NeurIPS 2025 Papers — Page 35
Conference on Neural Information Processing Systems · 5275 papers
Pessimistic Data Integration for Policy Evaluation
Xiangkun Wu (Zhejiang University), Chengchun Shi (London School of Economics and Political Science)
Reinforcement LearningTabularBiomedical Data
🎯 What it does: This paper studies how to combine historical control data with experimental data to enhance the effectiveness of A/B testing and proposes a lazy data integration method.
Photography Perspective Composition: Towards Aesthetic Perspective Recommendation
Lujian Yao (vivo Mobile Communication Co), Peng-Tao Jiang (vivo Mobile Communication Co)
GenerationRecommendation SystemReinforcement Learning from Human FeedbackReinforcement LearningDiffusion modelImageVideo
🎯 What it does: This paper proposes a Perspective Photography Composition (PPC) method that guides users from suboptimal viewpoints to more aesthetically pleasing ones using perspective transformation videos, and constructs an automated PPC dataset and Perspective Quality Assessment (PQA) model.
PhysCtrl: Generative Physics for Controllable and Physics-Grounded Video Generation
Chen Wang (University of Pennsylvania), Lingjie Liu (University of Pennsylvania)
GenerationData SynthesisDiffusion modelVideoPoint CloudPhysics Related
🎯 What it does: Proposes the PhysCtrl framework to achieve image-to-video generation under the control of physical parameters and external forces;
PhysDiff-VTON: Cross-Domain Physics Modeling and Trajectory Optimization for Virtual Try-On
Shibin Mei (Huawei), Bingbing Ni (Shanghai Jiao Tong University)
GenerationOptimizationConvolutional Neural NetworkDiffusion modelImagePhysics Related
🎯 What it does: A virtual try-on framework called PhysDiff-VTON based on diffusion models is proposed, which integrates pose-guided fabric deformation, waveform-enhanced texture preservation, and energy minimization sampling to achieve realistic clothing try-on effects.
PhysDiff: A Physically-Guided Diffusion Model for Multivariate Time Series Anomaly Detection
Long Li (Northwest Normal University), Hongle Guo (Northwest Normal University)
Anomaly DetectionDiffusion modelTime SeriesFinance RelatedPhysics Related
🎯 What it does: Proposes PhysDiff, a physics-guided diffusion model for multivariate time series anomaly detection.
PhySense: Sensor Placement Optimization for Accurate Physics Sensing
Yuezhou Ma (Tsinghua University), Mingsheng Long (Tsinghua University)
OptimizationComputational EfficiencyTransformerFlow-based ModelTime SeriesPhysics Related
🎯 What it does: A two-stage framework called PhySense has been developed, which utilizes a flow matching reconstructor and projection gradient descent to optimize sensor placement, achieving a synergy between sparse observation reconstruction of physical fields and sensor deployment optimization.
Physics of Language Models: Part 4.1, Architecture Design and the Magic of Canon Layers
Zeyuan Allen-Zhu (Meta)
TransformerLarge Language ModelTextPhysics Related
🎯 What it does: This paper systematically evaluates the core capabilities of different language model architectures by constructing a controllable set of synthetic pre-training tasks (DEPO, BREVO, CAPO, MANO, LANO) and proposes a lightweight Canon layer to enhance the intra-layer horizontal information flow.
Physics-Constrained Flow Matching: Sampling Generative Models with Hard Constraints
Utkarsh Utkarsh, Christopher Vincent Rackauckas (Massachusetts Institute of Technology)
GenerationData SynthesisOptimizationComputational EfficiencyFlow-based ModelTime SeriesPhysics RelatedStochastic Differential Equation
🎯 What it does: The PCFM (Physics-Constrained Flow Matching) framework is proposed, which enforces any nonlinear equality constraints through constrained projection of intermediate states during inference based on a pre-trained flow matching generative model, achieving zero-shot hard constraints.
Physics-Driven Spatiotemporal Modeling for AI-Generated Video Detection
Shuhai Zhang (University of Science and Technology of China), Mingkui Tan (South China University of Technology)
GenerationAnomaly DetectionDiffusion modelScore-based ModelVideoBenchmarkPhysics Related
🎯 What it does: A spatiotemporal modeling method based on the principle of physical conservation is proposed, defining and utilizing the Normalized Spatiotemporal Gradient (NSG) to detect AI-generated videos.
Physics-informed machine learning with domain decomposition and global dynamics for three-dimensional intersecting flows
Leslie K Hwang
OptimizationComputational EfficiencyPhysics Related
🎯 What it does: This study investigates the use of multi-domain PINN for flow field prediction in high aspect ratio three-dimensional cross-flow channels, enhancing model accuracy through global mass conservation constraints.
Physics-informed Neural Operator for Pansharpening
Xinyang Liu (Southeast University), Bo Liu (University of Electronic Science and Technology of China)
RestorationTransformerImagePhysics Related
🎯 What it does: A physics-informed neural operator framework (PINO) is proposed to fuse high-resolution panchromatic images and low-resolution multispectral images for spatial-spectral joint reconstruction.
