NeurIPS 2025 Papers — Page 48
Conference on Neural Information Processing Systems · 5275 papers
Topology of Reasoning: Understanding Large Reasoning Models through Reasoning Graph Properties
Gouki Minegishi (University of Tokyo), Yutaka Matsuo (University of Tokyo)
Knowledge DistillationSupervised Fine-TuningGraph
🎯 What it does: By clustering the internal hidden states of large reasoning models to construct reasoning graphs, we systematically analyze their loops, diameters, and small-world characteristics, exploring the relationship between these graph structures and reasoning performance.
Topology-Aware Conformal Prediction for Stream Networks
Jifan Zhang (Northwestern University), Shixiang Zhu (Carnegie Mellon University)
Graph Neural NetworkGraphTime SeriesOrdinary Differential Equation
🎯 What it does: An adaptive consistency prediction framework STACI aimed at flow networks is designed, which implements multi-site joint uncertainty quantification using topology-aware inconsistency scores and adaptive confidence levels.
Topology-aware Graph Diffusion Model with Persistent Homology
Joonhyuk Park (POSTECH), Won Hwa Kim (Samsung Electronics)
GenerationData SynthesisGraph Neural NetworkDiffusion modelGraphAlzheimer's Disease
🎯 What it does: A topology information-based graph diffusion generation model TAGG is proposed, which encodes graphs using persistent homology and introduces topological constraints during the diffusion process.
Topology-Aware Learning of Tubular Manifolds via SE(3)-Equivariant Network on Ball B-Spline Curve
Jingxuan Wang (Beijing Normal University), Di Wang (Nanyang Technological University)
Graph Neural NetworkTransformerDiffusion modelBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A continuous tubular manifold representation based on spherical B-spline curves (BBSC) is proposed, and an SE(3)-BBSCformerGCN framework is constructed by combining SE(3)-equivariant networks with graph convolutional networks for learning geometric and topological features of tubular structures.
TopoPoint: Enhance Topology Reasoning via Endpoint Detection in Autonomous Driving
Yanping Fu (Institute of Computing Technology, Chinese Academy of Sciences), Feng Dai (Institute of Computing Technology, Chinese Academy of Sciences)
Autonomous DrivingGraph Neural NetworkTransformerPoint Cloud
🎯 What it does: This paper proposes a new framework called TopoPoint, which explicitly detects lane endpoints and jointly infers lane information, significantly improving the accuracy of road topology inference.
Tortoise and Hare Guidance: Accelerating Diffusion Model Inference with Multirate Integration
Yunghee Lee (Agency for Defense Development), Hoseong Kim (Agency for Defense Development)
GenerationComputational EfficiencyDiffusion modelImageOrdinary Differential EquationAudio
🎯 What it does: By accelerating diffusion model inference through multi-rate integration, a training-independent Tortoise and Hare Guidance (THG) method is proposed, significantly reducing the number of function evaluations while maintaining generation quality.
Touch in the Wild: Learning Fine-Grained Manipulation with a Portable Visuo-Tactile Gripper
Xinyue Zhu (Columbia University), Yunzhu Li (Columbia University)
Robotic IntelligenceConvolutional Neural NetworkTransformerDiffusion modelContrastive LearningImageMultimodality
🎯 What it does: A lightweight handheld visual-tactile gripper was designed, capable of synchronously collecting RGB images and flexible tactile signals, constructing a multimodal dataset of approximately 2.6 million frames, and achieving a transferable visual-tactile encoder through cross-modal self-supervised pre-training, which significantly improved performance in fine-grained robotic manipulation tasks.
Toward a Unified Geometry Understanding : Riemannian Diffusion Framework for Graph Generation and Prediction
Yisen Gao (Guangxi Normal University), Xianxian LI
GenerationData SynthesisDrug DiscoveryGraph Neural NetworkTransformerDiffusion modelGraph
🎯 What it does: This paper presents GeoMancer, a unified Riemannian geometric diffusion framework for graph generation and prediction tasks.
Toward Artificial Palpation: Representation Learning of Touch on Soft Bodies
Zohar Rimon (Technion Israel Institute of Technology), Aviv Tamar (Technion Israel Institute of Technology)
Anomaly DetectionRepresentation LearningRecurrent Neural NetworkContrastive LearningImageBiomedical DataMagnetic Resonance ImagingBenchmark
🎯 What it does: A self-supervised learning-based artificial palpation framework is proposed, which learns to encode tactile sequences of soft objects and can be used for shape reconstruction and change detection.
Toward Efficient Inference Attacks: Shadow Model Sharing via Mixture-of-Experts
Li Bai (Hong Kong Polytechnic University), Haibo Hu (Hong Kong Polytechnic University)
Safty and PrivacyComputational EfficiencyConvolutional Neural NetworkMixture of ExpertsImage
🎯 What it does: A shared Shadow model pool SHAPOOL based on Mixture-of-Experts is proposed to reduce the training cost of Shadow models.
Toward Human Deictic Gesture Target Estimation
Xu Cao (University of Illinois Urbana-Champaign), James Matthew Rehg
RecognitionObject DetectionTransformerImageText
🎯 What it does: This paper proposes a Transformer-based model called TransGesture, which predicts the target objects of human pointing gestures from images and constructs a large-scale GestureTarget dataset.
Toward Interpretable Evaluation Measures for Time Series Segmentation
Félix Chavelli, Michaël Thomazo
SegmentationExplainability and InterpretabilityTime Series
🎯 What it does: An interpretable time series segmentation evaluation method is proposed, addressing the shortcomings of traditional metrics that fail to capture the location and type of errors, with the design of two new metrics, WARI and SMS.
