiencyTransformerSupervised Fine-TuningText
🎯 What it does: Propose a zeroth-order optimizer FZOO for fine-tuning large language models, significantly reducing memory usage and accelerating convergence.
G-Merging: Graph Models Merging for Parameter-Efficient Multi-Task Knowledge Consolidation
Jun Chen (Shenzhen University), Xiao Luo (University of Wisconsin-Madison)
Computational EfficiencyKnowledge DistillationDrug DiscoveryGraph Neural NetworkMixture of ExpertsGraph
🎯 What it does: Propose G-Merging, a framework that merges parameters of multi-task fine-tuned GNN models, achieving parameter-efficient multi-task knowledge integration.
G-reasoner: Foundation Models for Unified Reasoning over Graph-structured Knowledge
Linhao Luo (Monash University), Shirui Pan (Griffith University)
RetrievalComputational EfficiencyKnowledge DistillationRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelTextGraphBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the G-REASONER framework, unifying graph-structured knowledge with language foundation models, introducing QuadGraph's four-layer abstraction to unify heterogeneous knowledge graphs, constructing a 34M-parameter graph foundation model (GFM), and achieving large-scale training and inference through mixed-precision training and distributed messaging.
G4Splat: Geometry-Guided Gaussian Splatting with Generative Prior
Junfeng Ni (Tsinghua University), Siyuan Huang (Tsinghua University)
GenerationDiffusion modelGaussian SplattingImageVideo
🎯 What it does: Propose the G4SPLAT method, combining plane geometry guidance with diffusion model generation priors to achieve high-quality 3D scene reconstruction from sparse views.
GAGA: Gaussianity-Aware Gaussian Approximation for Efficient 3D Molecular Generation
Jingxiang Qu (Stony Brook University), Yi Liu (Stony Brook University)
Drug DiscoveryDiffusion modelScore-based ModelBiomedical DataStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Proposes the GAGA (Gaussianity-Aware Gaussian Approximation) method, which identifies when data achieves sufficient Gaussianity on the Gaussian Probability Path, truncating redundant high-noise steps to accelerate training and sampling in 3D molecular generation models.
Gaia2: Benchmarking LLM Agents on Dynamic and Asynchronous Environments
Romain Froger (Meta SuperIntelligence Labs), Thomas Scialom (Meta SuperIntelligence Labs)
TransformerLarge Language ModelAgentic AITextBenchmark
🎯 What it does: Propose the GAIA2 benchmark for evaluating large language model agents in asynchronous, dynamic mobile application environments, achieving fine-grained assessment through write-level verification.
GaitSnippet: Gait Recognition Beyond Unordered Sets and Ordered Sequences
Saihui Hou (Beijing Normal University), Yongzhen Huang (Beijing Normal University)
RecognitionConvolutional Neural NetworkVideo
🎯 What it does: Proposes a snippet-based gait recognition framework called GaitSnippet, which integrates both short-term and long-term temporal information;
GALAX: Graph-Augmented Language Model for Explainable Reinforcement-Guided Subgraph Reasoning in Precision Medicine
Heming Zhang (Washington University in St. Louis), Fuhai Li (Washington University in St. Louis)
Explainability and InterpretabilityDrug DiscoveryGraph Neural NetworkLarge Language ModelReinforcement LearningTextMultimodalityGraphBiomedical DataBenchmarkRetrieval-Augmented Generation
🎯 What it does: Developed a graph-enhanced large language model named GALAX, leveraging reinforcement learning and a graph process reward model (GPRM) to achieve explainable subgraph reasoning in precision medicine, using multi-omics data, textual knowledge, and graph structures to predict patient-specific drug targets.
Game-RL: Synthesizing Multimodal Verifiable Game Data to Boost VLMs' General Reasoning
Jingqi Tong (Fudan University), Xuanjing Huang (Fudan University)
Data SynthesisLarge Language ModelReinforcement LearningVision Language ModelVideoTextMultimodality
🎯 What it does: Designed a verifiable vision-language reinforcement learning framework called Game-RL based on video games, and constructed a GameQA visual question answering dataset automatically generated by Code2Logic, using GRPO to train the Vision-Language Model (VLM);
GAPrune: Gradient-Alignment Pruning for Domain-Aware Embeddings
Yixuan Tang (Hong Kong University of Science and Technology), Yi Yang (Hong Kong University of Science and Technology)
Domain AdaptationComputational EfficiencyRepresentation LearningContrastive LearningTextFinance Related
🎯 What it does: Proposes a gradient-aligned pruning framework called GAPRUNE for domain-embedding models, which can maintain or enhance domain-specific performance while keeping the model sparse.