Physics-informed Reduced Order Modeling of Time-dependent PDEs via Differentiable Solvers
Nima Hosseini Dashtbayaz (University of Western Ontario), Nigel J. W. Morris (Autodesk Research)
TransformerTime SeriesPhysics RelatedOrdinary Differential Equation
🎯 What it does: A physical information-based reduced-order model called Φ-ROM has been developed, which utilizes a differentiable PDE solver to learn the dynamics of latent variables in a low-dimensional latent space and reconstructs the high-dimensional solution field through a decoder, achieving efficient predictions in both time and space.
Physics-informed Value Learner for Offline Goal-Conditioned Reinforcement Learning
Vittorio Giammarino (Purdue University), Ahmed H Qureshi
Reinforcement LearningTabularPhysics Related
🎯 What it does: A physical information regularization based on Eikonal PDE is proposed for value function learning in offline goal-conditioned reinforcement learning, and it is integrated into the HIQL framework.
PhysioWave: A Multi-Scale Wavelet-Transformer for Physiological Signal Representation
Yanlong Chen (ETH Zurich), Yawei Li (ETH Zurich)
ClassificationRepresentation LearningTransformerMultimodalityTime SeriesBiomedical DataElectrocardiogram
🎯 What it does: This paper proposes a PhysioWave framework based on adaptive wavelet decomposition and frequency-guided masking for multi-scale physiological signal representation, and conducts large-scale self-supervised pre-training and downstream multi-modal fusion on three modalities: EMG, ECG, and EEG.
PhysVLM-AVR: Active Visual Reasoning for Multimodal Large Language Models in Physical Environments
Weijie Zhou (Beijing Jiaotong University), Jinqiao Wang (ObjectEye Inc.)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelVision-Language-Action ModelImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: This paper proposes the Active Visual Reasoning (AVR) paradigm and constructs the corresponding CLEVR-AVR interactive benchmark and AVR-152k dataset. Based on this, the PhysVLM-AVR model is trained and released, achieving multi-step perception-reasoning-action closed-loop reasoning in partially observable environments.
PhySwin: An Efficient and Physically-Informed Foundation Model for Multispectral Earth Observation
Chong Tang (University of Southampton), Jagmohan Chauhan (University College London)
RecognitionObject DetectionSegmentationComputational EfficiencyTransformerImage
🎯 What it does: This paper proposes PhySwin, a physically informed and efficient foundational model for multi-spectral Earth observation.
PhysX-3D: Physical-Grounded 3D Asset Generation
Ziang Cao (Nanyang Technological University), Ziwei Liu (Nanyang Technological University)
GenerationData SynthesisVision Language ModelDiffusion modelImageMeshPhysics Related
🎯 What it does: A PhysXNet physical attribute annotation dataset and PhysXGen generation framework are proposed, enabling the generation of 3D assets from images with absolute scale, materials, operability, kinematics, and functional descriptions.
PID-controlled Langevin Dynamics for Faster Sampling on Generative Models
Hongyi Chen (Tsinghua University), Xiao-Ping Zhang (Tsinghua University)
GenerationComputational EfficiencyImageGraphStochastic Differential Equation
🎯 What it does: A PID control-based Langevin dynamics algorithm (PIDLD) was designed and validated to accelerate the sampling process of unsupervised generative models.
PiKE: Adaptive Data Mixing for Large-Scale Multi-Task Learning Under Low Gradient Conflicts
Zeman Li (University of Southern California), Vahab Mirrokni (Google Research)
OptimizationData-Centric LearningTransformerLarge Language ModelText
🎯 What it does: This paper proposes PiKE, an adaptive data mixing algorithm that dynamically adjusts sampling weights in multi-task learning (especially in large-scale language model pre-training) to enhance convergence speed and downstream performance.
Pin the Tail on the Model: Blindfolded Repair of User-Flagged Failures in Text-to-Image Services
Gefei Tan (Northwestern University), Xiao Wang (Brave Software)
GenerationComputational EfficiencyDiffusion modelImage
🎯 What it does: A secure model repair framework called SURE is proposed, allowing service providers and third-party repair institutions to collaboratively fix errors and biases in text-to-image diffusion models without disclosing user feedback, model weights, or repair knowledge.
PINN Balls: Scaling Second-Order Methods for PINNs with Domain Decomposition and Adaptive Sampling
Andrea Bonfanti (Basque Center for Applied Mathematics), Marco Ellero (Basque Center for Applied Mathematics)
OptimizationMixture of ExpertsPhysics Related
🎯 What it does: A scalable second-order mixed expert PINN model (PINN BALLS) is proposed, achieving high-precision PDE solutions through learnable domain decomposition and adaptive sampling.
PINNs with Learnable Quadrature
Sourav Pal (University of Wisconsin Madison), Vikas Singh (University of Wisconsin Madison)
TabularPhysics Related
🎯 What it does: A learnable efficient numerical integration rule (LearnQuad) is proposed in PINNs (Physics-Informed Neural Networks), achieving adaptive sampling and weak/strong form integration for low-dimensional PDE solving by parameterizing the adjustable weight function with neural network and utilizing the asymptotic expansion of orthogonal polynomials.
Pinpointing Attention-Causal Communication in Language Models
Gabriel Franco (Boston University), Mark Crovella (Boston University)
TransformerLarge Language ModelText
🎯 What it does: A new method is proposed for identifying low-dimensional signals related to causal attention mechanisms (attention-causal communication) in Transformer models, and for constructing communication graphs and feedback paths within the model using these signals.