Toward Relative Positional Encoding in Spiking Transformers
Changze Lv (Fudan University), Dongsheng Li (Microsoft Research Asia)
Spiking Neural NetworkTransformerImageTextTime Series
🎯 What it does: Two relative position encoding methods, Gray-PE and Log-PE, are proposed and applied to the spiking Transformer, improving the self-attention mechanism to XNOR logic.
Towards 3D Objectness Learning in an Open World
Taichi Liu (Rutgers University), Desheng Zhang (Rutgers University)
Object DetectionTransformerMixture of ExpertsImagePoint Cloud
🎯 What it does: This paper presents OP3Det, the first model for class-agnostic 3D object detection in open-world environments.
Towards a General Attention Framework on Gyrovector Spaces for Matrix Manifolds
Rui Wang (Jiangnan University), Ziheng Chen (University of Trento)
TransformerTime SeriesBiomedical DataBenchmark
🎯 What it does: The GyroAtt framework is proposed, which extends the self-attention mechanism to the general gyrovector space, achieving a unified treatment of various matrix manifolds such as SPD, SPSD, and Grassmannian.
Towards a Geometric Understanding of Tensor Learning via the t-Product
Andong Wang (RIKEN), Qibin Zhao (RIKEN)
RestorationOptimizationImageVideo
🎯 What it does: A geometric framework centered on t-product (t-manifold) is proposed, providing a unified theory from differential geometry to tensor algebra for high-dimensional tensor learning.
Towards a Golden Classifier-Free Guidance Path via Foresight Fixed Point Iterations
Kaibo Wang (Hong Kong University of Science and Technology), Yang Xiang (Hong Kong University of Science and Technology)
GenerationOptimizationImage
🎯 What it does: This paper views conditional guidance as a calibration process for the 'golden path' and provides a unified explanation of CFG and its variants through fixed-point iteration.
Towards a Pairwise Ranking Model with Orderliness and Monotonicity for Label Enhancement
Yunan Lu (Nanjing University of Science and Technology), Xiuyi Jia (Nanjing University of Aeronautics and Astronautics)
Auto EncoderGenerative Adversarial NetworkTabularBiomedical Data
🎯 What it does: A pairwise ranking model called PROM that satisfies probability monotonicity and order is proposed, and a generative label enhancement algorithm LE-PROM is constructed based on this model to directly learn label distributions from multi-label data.
Towards A Translative Model of Sperm Whale Vocalization
Orr Paradise (University of California Berkeley), Shafi Goldwasser (University of California Berkeley)
ClassificationGenerationDomain AdaptationTransformerSupervised Fine-TuningAudio
🎯 What it does: WhAM is proposed, a transformer-based model capable of generating and converting any audio into the audio style of whale coda, and providing classification functionality for coda.
Towards Accurate Time Series Forecasting via Implicit Decoding
Xinyu Li (University of Melbourne), Mingming Gong (University of Melbourne)
Recurrent Neural NetworkTransformerTime Series
🎯 What it does: Improved the decoding phase of time series forecasting by proposing the Implicit Forecaster module, which implicitly predicts future sequences using frequency waveforms.
Towards Better & Faster Autoregressive Image Generation: From the Perspective of Entropy
Xiaoxiao Ma (University of Science and Technology of China), Lin Ma (Meituan)
GenerationComputational EfficiencyImage
🎯 What it does: This study investigates the sampling problem in autoregressive image generation and proposes an entropy-based dynamic temperature control and entropy-aware inference acceleration strategy.
Towards Building Model/Prompt-Transferable Attackers against Large Vision-Language Models
Xiaowen Cai (Huazhong University of Science and Technology), Wei Hu (Peking University)
Adversarial AttackLarge Language ModelVision Language ModelImageMultimodality
🎯 What it does: A method for generating adversarial examples that can cross models and prompts is proposed for large visual-language models.
Towards Comprehensive Scene Understanding: Integrating First and Third-Person Views for LVLMs
Insu Lee (Seoul National University), Byonghyo Shim (Seoul National University)
RecognitionSegmentationData SynthesisTransformerLarge Language ModelVision Language ModelImageVideoMultimodalityBenchmarkChain-of-Thought
🎯 What it does: This paper proposes E3VQA, a visual question answering benchmark based on synchronized front and rear views, and designs M3CoT, an unsupervised multi-view scene graph fusion reasoning method, to enhance the understanding and reasoning capabilities of large visual language models (LVLM) when integrating first-person and third-person images.
Towards Doctor-Like Reasoning: Medical RAG Fusing Knowledge with Patient Analogy through Textual Gradients
Yuxing Lu (Peking University), Jinzhuo Wang (Peking University)
GenerationRetrievalOptimizationTransformerLarge Language ModelTextBiomedical DataElectronic Health RecordsRetrieval-Augmented Generation
🎯 What it does: The DoctorRAG framework is proposed, integrating medical knowledge bases and similar case experiences to achieve reasoning similar to that of doctors, and iteratively optimizing answers through multi-agent text gradient (Med-TextGrad).
Towards Dynamic 3D Reconstruction of Hand-Instrument Interaction in Ophthalmic Surgery
Ming Hu (Monash University), Zongyuan Ge (Monash University)
SegmentationPose EstimationVideoPoint CloudBenchmark
🎯 What it does: This study constructed the first RGB-D dynamic 3D reconstruction dataset for ophthalmic microsurgery, OphNet-3D, and proposed an automatic annotation process along with two benchmark tasks.
