International Conference on Learning Representations · 2207 papers
Flattery, Fluff, and Fog: Diagnosing and Mitigating Idiosyncratic Biases in Preference Models
Anirudh Bharadwaj (University of Pennsylvania), Mark Yatskar (New York University)
CodeData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Systematically evaluated the preference imbalance of language model preference models under five types of biases (length, structure, terminology, flattery, and vagueness), and reduced these biases through post-training using adversarial data augmentation (CDA).
🎯 What it does: Developed a dynamic neural audio codec called FlexiCodec that can achieve high-quality audio encoding at extremely low frame rates (3~12.5 Hz).
FlexRibbon: Joint Sequence and Structure Pretraining for Protein Modeling
Jianwei Zhu (Zhongguancun Academy), Tao Qin (Zhongguancun Academy)
CodeProtein Structure PredictionTransformerDiffusion modelBiomedical Data
🎯 What it does: Propose FlexRibbon, a protein foundation model that integrates sequence and three-dimensional structure, capable of simultaneously generating protein sequences and full-atom structures under single-sequence conditions;
🎯 What it does: Propose FLOCK, a knowledge graph foundation model that utilizes random walks and maintains probabilistic node-relation equivariance, for zero-shot link prediction.
🎯 What it does: Developed the floq method, parameterizing the Q-function as a time-varying velocity field, jointly trained with flow matching and TD targets, leveraging multi-step numerical integration to achieve scalable computational capacity;
Flow Actor-Critic for Offline Reinforcement Learning
Jongseong Chae (KAIST), Youngchul Sung (KAIST)
CodeFlow-based ModelImageBenchmark
🎯 What it does: Proposed a flow model integrated offline reinforcement learning algorithm called Flow Actor-Critic (FAC), which simultaneously utilizes the flow behavior proxy to regularize the actor and penalize the critic, and trains the first-order flow actor and dual-tower critic within the same architecture.
🎯 What it does: Propose a hybrid offline-online reinforcement learning method called FINO, combining flow matching with noise injection, which expands the policy distribution through noise injection and achieves efficient exploration via entropy-guided sampling.
Flow of Spans: Generalizing Language Models to Dynamic Span-Vocabulary via GFlowNets
Bo Xue (Shanghai Jiao Tong University), Zhouhan Lin (Shanghai Jiao Tong University)
CodeGenerationData SynthesisTransformerLarge Language ModelReinforcement LearningFlow-based ModelTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes FoSS, a span generation framework based on generative flow networks (GFlowNets), which constructs a directed acyclic graph (DAG) state space using a dynamic span vocabulary to achieve multi-path generation.
🎯 What it does: Proposed a single-step completion strategy (SSCP), achieving one-time action generation through flow-matching and completion vectors, and built SSCQL (offline RL) and GC-SSCP (goal-conditioned RL) algorithms based on this strategy.
🎯 What it does: Proposed Flow2GAN, a two-stage training framework that first learns generation capabilities using improved Flow Matching, then refines them with GAN to obtain high-quality audio generators for one, two, or four steps.
🎯 What it does: Developed FlowCast, a full-probability rainfall nowcasting model that leverages Conditional Flow Matching (CFM) to directly generate high-dimensional rainfall images from noise in a compressed latent space, achieving fast sampling.
🎯 What it does: Proposed a flow-based inverse problem solver named Flower, which generates posterior samples that satisfy measurement constraints through a three-step iterative process (flow-consistent target estimation, measurement-aware correction, and temporal progression) using a pre-trained flow model.
FLUX-Reason-6M & PRISM-Bench: A Million-Scale Text-to-Image Reasoning Dataset and Comprehensive Benchmark
Rongyao Fang (Chinese University of Hong Kong), Hongsheng Li (Chinese University of Hong Kong)
CodeData SynthesisVision Language ModelDiffusion modelImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Constructed the FLUX-Reason-6M dataset comprising 6 million images and 20 million bilingual descriptions, and designed the seven-track PRISM-Bench evaluation benchmark.
Fly-CL: A Fly-Inspired Framework for Enhancing Efficient Decorrelation and Reduced Training Time in Pre-trained Model-based Continual Representation Learning
Heming Zou (Tsinghua University), Xiangyang Ji (Tianjin University)
🎯 What it does: A continuous learning framework named Fly-CL based on the moth olfactory circuit was constructed, which uses sparse random projection and top-k operations to decorrelate features from a frozen pre-trained model, and then adopts streaming ridge regression for efficient class prototype learning.
🎯 What it does: Proposed a joint continuous-integer flow generation framework named FMIP for predicting high-quality solutions to mixed integer linear programming (MILP), incorporating a global guidance mechanism during inference;
🎯 What it does: This study, without altering the model structure, utilizes Direct Preference Optimization (DPO) and public reward models to construct preference data, aligning BrushNet and FLUX.1 Fill with preferences to improve image restoration effects;
🎯 What it does: Propose a training- and mask-free image editing framework that enables large-scale object shape transformation while preserving the background intact.