PIPE: Physics-Informed Position Encoding for Alignment of Satellite Images and Time Series in Typhoon Forecasting
Haobo Li (Hong Kong University of Science and Technology), Alexis Kai Hon Lau (Hong Kong University of Science and Technology)
TransformerSupervised Fine-TuningVision Language ModelImageMultimodalityTime SeriesPhysics Related
🎯 What it does: A multimodal weather forecasting task based on satellite images and numerical time series is proposed, along with a lightweight Physical Information Position Encoding (PIPE) method.
PipeFusion: Patch-level Pipeline Parallelism for Diffusion Transformers Inference
Jiarui Fang (Tencent), WANG Jiannan
GenerationTransformerDiffusion modelImage
🎯 What it does: This paper proposes PipeFusion, a patch-level pipeline parallel inference scheme for Diffusion Transformers, which significantly reduces inference latency and memory usage.
PIVNO: Particle Image Velocimetry Neural Operator
Jie Xu (Beijing Union University), Qinghua Cui (Peking University)
Image TranslationDomain AdaptationConvolutional Neural NetworkRecurrent Neural NetworkOptical FlowImage
🎯 What it does: A particle image velocimetry (PIV) framework based on neural operators (PIVNO) is proposed, which directly maps particle image pairs to flow fields, avoiding the traditional issues of cost volume construction and local matching.
Pixel Reasoner: Incentivizing Pixel Space Reasoning via Curiosity-Driven Reinforcement Learning
Alex Su (University of Waterloo), Wenhu Chen (University of Waterloo)
Reinforcement LearningVision Language ModelImage
🎯 What it does: By first performing warm-start instruction tuning on the model and then using curiosity-driven reinforcement learning, the VLM can actively execute visual operations (such as Zoom-in and Select-Frame) in pixel space for reasoning, significantly enhancing visual reasoning performance.
Pixel-Perfect Depth with Semantics-Prompted Diffusion Transformers
Gangwei Xu (Zhejiang University), Xin Yang (Zhejiang University)
GenerationDepth EstimationTransformerDiffusion modelImagePoint Cloud
🎯 What it does: Proposes Pixel-Perfect Depth, a monocular depth estimation model that directly generates diffusion in pixel space, capable of outputting point clouds without flying pixels.
PixPerfect: Seamless Latent Diffusion Local Editing with Discriminative Pixel-Space Refinement
Haitian Zheng (Adobe Research), Zhe Lin (Adobe Research)
Image HarmonizationRestorationGenerationDiffusion modelGenerative Adversarial NetworkImage
🎯 What it does: This paper presents PixPerfect, a general pixel-level refinement framework for correcting color, texture, and boundary defects in LDM local editing and inpainting.
Place Cells as Multi-Scale Position Embeddings: Random Walk Transition Kernels for Path Planning
Minglu Zhao (University of California, Los Angeles), Ying Nian Wu (University of California, Los Angeles)
OptimizationRobotic IntelligenceReinforcement LearningGraph
🎯 What it does: This paper proposes modeling the hippocampal cell population as a non-negative positional embedding, achieved through spectral decomposition of the multi-step random walk transition kernel, and utilizes the inner product of the embedding to approximate transition probabilities, representing a multi-scale cognitive map of the environment. Subsequently, a path planning algorithm based on gradient ascent and adaptive scale selection is designed and validated in simulated open fields and multi-obstacle mazes.
PLANA3R: Zero-shot Metric Planar 3D Reconstruction via Feed-forward Planar Splatting
Changkun Liu (Hong Kong University of Science and Technology), Tristan Braud (Hong Kong University of Science and Technology)
SegmentationPose EstimationDepth EstimationTransformerVision Language ModelImage
🎯 What it does: A zero-shot planar 3D reconstruction model PLANA3R is proposed, which can predict sparse 3D planar primitives and simultaneously estimate relative camera poses from two-view images without pose information, achieving metric-scale 3D reconstruction.
PlanarGS: High-Fidelity Indoor 3D Gaussian Splatting Guided by Vision-Language Planar Priors
Xirui Jin (Shanghai Jiao Tong University), Wenxian Yu (Shanghai Jiao Tong University)
RestorationSegmentationVision Language ModelGaussian SplattingPoint Cloud
🎯 What it does: This paper proposes the PlanarGS framework, which utilizes a visual language pre-trained model to extract planar priors and combines them with multi-view depth/normal priors to impose planar consistency and geometric supervision on 3D Gaussian Splatting, enhancing the high-fidelity reconstruction of large planar areas in indoor environments.
Planning and Learning in Average Risk-aware MDPs
Weikai Wang (GERAD and HEC Montreal), Erick Delage (GERAD and HEC Montreal)
OptimizationReinforcement LearningTabular
🎯 What it does: This paper proposes a risk-aware relative value iteration (RVI) algorithm for average cost Markov decision processes (MDPs) and designs two model-independent Q-learning algorithms: a multi-layer Monte Carlo (MLMC) Q-learning that supports general dynamic risk measures, and a UBSR Q-learning specifically for expected loss risk that can be learned offline.