Towards Effective Federated Graph Foundation Model via Mitigating Knowledge Entanglement
Yinlin Zhu (Sun Yat-sen University), Meikang Qiu (Augusta University)
Federated LearningGraph Neural NetworkSupervised Fine-TuningPrompt EngineeringGraphBiomedical Data
🎯 What it does: Proposes the FedGFM+ framework, which trains graphical foundation models using federated pre-training and fine-tuning, and eliminates knowledge entanglement through two mechanisms: Anchor-based Domain-Aware Initialization and Adaptive Domain-Sensitive Prompt Pool, enhancing cross-domain and cross-task generalization capabilities.
Towards foundational LiDAR world models with efficient latent flow matching
Tianran Liu (University of Toronto), Nicholas Rhinehart (University of Toronto)
SegmentationCompressionDomain AdaptationAutonomous DrivingTransformerDiffusion modelFlow-based ModelAuto EncoderPoint Cloud
🎯 What it does: This paper proposes a transferable foundational LiDAR world model that can achieve performance improvements in various downstream prediction tasks (sparse to dense beam adaptation, indoor-outdoor transfer, semantic occupancy prediction) with less labeled data, and significantly enhances training and inference efficiency through compression and flow matching techniques.
Towards Fully FP8 GEMM LLM Training at Scale
Alejandro Hernández-Cano (École Polytechnique Fédérale de Lausanne), Martin Jaggi (École Polytechnique Fédérale de Lausanne)
TransformerLarge Language ModelText
🎯 What it does: Proposed and implemented the FOG architecture, achieving complete FP8 GEMM training within the Transformer block, including the attention mechanism, significantly enhancing the throughput of large-scale LLM training.
Towards General Continuous Memory for Vision-Language Models
Wenyi WU, Biwei Huang (University of California)
RetrievalCompressionTransformerVision Language ModelMultimodalityRetrieval-Augmented Generation
🎯 What it does: A pluggable continuous memory module CoMEM is designed for visual-language models (VLM), using the VLM itself as a memory encoder, and utilizing a small amount of self-synthesized data with LoRA fine-tuning, only increasing parameters by 1.2%.
Towards General Modality Translation with Contrastive and Predictive Latent Diffusion Bridge
Nimrod Berman (Bosch AI Center), Omri Azencot (Technical University of Munich)
Image TranslationGenerationData SynthesisTransformerDiffusion modelAuto EncoderContrastive LearningImageMultimodalityAudio
🎯 What it does: A general modal translation framework LDDBM based on Latent Diffusion Bridge is proposed, which can achieve information conversion between modalities of different dimensions and structures.
Towards Generalizable 3D Human Pose Estimation via Ensembles on Flat Loss Landscapes
Jumin Han (Korea University), Seong-Whan Lee (Korea University)
Pose EstimationConvolutional Neural NetworkGraph Neural NetworkTransformerSupervised Fine-TuningImage
🎯 What it does: By introducing an adaptive scaling mechanism in 3D human pose estimation, the loss surface is smoothed, and multiple solutions are integrated on the flat surface to enhance generalization ability.
Towards Generalizable Detector for Generated Image
Qianshu Cai (University of Science and Technology of China), Xinmei Tian (University of Science and Technology of China)
Object DetectionAnomaly DetectionTransformerContrastive LearningImage
🎯 What it does: This paper proposes viewing the detection of generated images as OOD detection and designs a training-free detection framework called DEnD based on the differential energy of self-supervised models.
Towards Generalizable Multi-Policy Optimization with Self-Evolution for Job Scheduling
Inguk Choi (Korea Advanced Institute of Science and Technology), Hyun-Jung Kim (Korea Advanced Institute of Science and Technology)
OptimizationGraph Neural NetworkTransformerReinforcement LearningTabular
🎯 What it does: The MP-ASIL framework is proposed to improve the solution of job scheduling problems through multi-strategy self-supervised learning.
Towards Generalizable Retina Vessel Segmentation with Deformable Graph Priors
Ke Liu (Zhejiang University), Shangqi Gao (University of Cambridge)
SegmentationGraph Neural NetworkGaussian SplattingImageBiomedical Data
🎯 What it does: A variational Bayesian framework named GraphSeg is proposed, which combines deformable shape priors to achieve generalizable retinal vessel segmentation.
Towards Graph Foundation Models: Training on Knowledge Graphs Enables Transferability to General Graphs
Kai Wang (Nanyang Technological University), Yifei Shen (Microsoft Research Asia)
ClassificationRepresentation LearningGraph Neural NetworkGraph
🎯 What it does: Pre-training zero-shot inference on knowledge graphs, mapping node/edge/graph-level tasks to inference tasks through a unified graph knowledge graph format, proposing the Semantic Conditional Message Passing (SCMP) mechanism to achieve joint representation of structure and semantics, and constructing a graph-based model SCR that can be directly transferred to multi-domain graph tasks.
Towards Identifiability of Hierarchical Temporal Causal Representation Learning
Zijian Li (Carnegie Mellon University), Kun Zhang (Carnegie Mellon University)
GenerationRepresentation LearningFlow-based ModelAuto EncoderTime SeriesMagnetic Resonance ImagingFinance Related
🎯 What it does: This study investigates the identifiability of hierarchical latent variables in time series and proposes a recognition framework and generative model based on Causal Hierarchical Latent Dynamics (CHiLD).