Fore-Mamba3D: Mamba-based Foreground-Enhanced Encoding for 3D Object Detection
Zhiwei Ning (Shanghai Jiao Tong University), Wei Liu (Shanghai Jiao Tong University)
CodeObject DetectionAutonomous DrivingPoint Cloud
🎯 What it does: Proposes Fore-Mamba3D, a 3D object detection backbone network focusing on foreground voxels, combining foreground sampling, region-to-global sliding window, and semantic-geometric fusion.
Forget Many, Forget Right: Scalable and Precise Concept Unlearning in Diffusion Models
Kaiyuan Deng (University of Arizona), Xiaolong Ma (University of Arizona)
CodeGenerationDiffusion modelImage
🎯 What it does: Propose the ScaPre framework to achieve large-scale concept forgetting, utilizing closed-form updates in diffusion models to precisely erase multiple target concepts.
Foundation Models for Causal Inference via Prior-Data Fitted Networks
Yuchen Ma (LMU Munich), Stefan Feuerriegel (LMU Munich)
CodeTransformerTabular
🎯 What it does: Construct and train a base model called CausalFM based on Prior-Data Fitted Networks (PFNs) for Bayesian inference in various causal inference scenarios (backdoor, frontdoor, and instrumental variable).
🎯 What it does: This paper proposes FOUNDAD, a few-shot anomaly detection method based on a foundational visual encoder, which identifies abnormal regions by mapping features back to the natural image manifold through a non-linear projector.
🎯 What it does: Propose an Spiking Neural Network framework f-SNN that replaces integer-order ODEs with fractional-order ODEs, leveraging fractional-order calculus to capture the long-term memory characteristics of neurons.
🎯 What it does: Propose FragFM, a hierarchical molecular graph generation framework that first generates molecules at the fragment level and then losslessly reconstructs them to the atomic level using a fine-grained autoencoder.
Frame Guidance: Training-Free Guidance for Frame-Level Control in Video Diffusion Models
Sangwon Jang (KAIST), Sung Ju Hwang (DeepAuto.ai)
CodeGenerationDiffusion modelAuto EncoderVideo
🎯 What it does: Propose a training-agnostic Frame Guidance framework that achieves controllable video generation by applying gradient guidance to a few frames, supporting multiple conditions such as keyframes, style, looping, depth, and sketches.
CodeComputational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposes the FreeKV framework, combining algorithmic and system-level improvements to achieve an efficient, training-free solution for KV cache retrieval.
Frequency-Balanced Retinal Representation Learning with Mutual Information Regularization
Seunghoon Lee (VUNO Inc), Doohyun Park (VUNO Inc)
CodeRepresentation LearningAuto EncoderImageBiomedical Data
🎯 What it does: Propose a retinal image pre-training framework named RetMAE, which learns frequency-balanced visual features by incorporating high-frequency mutual information (MI) regularization into the Masked Autoencoder (MAE).
FRIEDA: Benchmarking Multi-Step Cartographic Reasoning in Vision-Language Models
Jiyoon Pyo (University of Minnesota Twin Cities), Yao-Yi Chiang (University of Minnesota Twin Cities)
CodeLarge Language ModelVision Language ModelMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Constructed the FRIEDA benchmark, focusing on multi-map, multi-step map reasoning tasks to evaluate the ability of large vision-language models in reading geographic information charts.
From Assistant to Independent Developer — Are GPTs Ready for Software Development?
Dezhi Ran (Peking University), Tao Xie (Peking University)
CodeAI Code AssistantTransformerLarge Language ModelAgentic AITextBenchmark
🎯 What it does: This paper proposes and implements APPFORGE, a benchmark for evaluating the ability of large language models to complete real Android application development from scratch.
From atom to space: A region-based readout function for spatial properties of materials
Jiawen Zou (Fudan University), Bo Yan (Fudan University)
CodeGraph Neural NetworkTransformerGraphPhysics Related
🎯 What it does: Proposed a spatial node-based readout function called SpatialRead, improving the prediction of spatially separable attributes (e.g., gas adsorption, pore size) by graph neural networks.
CodeRepresentation 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 Concepts to Components: Concept-Agnostic Attention Module Discovery in Transformers
Jingtong Su (New York University), Karen Ullrich (Meta AI)
CodeExplainability 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)
CodeTransformerLarge 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 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)
CodeSafty 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 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)
CodeOptimizationComputational 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)
CodeExplainability 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 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)
CodeRepresentation 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)
CodeData 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)
CodeGenerationLarge 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.