GAR: Generative Adversarial Reinforcement Learning for Formal Theorem Proving
Ruida WANG, Tong Zhang (University of Illinois Urbana Champaign)
AI Code AssistantReinforcement LearningGenerative Adversarial NetworkTextChain-of-Thought
🎯 What it does: Propose the GAR framework, jointly training the problem generator and prover through adversarial reinforcement learning to achieve implicit curriculum learning, significantly improving formal proof performance.
GARLIC: Graph Attention-based Relational Learning of Multivariate Time Series in Intensive Care
Ruirui Wang (University of Zürich), Diego Paez-Granados (ETH Zürich)
ClassificationExplainability and InterpretabilityRecurrent Neural NetworkGraph Neural NetworkAuto EncoderTime SeriesBiomedical DataElectronic Health Records
🎯 What it does: Propose a graph attention-based multivariate time series learning model called GARLIC, specifically designed to handle irregular sampling and missing data in ICU settings, while providing built-in interpretability;
GarmentGPT: Compositional Garment Pattern Generation via Discrete Latent Tokenization
Fangsheng Weng (ChimerAI), Xiaoguang Han (Chinese University of Hong Kong)
GenerationTransformerVision Language ModelAuto EncoderImageTextMultimodalityBenchmark
🎯 What it does: Proposes a framework called GarmentGPT that utilizes RVQ-VAE discretization and a vision-language model to generate and edit sewing patterns, enabling the creation of structured garment patterns from text and images.
GAS: Enhancing Reward-Cost Balance of Generative Model-assisted Offline Safe RL
Zifan LIU, Jun Zhang (Hong Kong University of Science and Technology)
OptimizationReinforcement LearningSequential
🎯 What it does: Propose Goal-Assisted Stitching (GAS), which achieves a balance between reward maximization and constraint satisfaction in offline safe reinforcement learning using generative models.
GAS: Improving Discretization of Diffusion ODEs via Generalized Adversarial Solver
Aleksandr Oganov (Hse University), Aibek Alanov (Hse University)
GenerationKnowledge DistillationDiffusion modelGenerative Adversarial NetworkImage
🎯 What it does: Proposes Generalized Adversarial Solver (GAS), a low NFE diffusion model sampling method that combines generalized solver parameterization, teacher-student distillation, and adversarial training.
Gauge Flow Matching: Efficient Constrained Generative Modeling over General Convex Set and Beyond
Xinpeng Li (City University of Hong Kong), Minghua Chen (City University of Hong Kong)
GenerationOptimizationFlow-based ModelImageTime Series
🎯 What it does: Propose the Gauge Flow Matching framework, which transforms any compact convex constraint domain into a unit ball via gauge mapping, and achieves strictly constrained generation by combining flow matching with reflection.
Gauge-invariant representation holonomy
Vasileios Sevetlidis (Athena Research Center), George Pavlidis (Athena Research Center)
Explainability and InterpretabilityRepresentation LearningConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a gauge-invariant connection metric called representation holonomy, which measures the 'twisting' degree of hierarchical representations in deep networks along input paths, and provides a practically computable estimator.
Gaussian certified unlearning in high dimensions: A hypothesis testing approach
Aaradhya Pandey (Princeton University), Sanjeev Kulkarni (Princeton University)
OptimizationImageTextTabular
🎯 What it does: This paper studies the problem of machine unlearning under high-dimensional settings (p≈n), proposing a new concept of privacy verifiability called ε-Gaussian verifiability. It proves that under this framework, a single noisy Newton step can simultaneously satisfy privacy and accuracy requirements.
GaussianFusion: Unified 3D Gaussian Representation for Multi-Modal Fusion Perception
Xiao Zhao (Tencent), Kuifeng Su (Tencent)
Autonomous DrivingGaussian SplattingImageMultimodalityPoint Cloud
🎯 What it does: Propose a multimodal fusion framework called GaussianFusion based on a unified 3D Gaussian representation, which can map camera and LiDAR features into a continuous space for fusion and support multi-task 3D perception.
GAVEL: Towards Rule-Based Safety through Activation Monitoring
Shir Rozenfeld (Ben Gurion University Negev), Yisroel Mirsky (Ben Gurion University Negev)
Safty and PrivacyExplainability and InterpretabilityRecurrent Neural NetworkTextBenchmark
🎯 What it does: Proposed the GAVEL framework, which monitors internal activations of large language models using cognitive elements (CEs) and implements real-time safety protection through rule-based logic;
GCGNet: Graph-Consistent Generative Network for Time Series Forecasting with Exogenous Variables
Zhengyu Li (East China Normal University), Bin Yang (East China Normal University)
Graph Neural NetworkAuto EncoderTime Series
🎯 What it does: Propose a graph consistent generative network named GCGNet for time series prediction with historical and future exogenous variables, jointly modeling spatiotemporal and channel correlations, and achieving robust prediction through variational generators, graph structure alignment, and graph refinement modules.