Towards Implicit Aggregation: Robust Image Representation for Place Recognition in the Transformer Era
Feng Lu (Tsinghua University), Chun Yuan (Tsinghua University)
RecognitionRetrievalTransformerImage
🎯 What it does: We propose an implicit aggregation method called ImAge that generates global descriptors in the Transformer backbone using learnable aggregation tokens without the need for explicit aggregators.
Towards Interpretability Without Sacrifice: Faithful Dense Layer Decomposition with Mixture of Decoders
James Oldfield (University of Wisconsin-Madison), Grigorios Chrysos (Cyprus Institute)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: This paper proposes a hierarchical sparsification framework called Mixture of Decoders (MxD) to decompose the MLP layers in large language models without sacrificing accuracy.
Towards Interpretable and Efficient Attention: Compressing All by Contracting a Few
Qishuai Wen (Beijing University of Posts and Telecommunications), Chun-Guang Li (Beijing University of Posts and Telecommunications)
ClassificationSegmentationExplainability and InterpretabilityComputational EfficiencyTransformerAuto EncoderImage
🎯 What it does: An optimization-target-based attention mechanism called Contract-and-Broadcast Self-Attention (CBSA) is proposed, achieving a unification of interpretability and efficiency, and linearizing the attention computation complexity through representative token compression.
Towards Irreversible Attack: Fooling Scene Text Recognition via Multi-Population Coevolution Search
Jingyu Li (Sun Yat-sen University), Xiaochun Cao (Sun Yat-sen University)
RecognitionOptimizationAdversarial AttackRecurrent Neural NetworkImage
🎯 What it does: A black-box pixel-level attack method for scene text recognition models is proposed—Multi-Population Co-evolutionary Search (MPCS), which can cause the model to predict more incorrect characters while maintaining visual semantic integrity.
Towards Large-Scale In-Context Reinforcement Learning by Meta-Training in Randomized Worlds
Fan Wang (Shenzhen Institute of Artificial Intelligence and Robotics for Society), Haifeng Wang (Baidu Inc)
Meta LearningTransformerReinforcement LearningSequential
🎯 What it does: Proposes the AnyMDP task generator and the OmniRL framework to achieve large-scale ICRL training.
Towards Minimizing Feature Drift in Model Merging: Layer-wise Task Vector Fusion for Adaptive Knowledge Integration
Wenju Sun (Beijing Jiaotong University), Boyang Li (Nanyang Technological University)
OptimizationTransformerImageMultimodality
🎯 What it does: This paper proposes the Layer-wise Optimal Task Vector Merging (LOT Merging) method, which achieves multi-task model merging by minimizing feature drift at the layer level.
Towards Multi-Table Learning: A Novel Paradigm for Complementarity Quantification and Integration
Junyu Zhang (Beijing Institute of Technology), Changsheng Li (Beijing Institute of Technology)
ClassificationRecommendation SystemTransformerTabularBiomedical DataFinance Related
🎯 What it does: A unified paradigm for multi-table learning is proposed, systematically quantifying inter-table complementarity and designing ATCA-Net to achieve adaptive table encoding and cross-table attention fusion.
Towards Multiscale Graph-based Protein Learning with Geometric Secondary Structural Motifs
Shih-Hsin Wang (University of Utah), Bao Wang (University of Utah)
Protein Structure PredictionGraph Neural NetworkGraphBiomedical Data
🎯 What it does: This paper proposes a multi-scale hierarchical graph representation constructed based on protein secondary structures (α-helix, β-sheet, loop, etc.), and designs a two-stage graph neural network on this basis for efficient learning of protein structure and function.
Towards Physical Understanding in Video Generation: A 3D Point Regularization Approach
Yunuo Chen (University of California), Anil Kag (Snap Inc)
GenerationData SynthesisDiffusion modelVideoPoint Cloud
🎯 What it does: By incorporating sparse 3D point trajectories into video generation models, this approach achieves modeling and constraints of 3D shapes and motions, significantly reducing deformation and object distortion.
Towards Physics-informed Spatial Intelligence with Human Priors: An Autonomous Driving Pilot Study
Guanlin Wu (Johns Hopkins University), Hao Frank Yang (Johns Hopkins University)
Autonomous DrivingTransformerLarge Language ModelImageMultimodalityBenchmark
🎯 What it does: Design and evaluate a grid-based visual-spatial intelligence representation (SIG) and its evaluation metrics, construct the SIGBench benchmark, and conduct zero-shot and few-shot inference experiments using multimodal large language models in autonomous driving scenarios.
Towards Pre-trained Graph Condensation via Optimal Transport
Yeyu Yan (Beijing Jiaotong University), Kunlun He (Chinese PLA General Hospital)
CompressionExplainability and InterpretabilityGraph Neural NetworkGraph
🎯 What it does: A pre-training graph condensation method (PreGC) is proposed, which can generate reusable compressed graphs without relying on specific tasks and model architectures.
Towards Predicting Any Human Trajectory In Context
Ryo Fujii, Ryo Hachiuma
Object TrackingTransformerVideo
🎯 What it does: The TrajICL framework is proposed, achieving pedestrian trajectory prediction in different scenarios without the need for fine-tuning.