Planning with Quantized Opponent Models
XiaoPeng Yu, Zongqing Lu (Peking University)
OptimizationReinforcement LearningAuto EncoderTabularBenchmark
🎯 What it does: This paper proposes Quantized Opponent Models (QOM), which implement a discrete opponent type library based on Bayesian inference in partially observable multi-agent games and combine it with MCTS for Bayesian-aware planning.
Planning without Search: Refining Frontier LLMs with Offline Goal-Conditioned RL
Joey Hong (University of California Berkeley), Sergey Levine (University of California Berkeley)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: For tasks that require multi-round interaction, a method is proposed to assist large language models in self-improving their reasoning through an offline reinforcement learning-trained goal-conditioned value function (natural language critic), avoiding direct RL fine-tuning of LLMs.
PlanU: Large Language Model Reasoning through Planning under Uncertainty
Ziwei Deng (Xiamen University), Cheng Wang (National University of Defense Technology)
Large Language ModelReinforcement LearningWorld ModelTextBenchmark
🎯 What it does: A planning method named PlanU based on LLM‑based MCTS is proposed, capable of making multi-step decisions in the presence of environmental uncertainty and LLM generation uncertainty.
Plasticity as the Mirror of Empowerment
David Abel, Satinder Singh
🎯 What it does: This paper conducts a systematic study of the plasticity of agents and empowerment from the perspective of information theory, proposing a general measure of Generalized Directed Information (GDI) and using it to define plasticity, as well as to elucidate the relationship and inherent tension between plasticity and empowerment.
PlayerOne: Egocentric World Simulator
Yuanpeng Tu (Hong Kong University), Hengshuang Zhao (Hong Kong University)
GenerationData SynthesisPose EstimationTransformerDiffusion modelAuto EncoderImageVideo
🎯 What it does: The first real-world egocentric world simulator, PlayerOne, has been constructed, which can generate immersive videos that are synchronized with motion and consistent in perspective based on first-person images provided by users and human motion sequences captured by external cameras.
PLD: A Choice-Theoretic List-Wise Knowledge Distillation
Ejafa Bassam (Peking University), Kaigui Bian (Peking University)
Object DetectionKnowledge DistillationImage
🎯 What it does: A list-based knowledge distillation method (PLD) based on the Plackett-Luce model is proposed, using the complete category ranking of the teacher as the distillation target, unifying cross-entropy and distillation loss.
PLEIADES: Building Temporal Kernels with Orthogonal Polynomials
Yan Ru Pei (NVIDIA Corporation), Olivier JMD Coenen
ClassificationObject DetectionConvolutional Neural NetworkVideo
🎯 What it does: This paper proposes a neural network called PLEIADES, which uses orthogonal polynomial expansion for adaptive temporal convolution kernels, suitable for low-latency spatiotemporal classification and detection with event cameras.
Plenodium: Underwater 3D Scene Reconstruction with Plenoptic Medium Representation
Changguang WU, Jinhui Tang (Nanjing Forestry University)
RestorationGenerationGaussian SplattingImage
🎯 What it does: Proposes the Plenodium framework, specifically designed for underwater 3D scene reconstruction and image de-watering.
PLMTrajRec: A Scalable and Generalizable Trajectory Recovery Method with Pre-trained Language Models
Tonglong Wei (Beijing Jiaotong University), Huaiyu Wan (Beijing Jiaotong University)
RestorationDomain AdaptationTransformerLarge Language ModelPrompt EngineeringTime Series
🎯 What it does: A scalable general trajectory recovery method based on pre-trained language models (PLMTrajRec) has been developed, capable of recovering missing points in sparse trajectories and achieving cross-domain generalization across different sampling intervals.
pLSTM: parallelizable Linear Source Transition Mark networks
Korbinian Pöppel (Johannes Kepler University Linz), Sepp Hochreiter (Johannes Kepler University Linz)
Super ResolutionRecurrent Neural NetworkGraph Neural NetworkImageGraph
🎯 What it does: A parallelizable linear source-transition-labeling network (pLSTM) is proposed for long-range information propagation in multidimensional data (images, graphs).
Plug-and-Play Context Feature Reuse for Efficient Masked Generation
Xuejie Liu (Peking University), Yitao Liang (Peking University)
GenerationComputational EfficiencyTransformerImage
🎯 What it does: A pluggable ReCAP module is proposed, which alternates between full evaluation and lightweight evaluation during the inference process of Masked Generative Models (MGMs) to reduce the computational load at each step and improve generation efficiency.
Plug-and-play Feature Causality Decomposition for Multimodal Representation Learning
Ye Liu (South China University of Technology), Hongmin Cai (South China University of Technology)
Representation LearningContrastive LearningMultimodality
🎯 What it does: A pluggable multimodal feature causal decomposition (FCD) module has been designed and implemented to decompose unimodal features into three parts: modality-invariant, unique (complementary), and redundant (noise) without altering the original model structure, and to combine it with a fusion module to generate more robust multimodal representations.
PMLF: A Physics-Guided Multiscale Loss Framework for Structurally Heterogeneous Time Series
Xinghong Chen (Fujian Normal University), Guannan Chen (Fujian Normal University)
TransformerTime SeriesPhysics Related
🎯 What it does: A physics-based multi-scale loss framework is proposed for seasonal and trend decomposition of time series, supervised by quadratic loss and logarithmic loss respectively.