🎯 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 Pixels to Words -- Towards Native Vision-Language Primitives at Scale
Haiwen Diao (Nanyang Technological University), Ziwei Liu (Nanyang Technological University)
CodeTransformerLarge 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 Reproduction to Replication: Evaluating Research Agents with Progressive Code Masking
Gyeongwon James Kim (Carnegie Mellon University), Daniel Fried (Carnegie Mellon University)
CodeAI 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 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)
CodeOptimizationBenchmark
🎯 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)
CodePose 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 Text to Talk: Audio-Language Model Needs Non-Autoregressive Joint Training
Tianqiao Liu (Jinan University), Zitao Liu (Peking University)
CodeGenerationTransformerLarge 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 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)
CodeGenerationTransformerLarge 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)
CodeGenerationData-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.
FrontierCO: Real-World and Large-Scale Evaluation of Machine Learning Solvers for Combinatorial Optimization
Shengyu Feng (Carnegie Mellon University), Yiming Yang (Carnegie Mellon University)
CodeOptimizationGraph 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.
Frozen Priors, Fluid Forecasts: Prequential Uncertainty for Low-Data Deployment with Pretrained Generative Models
Fernando Ruiz-Mazo (Aalto University), Vikas K Garg (Aalto University)
CodeGenerationData 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.
FS-KAN: Permutation Equivariant Kolmogorov-Arnold Networks via Function Sharing
Ran Elbaz (Technion Israel Institute of Technology), Haggai Maron (Technion Israel Institute of Technology)
CodeClassificationPoint 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;
🎯 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.
🎯 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);
Fusing Pixels and Genes: Spatially-Aware Learning in Computational Pathology
Minghao Han (Fudan University), Lihua Zhang (Fudan University)
CodeRepresentation LearningAuto EncoderContrastive LearningMultimodalityBiomedical Data
🎯 What it does: Proposed the STAMP framework, which performs multimodal representation learning by integrating histopathological images and spatial transcriptomics.
🎯 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.
🎯 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)
CodeComputational 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.
🎯 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.
CodeData 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)
CodeDomain 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.
🎯 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.
GarmentGPT: Compositional Garment Pattern Generation via Discrete Latent Tokenization
Fangsheng Weng (ChimerAI), Xiaoguang Han (Chinese University of Hong Kong)
CodeGenerationTransformerVision 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.
🎯 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.
GAVEL: Towards Rule-Based Safety through Activation Monitoring
Shir Rozenfeld (Ben Gurion University Negev), Yisroel Mirsky (Ben Gurion University Negev)
CodeSafty 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)
CodeGraph 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)
CodeGenerationData 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.
🎯 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)
CodeOptimizationGraph 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;
CodeLarge 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.
GenCtrl -- A Formal Controllability Toolkit for Generative Models
Emily Cheng (Apple), Xavier Suau (Apple)
CodeGenerationImageText
🎯 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.
Generalizable End-to-End Tool-Use RL with Synthetic CodeGym
Weihua Du (Carnegie Mellon University), Jiecao Chen (ByteDance)
CodeAI 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)
CodeOptimizationMeta 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 of RLVR Using Causal Reasoning as a Testbed
Brian Lu (Johns Hopkins University), Hongyuan Mei (Toyota Technological Institute at Chicago)
CodeExplainability 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 Spherical Neural Operators: Green’s Function Formulation
Hao Tang (University of Dundee), Chao Li (University of Dundee)
CodeConvolutional 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)
CodeRecommendation 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)
CodeGenerationData 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.
🎯 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.
🎯 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)
CodeGenerationReinforcement 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.
Andy Liu (Carnegie Mellon University), Max Kleiman-Weiner (University of Washington)
CodeExplainability 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.
GenSR: Symbolic regression based on equation generative space
Qian Li (Shanghai Jiao Tong University), Yuntian Chen (Imperial College London)
CodeGenerationOptimizationTransformerAuto 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.
CodeLarge 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;
GeoGramBench: Benchmarking the Geometric Program Reasoning in Modern LLMs
Shixian Luo (Li Auto Inc), Yong Wu (Li Auto Inc)
CodeAI 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 Constraints for Small Language Models to Understand and Expand Scientific Taxonomies
Liri Fang (University of Illinois Urbana Champaign), Vetle I Torvik (Meta Ai)
CodeKnowledge 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.
🎯 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.
Yuzhong Zhao (University of Chinese Academy of Sciences), Furu Wei (Microsoft Research)
CodeOptimizationLarge 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.
🎯 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.
GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning
Lakshya A Agrawal (UC Berkeley), Omar Khattab (MIT)
CodeOptimizationTransformerLarge 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)
CodeLarge 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;
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)
CodeTransformerLarge 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.
🎯 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: Hallucination-Inducing Image Generation for Multimodal LLMs
Aryan Yazdan Parast (University of Melbourne), NAVEED AKHTAR
CodeGenerationData 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).
CodeSafty 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;
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)
CodeGenerationLarge 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)
CodeAI 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;
🎯 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.
Global-Recent Semantic Reasoning on Dynamic Text-Attributed Graphs with Large Language Models
Yunan Wang (Beihang University), Ziwei Zhang (Beihang University)
CodeRecurrent 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;