Towards Principled Unsupervised Multi-Agent Reinforcement Learning
Riccardo Zamboni (Politecnico di Milano), Marcello Restelli (Politecnico di Milano)
Reinforcement Learning
🎯 What it does: Proposes a multi-agent unsupervised pre-training framework based on state entropy maximization, defining three objectives: joint, marginal, and mixed;
Towards Prospective Medical Image Reconstruction via Knowledge-Informed Dynamic Optimal Transport
Taoran Zheng (Xi'an Jiaotong University), Zongben Xu (Xi'an Jiaotong University)
RestorationOptimizationRecurrent Neural NetworkAuto EncoderImageBiomedical DataMagnetic Resonance ImagingComputed TomographyOrdinary Differential Equation
🎯 What it does: The KIDOT framework is proposed, modeling medical image reconstruction as a dynamic optimal transport process constrained by imaging physics, and learning to reconstruct from unpaired data through neural networks;
Towards Provable Emergence of In-Context Reinforcement Learning
Jiuqi Wang (University of Virginia), Shangtong Zhang (University of Virginia)
TransformerReinforcement LearningTabular
🎯 What it does: The study shows that after pre-training with reinforcement learning, the Transformer model can achieve TD learning through context, and proves that its parameters are the global optimal solution for the pre-training loss.
Towards Realistic Earth-Observation Constellation Scheduling: Benchmark and Methodology
Luting Wang (Beihang University), Si Liu (Beihang University)
OptimizationTransformerReinforcement LearningBenchmark
🎯 What it does: The AEOS-Bench benchmark and AEOS-Former scheduling model are proposed for real AEOS constellation scheduling problems, supporting large-scale, dynamic, and constrained scenarios.
Towards Reliable and Holistic Visual In-Context Learning Prompt Selection
Wenxiao Wu (Huazhong University of Science and Technology), Yanwei Fu (Fudan University)
Object DetectionSegmentationPrompt EngineeringImage
🎯 What it does: Proposes RH-Partial2Global, improving the reliability and comprehensiveness of prompt selection in Visual Context Learning (VICL).
Towards Reliable Code-as-Policies: A Neuro-Symbolic Framework for Embodied Task Planning
Sanghyun Ahn (Sungkyunkwan University), Honguk Woo (Sungkyunkwan University)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningMultimodality
🎯 What it does: The NESYRO framework is constructed, combining neural networks and symbolic reasoning, recursively executing code verification and interactive validation, using LLM to generate executable code and actively detect missing observations, enhancing the reliability of robotic task planning.
Towards Reliable Identification of Diffusion-based Image Manipulations
Alex Costanzino, Fabio Pizzati
Anomaly DetectionData-Centric LearningTransformerDiffusion modelContrastive LearningImageBenchmark
🎯 What it does: This paper proposes a method for image forgery detection and localization specifically targeting diffusion model inpainting attacks, called RADAR.
Towards Reliable LLM-based Robots Planning via Combined Uncertainty Estimation
Shiyuan Yin (Henan University of Technology), Xuelong Li (China Telecom)
Knowledge DistillationRobotic IntelligenceTransformerLarge Language ModelText
🎯 What it does: The CURE framework is proposed to provide fine-grained uncertainty estimation for robot planning generated by large language models (LLMs) to enhance the reliability of planning.
Towards Resilient Safety-driven Unlearning for Diffusion Models against Downstream Fine-tuning
Boheng Li (Nanyang Technological University), Tianwei Zhang (Nanyang Technological University)
GenerationMeta LearningDiffusion modelImage
🎯 What it does: This paper proposes ResAlign, a security-driven robust zero-shot framework for text-to-image diffusion models, aimed at maintaining model safety after downstream fine-tuning.
Towards Robust Parameter-Efficient Fine-Tuning for Federated Learning
Xiuwen Fang (Wuhan University), Mang Ye (Wuhan University)
Federated LearningTransformerSupervised Fine-TuningImage
🎯 What it does: This paper proposes a robust parameter-efficient fine-tuning framework (RFedLR) specifically designed for scenarios with label noise in federated learning, combining sensitivity-aware robust tuning (SRT) and adaptive LoRA aggregation (AFLA);
Towards Robust Pseudo-Label Learning in Semantic Segmentation: An Encoding Perspective
Wangkai Li (University of Science and Technology of China), Tianzhu Zhang (University of Science and Technology of China)
SegmentationDomain AdaptationContrastive LearningImage
🎯 What it does: In response to pseudo-label learning in semantic segmentation, the authors propose using Error-Correcting Output Codes (ECOC) for fine-grained multi-bit binary encoding of categories, thereby enhancing the robustness of pseudo-labels and improving model training.
Towards Robust Uncertainty Calibration for Composed Image Retrieval
Yifan Wang (Tsinghua University), Chun Yuan (Tsinghua University)
RetrievalTransformerVision Language ModelContrastive LearningImageText
🎯 What it does: A robust uncertainty calibration framework RUNC is proposed to improve image retrieval in dual-modal queries.
Towards Robust Zero-Shot Reinforcement Learning
Kexin ZHENG, Xianyuan Zhan (Tsinghua University)
TransformerReinforcement LearningDiffusion modelMultimodality
🎯 What it does: A new zero-shot RL method based on the Forward-Backward (FB) framework, called BREEZE, is proposed to address issues such as scale inconsistency, bias, and outlier estimation in FB methods.
Towards Self-Refinement of Vision-Language Models with Triangular Consistency
Yunlong Deng (Mohamed bin Zayed University of Artificial Intelligence), Kun Zhang (Carnegie Mellon University)
GenerationData SynthesisOptimizationTransformerSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: A self-improvement framework based on the principle of triangular consistency is proposed, which automatically generates high-quality image-question-answer (IQA) pairs from unlabeled images and iteratively fine-tunes the visual language model (VLM) based on this.