PMQ-VE: Progressive Multi-Frame Quantization for Video Enhancement
ZhanFeng Feng (University of Science and Technology of China), Zheng-Jun Zha (University of Science and Technology of China)
RestorationSuper ResolutionKnowledge DistillationTransformerVideo
🎯 What it does: This paper proposes a Progressive Multi-Frame Quantization framework (PMQ-VE) for multi-frame video enhancement tasks, which consists of a rough stage called Backtracking-based Multi-Frame Quantization (BMFQ) and a refinement stage called Progressive Multi-Teacher Distillation (PMTD), achieving low-bit-width quantization for tasks such as STVSR, VSR, and VFI.
PocketSR: The Super-Resolution Expert in Your Pocket Mobiles
Haoze Sun (Tsinghua University), Wenbo Li (Joy Future Academy)
RestorationSuper ResolutionDiffusion modelAuto EncoderImage
🎯 What it does: Proposes PocketSR, a lightweight single-step generative super-resolution model;
POCO: Scalable Neural Forecasting through Population Conditioning
Yu Duan (Massachusetts Institute of Technology), Kanaka Rajan (Harvard Medical School)
Time SeriesBiomedical Data
🎯 What it does: A POCO model that combines a single neuron predictor and a global population encoder is proposed for cell-level time series prediction on multi-animal multi-session calcium imaging data.
PoE-World: Compositional World Modeling with Products of Programmatic Experts
Wasu Top Piriyakulkij (Cornell University), Kevin Ellis (Cornell University)
Robotic IntelligenceLarge Language ModelReinforcement LearningMixture of ExpertsWorld ModelVideo
🎯 What it does: Utilizing large language models to generate small program experts and representing the world model as an exponentially weighted product of these experts, learning interpretable and scalable symbolic world models, which are then used for planning and pre-training strategies in Atari games.
PoGDiff: Product-of-Gaussians Diffusion Models for Imbalanced Text-to-Image Generation
Ziyan Wang (Georgia Institute of Technology), Hao Wang (Rutgers University)
GenerationData SynthesisDiffusion modelImageText
🎯 What it does: A fine-tuning method named PoGDiff is proposed to address the poor generation performance of minority classes in text-to-image generation tasks caused by imbalanced data distribution.
Point Cloud Synthesis Using Inner Product Transforms
Ernst Röell (AIDOS Lab, University of Fribourg), Bastian Rieck (Institute of AI for Health, Helmholtz Munich)
GenerationData SynthesisComputational EfficiencyConvolutional Neural NetworkAuto EncoderPoint Cloud
🎯 What it does: A two-step point cloud generation framework based on Inner Product Transformation (IPT) is proposed, which first encodes the point cloud into a two-dimensional image and then reconstructs the point cloud through a convolutional decoder, achieving reconstruction, generation, and downsampling.
Point or Line? Using Line-based Representation for Panoptic Symbol Spotting in CAD Drawings
Xingguang Wei (University of Science and Technology of China), Hongjie Zhang (Shanghai AI Laboratory)
Object DetectionSegmentationTransformerImage
🎯 What it does: For the task of panoramic symbol localization in CAD drawings, VecFormer is proposed to construct type-independent line segment representations through line segment sampling, completing symbol detection and segmentation with joint instance and semantic branches.
Point-MaDi: Masked Autoencoding with Diffusion for Point Cloud Pre-training
Xiaoyang Xiao (Xi'an Jiaotong University), Shaoyi Du (Xi'an Jiaotong University)
ClassificationObject DetectionSegmentationTransformerDiffusion modelAuto EncoderPoint Cloud
🎯 What it does: A dual diffusion mask autoencoder called Point-MaDi is proposed for point cloud pre-training, which eliminates position encoding leakage and learns global semantics and local geometry.
Point-RFT: Improving Multimodal Reasoning with Visually Grounded Reinforcement Finetuning
Minheng Ni, Lijuan Wang
TransformerSupervised Fine-TuningReinforcement LearningVision Language ModelMultimodalityChain-of-Thought
🎯 What it does: Proposes the Point-RFT framework, achieving enhanced fine-tuning of visual document understanding through visually guided Chain-of-Thought (CoT).
Point3R: Streaming 3D Reconstruction with Explicit Spatial Pointer Memory
Yuqi Wu (Tsinghua University), Jiwen Lu (Tsinghua University)
Pose EstimationDepth EstimationAutonomous DrivingTransformerImageVideoPoint Cloud
🎯 What it does: An online streaming 3D reconstruction framework called Point3R is proposed, which utilizes explicit spatial pointer memory to achieve global point cloud generation from continuous images.
Point4Bit: Post Training 4-bit Quantization for Point Cloud 3D Detection
Jianyu Wang (Tongji University), Sifan Zhou (Carnegie Mellon University)
Object DetectionCompressionAutonomous DrivingPoint Cloud
🎯 What it does: A post-training 4-bit quantization framework called Point4bit is proposed for voxelized 3D object detection, achieving nearly lossless accuracy compression without the need for retraining or labels.