GDGB: A Benchmark for Generative Dynamic Text-Attributed Graph Learning
Jie Peng (Renmin University of China), Chuntao Hong (Ant Group)
GenerationData SynthesisLarge Language ModelAgentic AITextGraphBenchmark
🎯 What it does: Constructed GDGB, a high-quality text attribute dynamic graph benchmark, defined two generation tasks (TDGG and IDGG), and proposed the GAG-General framework for evaluation.
GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks
Tejal Patwardhan (OpenAI), Jerry Tworek (OpenAI)
TransformerLarge Language ModelPrompt EngineeringMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the GDPval benchmark, which evaluates AI models' performance on economic value knowledge tasks using real-world tasks crafted by industry experts.
GDR-learners: Orthogonal Learning of Generative Models for Potential Outcomes
Valentyn Melnychuk (LMU Munich), Stefan Feuerriegel (LMU Munich)
GenerationData SynthesisDiffusion modelFlow-based ModelAuto EncoderGenerative Adversarial NetworkImageTabularBenchmark
🎯 What it does: Proposed a new generative doubly robust learning framework (GDR-learners) for estimating the conditional distribution of potential outcomes (CDPOs)
Gelato: Graph Edit Distance via Autoregressive Neural Combinatorial Optimization
Paolo Pellizzoni (Max Planck Institute of Biochemistry), Karsten Borgwardt (Max Planck Institute of Biochemistry)
OptimizationGraph Neural NetworkGraph
🎯 What it does: Developed GELATO, a graph neural network-based autoregressive model that progressively predicts and constructs node matches for graph edit distance (GED), yielding an approximate GED solution;
GEM: A Gym for Generalist LLMs
Zichen Liu, Min Lin (Sea AI Lab)
Large Language ModelReinforcement LearningTextBenchmark
🎯 What it does: Proposes GEM (General Experience Maker) — a unified multi-round reinforcement learning (RL) environment framework for large language models (LLMs), providing baseline algorithms (REINFORCE+ReBN, PPO, GRPO, REINFORCE) and complete training scripts on 24 multi-round tasks, demonstrating seamless integration of GEM with five major RL frameworks.
Gen-DFL: Decision-Focused Generative Learning for Robust Decision Making
Prince Zizhuang Wang (Carnegie Mellon University), Shixiang Zhu (Carnegie Mellon University)
OptimizationFlow-based ModelGenerative Adversarial NetworkContrastive LearningTabularSequentialFinance Related
🎯 What it does: This paper proposes a generative model-based decision focusing learning framework, Gen-DFL, for robust decision-making in high-dimensional risk-sensitive scenarios.
gen2seg: Generative Models Enable Generalizable Instance Segmentation
Om Khangaonkar (University of California, Davis), Hamed Pirsiavash (University of California, Davis)
SegmentationDiffusion modelAuto EncoderImage
🎯 What it does: By refining pre-trained generative models (Stable Diffusion and MAE) using only a small amount of human-annotated indoor furniture and vehicle masks, a model capable of instance segmentation in zero-shot scenarios was constructed, enabling accurate segmentation of unseen object categories and image styles;
GenCape: Structure-Inductive Generative Modeling for Category-Agnostic Pose Estimation
Jiyong Rao (Tongji University), Shengjie Zhao (Tongji University)
Pose EstimationGraph Neural NetworkTransformerAuto EncoderImage
🎯 What it does: Proposes the GenCape framework, using a generative approach to learn instance-specific joint relationships, achieving category-agnostic pose estimation without skeletal priors.
GenCompositor: Generative Video Compositing with Diffusion Transformer
Shuzhou Yang (Peking University), Jian Zhang (Peking University)
Image HarmonizationGenerationData SynthesisTransformerDiffusion modelAuto EncoderVideo
🎯 What it does: Proposed and implemented the 'Generative Video Compositing' task, which can automatically inject dynamic foreground elements from other video sources into target videos based on user-defined controls such as trajectory and size, while maintaining background consistency.
GenCP: Towards Generative Modeling Paradigm of Coupled physics
Tianrun Gao (Westlake University), Tailin Wu (Westlake University)
Flow-based ModelBenchmarkPhysics Related
🎯 What it does: Propose GenCP, a generative coupled physics modeling paradigm that utilizes flow matching and operator splitting to transition from decoupled training to coupled inference.
GenCtrl -- A Formal Controllability Toolkit for Generative Models
Emily Cheng (Apple), Xavier Suau (Apple)
GenerationImageText
🎯 What it does: Proposes a control theory-based framework for theoretical analysis and empirical evaluation of controllability (reachability and controllability) in any black-box generative model (LLM, T2IM, etc.), and provides a PAC algorithm based on Monte Carlo sampling to estimate reachable and controllable sets.