Towards Single-Source Domain Generalized Object Detection via Causal Visual Prompts
Chen Li (Huazhong University of Science and Technology), Xinzhong Zhu (Zhejiang Normal University)
Object DetectionDomain AdaptationPrompt EngineeringImage
🎯 What it does: The Cauvis method is proposed, which achieves reverse causal adjustment in single-source domain generalization object detection through visual prompts and cross-attention, and incorporates a dual-branch adapter to decouple causal features from high-frequency domain features.
Towards Straggler-Resilient Split Federated Learning: An Unbalanced Update Approach
Dandan Liang (Rochester Institute of Technology), Haibo Yang (Rochester Institute of Technology)
OptimizationFederated LearningConvolutional Neural NetworkSupervised Fine-TuningImageText
🎯 What it does: This paper proposes MU-SplitFed, a method to alleviate stragglers in Split Federated Learning through server-side unbalanced updates and zeroth-order optimization, significantly reducing the number of communication rounds and decoupling training time from the slowest client.
Towards Syn-to-Real IQA: A Novel Perspective on Reshaping Synthetic Data Distributions
Aobo Li (Xidian University), Weisheng Dong (Xidian University)
Data SynthesisDomain AdaptationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a new framework called SynDR-IQA, which enhances the cross-domain generalization ability of no-reference image quality assessment (BIQA) models by reshaping the distribution of synthetic data.
Towards the Resistance of Neural Network Fingerprinting to Fine-tuning
Ling Tang (Shanghai Jiao Tong University), Quanshi Zhang (Shanghai Jiao Tong University)
RecognitionConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a network fingerprint embedding method based on the frequency components of convolution kernels, theoretically proving that these components remain unchanged during the fine-tuning process, thereby achieving robust protection of model copyrights.
Towards Thinking-Optimal Scaling of Test-Time Compute for LLM Reasoning
Wenkai Yang (Renmin University of China), Furu Wei (Microsoft Research)
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkChain-of-Thought
🎯 What it does: This paper studies the test-time scalability of chain-of-thought (CoT) length in reasoning tasks using large language models, finding that overly long CoTs can lead to a decline in reasoning performance. It then proposes the Thought Optimal Expansion (TOPS) strategy, allowing the model to determine the necessary CoT length on its own and achieve more efficient and effective System 2 reasoning through self-improvement. Ultimately, it achieves better performance than existing distilled o1 models on multiple mathematical reasoning benchmarks.
Towards Understanding Safety Alignment: A Mechanistic Perspective from Safety Neurons
Jianhui Chen (Tsinghua University), Juanzi Li (Tsinghua University)
Safty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Through mechanism interpretability methods, identify and validate safety neurons in large language models, revealing the competitive relationship between safety and usefulness, and construct safety protections based on these neurons to preemptively detect harmful outputs.
Towards Understanding the Mechanisms of Classifier-Free Guidance
Xiang Li (University of Michigan), Qing Qu (Michigan State University)
GenerationDiffusion modelContrastive LearningImage
🎯 What it does: This paper reveals the core mechanism of Classifier-Free Guidance (CFG) in improving generation quality through an analytical approach in a simplified linear diffusion model, and validates this analysis on a real nonlinear diffusion model.
Towards Understanding Transformers in Learning Random Walks
Wei Shi (University of Hong Kong), Yuan Cao (University of Hong Kong)
TransformerGraph
🎯 What it does: This paper theoretically studies the performance of a single-layer Transformer in learning the circular random walk task, proving that optimal predictions can be achieved under gradient descent training, and providing interpretability of self-attention and value matrices; it also explores the failures caused by edge cases (p=0 or 1) and validates the theoretical conclusions through experiments.
Towards Unified and Lossless Latent Space for 3D Molecular Latent Diffusion Modeling
Yanchen Luo (University of Science and Technology of China), Xiang Wang (University of Science and Technology of China)
GenerationData SynthesisDrug DiscoveryTransformerDiffusion modelAuto EncoderMultimodality
🎯 What it does: This paper proposes the UAE-3D multimodal VAE and UDM-3D joint latent diffusion model, which compresses the originally separate invariant and equivariant modalities into the same latent space, achieving efficient 3D molecular generation.
Towards Unified Multimodal Interleaved Generation via Group Relative Policy Optimization
Ming Nie (Fudan University), Li Zhang (Fudan University)
GenerationOptimizationReinforcement LearningVision Language ModelImageTextMultimodality
🎯 What it does: Through the warm-up phase and training based on GRPO reinforcement learning, the unified visual language model is able to perform multimodal interactive generation (text and images appearing alternately) in a high-quality and coherent manner.
Towards Unsupervised Domain Bridging via Image Degradation in Semantic Segmentation
Wangkai Li (University of Science and Technology of China), Tianzhu Zhang (University of Science and Technology of China)
SegmentationDomain AdaptationConvolutional Neural NetworkTransformerDiffusion modelImage
🎯 What it does: This paper proposes DiDA, an unsupervised domain bridging framework based on image degradation, aimed at enhancing the cross-domain generalization ability of semantic segmentation.
Towards Unsupervised Open-Set Graph Domain Adaptation via Dual Reprogramming
Zhen Zhang (Nanjing University), Bingsheng He (National University of Singapore)
Domain AdaptationGraph Neural NetworkGraph
🎯 What it does: This study investigates the problem of unsupervised open set graph domain adaptation, where the target graph contains new categories that do not exist in the source graph.