PointMAC: Meta-Learned Adaptation for Robust Test-Time Point Cloud Completion
Linlian Jiang (Concordia University), Yang Wang (Concordia University)
GenerationDomain AdaptationMeta LearningTransformerPoint Cloud
🎯 What it does: Proposes the PointMAC framework, which utilizes meta-learning driven self-supervised adaptation during testing to achieve sample-specific dynamic inference for point cloud completion tasks.
PointMapPolicy: Structured Point Cloud Processing for Multi-Modal Imitation Learning
Xiaogang Jia (Karlsruhe Institute of Technology), Gerhard Neumann (Karlsruhe Institute of Technology)
Robotic IntelligenceReinforcement Learning from Human FeedbackConvolutional Neural NetworkRecurrent Neural NetworkReinforcement LearningDiffusion modelMultimodalityPoint Cloud
🎯 What it does: This paper proposes PointMapPolicy, which projects depth maps into 2D point mappings and fuses them with RGB to achieve multi-modal imitation learning using an EDM-based diffusion strategy.
PointTruss: K-Truss for Point Cloud Registration
Yue Wu (Xidian University), wenping ma
Pose EstimationAutonomous DrivingGraph Neural NetworkPoint Cloud
🎯 What it does: A point cloud registration method based on graph theory called PointTruss is proposed, which uses triangle-supported k-truss decomposition to filter inliers and estimate rigid transformations.
Poison as Cure: Visual Noise for Mitigating Object Hallucinations in LVMs
Kejia Zhang (Xiamen University), Huan Wang (Westlake University)
Object DetectionOptimizationAdversarial AttackTransformerVision Language ModelImageMultimodality
🎯 What it does: A training-free method based on Visual Adversarial Perturbations (VAP) is proposed, which injects subtle noise into visual inputs through adversarial optimization to reduce object hallucinations in large visual-language models (LVMs);
Polar Sparsity: High Throughput Batched LLM Inferencing with Scalable Contextual Sparsity
Susav Shrestha (Texas A&M University), A. L. Narasimha Reddy (Texas A&M University)
TransformerLarge Language ModelText
🎯 What it does: For large-scale batch inference of LLMs, Polar Sparsity is proposed, achieving contextual sparsity by dynamically activating sub-networks and heads in the MLP and attention layers.
PoLAR: Polar-Decomposed Low-Rank Adapter Representation
Kai Lion (ETH Zurich), Niao He (ETH Zurich)
OptimizationTransformerLarge Language ModelText
🎯 What it does: Proposes PoLAR parameterization and landing optimization to address the issue of insufficient directional diversity caused by the low stable rank of LoRA, achieving faster convergence.
PolarQuant: Leveraging Polar Transformation for Key Cache Quantization and Decoding Acceleration
Songhao Wu (Renmin University of China), Rui Yan (Renmin University of China)
CompressionComputational EfficiencyTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: The PolarQuant method is proposed to perform polar coordinate quantization on the key vectors in the KV cache of large language models, thereby reducing memory usage and accelerating decoding.
Policy Compatible Skill Incremental Learning via Lazy Learning Interface
Daehee Lee (Sungkyunkwan University), Honguk Woo (Sungkyunkwan University)
Robotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningMultimodality
🎯 What it does: A framework named SIL-C is proposed to maintain compatibility between policies and skills during the Skill Incremental Learning (SIL) process, allowing improvements in learned lower-level skills to seamlessly enhance the performance of higher-level policies without the need for retraining or structural adjustments.
Policy Gradient Methods Converge Globally in Imperfect-Information Extensive-Form Games
Fivos Kalogiannis (University of California San Diego), Gabriele Farina (Massachusetts Institute of Technology)
OptimizationReinforcement Learning
🎯 What it does: It is proven that in zero-sum generalized form games with imperfect information, three types of policy gradient methods (direct parameterization, softmax parameterization, and natural policy gradient) converge to an ε-Nash equilibrium in the final iteration, and polynomial time iteration and sample complexity are provided.
Policy Optimized Text-to-Image Pipeline Design
Uri Gadot (Technion), Shie Mannor (Technion)
GenerationOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This paper proposes a reinforcement learning-based text-to-image pipeline generation framework called FlowRL, which achieves the automatic generation of high-quality and diverse ComfyUI workflows through supervised fine-tuning of language models, reward model prediction, and GRPO optimization.
PolyJuice Makes It Real: Black-Box, Universal Red Teaming for Synthetic Image Detectors
Sepehr Dehdashtian (Michigan State University), Vishnu Boddeti
GenerationData SynthesisAdversarial AttackSupervised Fine-TuningGenerative Adversarial NetworkImageText
🎯 What it does: PolyJuice has been developed, a black-box general red team method for synthetic image detectors, designed to generate synthetic images that can mislead detectors.
Polyline Path Masked Attention for Vision Transformer
Zhongchen Zhao (Xi'an Jiaotong University), Deyu Meng (Xi'an Jiaotong University)
ClassificationObject DetectionSegmentationTransformerImage
🎯 What it does: Proposes Polyline Path Masked Attention (PPMA), which extends the structured mask of Mamba2 to two-dimensional images, embedding it in ViT to achieve explicit modeling of spatial adjacency.
PolyPose: Deformable 2D/3D Registration via Polyrigid Transformations
Vivek Gopalakrishnan (Massachusetts Institute of Technology), Polina Golland (Massachusetts Institute of Technology)
Pose EstimationOptimizationImageBiomedical DataComputed Tomography
🎯 What it does: A two-dimensional/three-dimensional deformable registration method called PolyPose is proposed, based on a polyrigid framework, to estimate the pose and deformation of a 3D CT volume from a minimal number of 2D X-ray images.
PolyVivid: Vivid Multi-Subject Video Generation with Cross-Modal Interaction and Enhancement
Teng Hu (Shanghai Jiao Tong University), Ran Yi (Shanghai Jiao Tong University)
GenerationData SynthesisTransformerLarge Language ModelVision Language ModelVideoTextMultimodality
🎯 What it does: PolyVivid, through a multi-subject video customization framework, can generate high-quality videos with consistent identities using multiple subject images and text prompts as input;
Pool Me Wisely: On the Effect of Pooling in Transformer-Based Models
Sofiane ENNADIR, Lele Cao (Microsoft Gaming)
ClassificationRestorationSegmentationData SynthesisTransformerImageTextTime Series
🎯 What it does: This paper systematically studies the role of pooling layers in the Transformer model, providing a theoretical framework and closed-form upper bounds for pooling expressiveness, and empirically validating it in visual, natural language, and time series tasks.
Pose Splatter: A 3D Gaussian Splatting Model for Quantifying Animal Pose and Appearance
Jack Goffinet (Duke University), David Carlson
Pose EstimationGaussian SplattingVideo
🎯 What it does: Proposes the Pose Splatter framework, which utilizes shape sculpting and 3D Gaussian splatting to achieve unlabelled, non-frame-by-frame optimization of 3D pose and appearance reconstruction for animals.
PoseCrafter: Extreme Pose Estimation with Hybrid Video Synthesis
Qing Mao (Northwestern Polytechnical University), Gim Hee Lee (National University of Singapore)
Data SynthesisPose EstimationDiffusion modelVideo
🎯 What it does: To address the problem of extreme perspective changes (where images have little or no overlap), the PoseCrafter framework is proposed, which assists camera pose estimation by synthesizing intermediate views.
Positional Fragility in LLMs: How Offset Effects Reshape Our Understanding of Memorization Risks
Yixuan Xu (ETH Zürich), Imanol Schlag (ETH Zürich)
TransformerLarge Language ModelText
🎯 What it does: The study investigated the verbatim memory behavior of large language models under different context shifts, finding that short prefixes are most likely to trigger memory at the start of the window, and that shifts lead to memory decay.
Post Hoc Regression Refinement via Pairwise Rankings
Kevin Tirta Wijaya (Max Planck Institute for Informatics), Vahid Babaei (Max Planck Institute for Informatics)
OptimizationDrug DiscoveryLarge Language ModelTabularAgriculture Related
🎯 What it does: A post-hoc correction framework called RankRefine is proposed, which enhances the accuracy of continuous attribute predictions in low-data environments by combining the predictions of a benchmark regressor with pairwise ranking information generated by an external ranker (such as LLMs or experts).
Posterior Contraction for Sparse Neural Networks in Besov Spaces with Intrinsic Dimensionality
Kyeongwon Lee (University of Maryland), Seonghyun Jeong (Yonsei University)
🎯 What it does: It is proven that sparse Bayesian neural networks can achieve optimal posterior contraction rates in anisotropic Besov spaces and their hierarchical combination structures, with the contraction rate depending solely on the intrinsic dimension of the function, overcoming the curse of dimensionality.
Posterior Sampling by Combining Diffusion Models with Annealed Langevin Dynamics
Zhiyang Xun (University of Texas at Austin), Eric Price (Microsoft Research)
RestorationSuper ResolutionDiffusion modelImageStochastic Differential Equation
🎯 What it does: A posterior sampling algorithm that combines diffusion models with annealed Langevin dynamics is proposed, suitable for local or global log-convex prior distributions.
Power Lines: Scaling laws for weight decay and batch size in LLM pre-training
Shane Bergsma (Cerebras Systems), Joel Hestness (Cerebras Systems)
TransformerLarge Language ModelText
🎯 What it does: This study investigates and verifies the scaling laws of weight decay λ and batch size B in LLM pre-training with respect to model size N and data volume D. It proposes a weight decay prediction formula based on the AdamW time scale τ and provides optimal and critical scaling rules for batch size.
PPMStereo: Pick-and-Play Memory Construction for Consistent Dynamic Stereo Matching
Yun Wang (City University of Hong Kong), Junjie Hu (Chinese University of Hong Kong Shenzhen)
Depth EstimationOptimizationRecurrent Neural NetworkOptical FlowVideo
🎯 What it does: A dynamic stereo matching framework called PPMStereo based on Pick-and-Play Memory is proposed to achieve spatiotemporal consistency in disparity estimation for long-sequence videos.
Practical and Effective Code Watermarking for Large Language Models
Zhimeng Guo (Pennsylvania State University), Minhao Cheng (Pennsylvania State University)
GenerationAI Code AssistantTransformerLarge Language ModelText
🎯 What it does: A practical watermarking framework ACW suitable for large language model code generation is proposed, which can embed and detect watermarks without exposing the LLM or prompt information.