Towards Unsupervised Training of Matching-based Graph Edit Distance Solver via Preference-aware GAN
Wei Huang (University of New South Wales), Xuemin Lin (Shanghai Jiaotong University)
Graph Neural NetworkDiffusion modelGenerative Adversarial NetworkGraph
🎯 What it does: An unsupervised training framework for Graph Edit Distance (GED) called GEDRanker is proposed, which utilizes a GAN discriminator to guide the matching model in generating high-quality node matching matrices and can recover the edit paths.
Towards Visualization-of-Thought Jailbreak Attack against Large Visual Language Models
Hongqiong Zhong (Alibaba Group), Kaifu Zhang (Alibaba Group)
Adversarial AttackLarge Language ModelPrompt EngineeringVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: An automated jailbreak framework named Visualization-of-Thought Attack (VoTA) is proposed, which induces VLM to generate unsafe content through image sequences.
ToxicTextCLIP: Text-Based Poisoning and Backdoor Attacks on CLIP Pre-training
Xin Yao (Central South University), Ming Zhao (Central South University)
Adversarial AttackTransformerVision Language ModelImageText
🎯 What it does: Proposes the ToxicTextCLIP framework, which achieves control over the model during the CLIP pre-training phase through text poisoning and backdoor attacks, mainly including a background-aware text selector and a background-driven text enhancer.
TP-MDDN: Task-Preferenced Multi-Demand-Driven Navigation with Autonomous Decision-Making
Shanshan Li (Fudan University), Xiangyang Xue (Fudan University)
Autonomous DrivingRobotic IntelligenceTransformerLarge Language ModelAgentic AITextMultimodalityPoint CloudBenchmark
🎯 What it does: Proposed the TP-MDDN multi-demand task instruction benchmark, and designed AWMSystem, MASMap, dual-rhythm action generator, and adaptive error corrector to achieve long-term navigation with autonomous decision-making;
TPP-SD: Accelerating Transformer Point Process Sampling with Speculative Decoding
Shukai Gong (Renmin University of China), Feng Zhou (Renmin University of China)
GenerationComputational EfficiencyTransformerLarge Language ModelTime SeriesSequential
🎯 What it does: The TPP-SD method is proposed, which accelerates event generation by introducing speculative decoding in the sampling of Transformer temporal point processes.
TRACE: Contrastive learning for multi-trial time series data in neuroscience
Lisa Schmors (Hertie Institute for AI in Brain Health), Philipp Berens (Hertie Institute for AI in Brain Health)
Anomaly DetectionRepresentation LearningContrastive LearningTime SeriesBiomedical Data
🎯 What it does: This paper studies a contrastive learning visualization framework TRACE that utilizes multi-trial time series neural data.
TRACE: Grounding Time Series in Context for Multimodal Embedding and Retrieval
Jialin Chen (Yale University), Rex Ying (Yale University)
ClassificationRetrievalTransformerContrastive LearningMultimodalityTime SeriesRetrieval-Augmented Generation
🎯 What it does: TRACE is proposed, a multimodal retriever capable of dual-layer alignment between multivariate time series and corresponding textual descriptions, and it can also serve as a powerful encoder for prediction and classification tasks.
Tracing Back the Malicious Clients in Poisoning Attacks to Federated Learning
Yuqi Jia (Duke University), Neil Zhenqiang Gong
Federated LearningAdversarial AttackImageText
🎯 What it does: A method for tracing malicious clients in federated learning after deployment, called FLForensics, is proposed, which can detect misclassified target inputs and locate the malicious clients responsible for those misclassifications after the model is deployed.
Tracing the Representation Geometry of Language Models from Pretraining to Post-training
Melody Zixuan Li (McGill University), Blake Aaron Richards
Representation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Using spectral analysis methods, we quantify the geometric evolution of representations in large language models during pre-training and post-training phases, discovering and explaining three non-monotonic geometric stages (warmup, entropy-seeking, compression-seeking), and relate them to short-term memory, long-range generalization, and the performance of different post-training strategies (SFT, DPO, RLVR).
Tracing the Roots: Leveraging Temporal Dynamics in Diffusion Trajectories for Origin Attribution
Andreas Floros (Imperial College London), Pier Luigi Dragotti (Imperial College London)
ClassificationGenerationDiffusion modelImage
🎯 What it does: Conduct a temporal analysis of the complete trajectories of diffusion models, constructing a unified framework for member inference and model attribution; determine whether an image is a member of the model training set, a model-generated sample, or external data by utilizing the temporal dynamics of the trajectories.
Track, Inpaint, Resplat: Subject-driven 3D and 4D Generation with Progressive Texture Infilling
Shuhong Zheng (University of Toronto), Igor Gilitschenski (University of Toronto)
Object TrackingGenerationData SynthesisVision Language ModelDiffusion modelVideo
🎯 What it does: The TIRE method is proposed, achieving subject-based 3D/4D generation through a three-stage progressive texture filling of Track-Inpaint-Resplat, significantly enhancing identity preservation and geometric quality.
Track3R: Joint Point Map and Trajectory Prior for Spatiotemporal 3D Understanding
Seong Hyeon Park (Korea Advanced Institute of Science and Technology), Jinwoo Shin (Korea Advanced Institute of Science and Technology)
Pose EstimationAutonomous DrivingTransformerContrastive LearningImageVideo
🎯 What it does: The Track3R framework is proposed, which jointly predicts 3D point mapping and motion trajectories.
Tracking and Understanding Object Transformations
Yihong Sun (Cornell University), Bharath Hariharan (Cornell University)
Object TrackingSegmentationGraph Neural NetworkTransformerLarge Language ModelVideo
🎯 What it does: Achieve continuous tracking of objects in videos during state changes, and detect and describe these state transformations.
TrackingWorld: World-centric Monocular 3D Tracking of Almost All Pixels
Jiahao Lu (Hong Kong University of Science and Technology), Yuan Liu (Hong Kong University of Science and Technology)
Object TrackingDepth EstimationOptimizationSimultaneous Localization and MappingOptical FlowVideo
🎯 What it does: For monocular video, dense 3D tracking in the world coordinate system has been achieved for almost all pixels.
Tractable Multinomial Logit Contextual Bandits with Non-Linear Utilities
Taehyun Hwang (Seoul National University), Min-hwan Oh (Seoul National University)
Recommendation SystemOptimizationReinforcement LearningTabular
🎯 What it does: This paper studies the context bandit problem of polynomial logistic regression (MNL) under nonlinear utility functions (such as neural networks) and proposes a computationally feasible UCB algorithm that is provably convergent to ˜O(√T).
TractoTransformer: Diffusion MRI Streamline Tractography using CNN and Transformer Networks
Itzik Waizman (Ben-Gurion University of the Negev), Tammy Riklin Raviv (Ben-Gurion University of the Negev)
Convolutional Neural NetworkTransformerBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A method for white matter fiber tracking called TractoTransformer, which integrates 3D CNN and Transformer, has been developed to reconstruct white matter fiber bundles from diffusion MRI.
Tradeoffs between Mistakes and ERM Oracle Calls in Online and Transductive Online Learning
Idan Attias (Institute for Data Econometrics Algorithms and Learning), Arvind Ramaswami (Purdue University)
🎯 What it does: This paper studies the trade-off between error (or regret) and the number of oracle calls when learners can only interact with the concept class through ERM or weak consistency oracle in online and transductive online learning.
Train on Pins and Test on Obstacles for Rectilinear Steiner Minimum Tree
Xingbo Du (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
OptimizationReinforcement LearningGraph
🎯 What it does: A reinforcement learning-based OAREST framework is proposed, capable of generating optimal obstacle-avoiding straight-line Steiner trees without trained obstacles.
Train to Defend: First Defense Against Cryptanalytic Neural Network Parameter Extraction Attacks
Ashley Kurian (North Carolina State University), Aydin Aysu (North Carolina State University)
Adversarial AttackConvolutional Neural NetworkImage
🎯 What it does: By incorporating weight similarity regularization during training, the weights of neurons within the same layer become more similar, thereby suppressing parameter extraction attacks based on cryptanalysis while maintaining zero additional overhead during the inference phase.
Train with Perturbation, Infer after Merging: A Two-Stage Framework for Continual Learning
Haomiao Qiu (Harbin Institute of Technology), Liqiang Nie (Harbin Institute of Technology)
ClassificationDomain AdaptationTransformerSupervised Fine-TuningImage
🎯 What it does: A continuous learning framework P&M is proposed, which performs model fusion after each task using convex combinations and task vector perturbations to reduce catastrophic forgetting and enhance generalization.
Trained Mamba Emulates Online Gradient Descent in In-Context Linear Regression
Jiarui Jiang (Harbin Institute of Technology), Liqiang Nie (Harbin Institute of Technology)
OptimizationTransformerTabular
🎯 What it does: This study investigates the learning mechanism of Mamba in the context of linear regression tasks, proving that it can achieve ICL through a variant of online gradient descent.
Training a Scientific Reasoning Model for Chemistry
Siddharth Narayanan, Andrew White
Drug DiscoveryReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
🎯 What it does: A 24B parameter inference-based large language model, ether0, has been proposed and trained, focusing on tasks in the field of chemistry such as molecular design, synthesis, and editing. The model can perform reasoning in natural language and output SMILES molecular structures.
Training Language Models to Generate Quality Code with Program Analysis Feedback
Feng Yao (University of California San Diego), Jingbo Shang (Microsoft Research)
AI Code AssistantReinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes the REAL framework, which utilizes program analysis as a reward signal to train LLMs through reinforcement learning to generate code that is both functionally correct and safe, as well as maintainable.
Training Language Models to Reason Efficiently
Daman Arora (Carnegie Mellon University), Andrea Zanette (Carnegie Mellon University)
Computational EfficiencyTransformerLarge Language ModelReinforcement LearningTextChain-of-Thought
🎯 What it does: Training large inference models to reduce unnecessary computations during inference mainly by shortening the length of chain-of-thought (CoT) to lower inference costs.
Training Robust Graph Neural Networks by Modeling Noise Dependencies
Yeonjun In (KAIST), Chanyoung Park (KAIST)
Graph Neural NetworkGraph
🎯 What it does: In the study of robustness in graph neural networks, a Dependency Noise (DANG) model is proposed, and DA-GNN is constructed to capture the causal relationships of noise through variational inference.
Training the Untrainable: Introducing Inductive Bias via Representational Alignment
Vighnesh Subramaniam (Massachusetts Institute of Technology), Andrei Barbu (Massachusetts Institute of Technology)
ClassificationKnowledge DistillationConvolutional Neural NetworkRecurrent Neural NetworkImageTextSequential
🎯 What it does: A Guidance method is proposed, which enables traditional non-trainable networks (such as fully connected networks, non-residual CNNs, and simple RNNs) to learn effectively by adding hierarchical representation similarity regularization to the target network.