Practical Bayes-Optimal Membership Inference Attacks
Marcus Lassila (Chalmers University of Technology), Alexandre Graell i Amat (Chalmers University of Technology)
Safty and PrivacyAdversarial AttackGraph Neural NetworkGraph
🎯 What it does: This paper studies and implements Bayesian optimal membership inference attacks (MIA) for i.i.d. and graph-structured data, particularly focusing on node-level MIA for Graph Neural Networks (GNN).
Practical do-Shapley Explanations with Estimand-Agnostic Causal Inference
Álvaro Parafita (Barcelona Supercomputing Center), Francisco J. Cazorla (Barcelona Supercomputing Center)
Explainability and InterpretabilityComputational EfficiencyGraph Neural NetworkFlow-based ModelGraph
🎯 What it does: This paper proposes the use of estimand-agnostic methods (EA) and frontier reducible algorithms (FRA) to efficiently compute do-SHAP values on complex causal graphs, and validates their interpretability on real data.
Practical Kernel Selection for Kernel-based Conditional Independence Test
Wenjie Wang (University of Melbourne), Feng Liu (University of Melbourne)
OptimizationHyperparameter SearchTabular
🎯 What it does: This paper studies and implements a kernel parameter selection method based on power ratio (Power) to improve the kernel parameter setting in the Kernel-based Conditional Independence Test (KCI), thereby enhancing the effectiveness and robustness of CI testing.
Pragmatic Heterogeneous Collaborative Perception via Generative Communication Mechanism
Junfei Zhou (Southwest Jiaotong University), Jianping Wang (City University of Hong Kong)
Object DetectionAutonomous DrivingConvolutional Neural NetworkDiffusion modelMultimodality
🎯 What it does: A generative communication mechanism called GenComm is proposed for heterogeneous multi-agent collaborative perception.
Praxis-VLM: Vision-Grounded Decision Making via Text-Driven Reinforcement Learning
Zhe Hu (Hong Kong Polytechnic University), Yu Yin (Case Western Reserve University)
Autonomous DrivingRobotic IntelligenceTransformerReinforcement LearningVision Language ModelVideoText
🎯 What it does: This paper proposes a text-driven reinforcement learning method called Praxis-VLM, which enables visual language models to possess transferable reasoning and decision-making capabilities in visual decision-making tasks.
Pre-trained Large Language Models Learn to Predict Hidden Markov Models In-context
Yijia Dai (Cornell University), Jennifer J. Sun (Cornell University)
TransformerLarge Language ModelTextSequential
🎯 What it does: The study investigates the prediction of sequences generated by hidden Markov models (HMM) using pre-trained large language models (LLM) through in-context learning (ICL), finding that their performance can approach the theoretically optimal Viterbi algorithm.
Pre-Trained Policy Discriminators are General Reward Models
Shihan Dou (Shanghai AI Laboratory), Kai Chen (Shanghai AI Laboratory)
Reinforcement Learning from Human FeedbackTransformerReinforcement LearningContrastive LearningText
🎯 What it does: Proposes the POLAR framework, redefining the reward model as a policy discriminator, utilizing unsupervised pre-training and a small amount of manually labeled fine-tuning to achieve relative evaluation of different policies;
PREAMBLE: Private and Efficient Aggregation via Block Sparse Vectors
Hilal Asi, Kunal Talwar
Safty and PrivacyComputational EfficiencyImage
🎯 What it does: This paper proposes a private high-dimensional vector aggregation protocol (PREAMBLE) under the Prio two-server model. By compressing vectors into block-sparse form and combining distributed point functions (DPF) with sampling techniques, it achieves low communication, low computation, and supports differential privacy in aggregation.
Precise Asymptotics and Refined Regret of Variance-Aware UCB
Yingying Fan (University of Southern California), Zhengyuan Zhou (New York University)
OptimizationReinforcement Learning from Human Feedback
🎯 What it does: This paper provides a precise asymptotic and non-asymptotic analysis of the arm pulling rate of the UCB-Var algorithm in multi-armed bandits.
Precise Diffusion Inversion: Towards Novel Samples and Few-Step Models
Jing Zuo (Beijing University of Posts and Telecommunications), Yonggang Qi (Beijing University of Posts and Telecommunications)
RestorationGenerationDiffusion modelImage
🎯 What it does: A testing optimization framework named PreciseInv is proposed, which quickly and accurately maps real images back to the latent space of diffusion models, enabling high-quality image reconstruction and editing.
Precise Information Control in Long-Form Text Generation
Jacqueline He (Paul G. Allen School of Computing Science and Engineering University of Washington), Luke Zettlemoyer (Paul G. Allen School of Computing Science and Engineering University of Washington)
GenerationOptimizationTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed the Precise Information Control (PIC) framework, constructed the PIC-Bench benchmark, and based on this, trained the PIC-LM 8B model through weakly supervised preference data construction and length normalization direct preference optimization;
Preconditioned Langevin Dynamics with Score-based Generative Models for Infinite-Dimensional Linear Bayesian Inverse Problems
Lorenzo Baldassari (University of Basel), Maarten V. de Hoop (Rice University)
Score-based ModelStochastic Differential Equation
🎯 What it does: In infinite-dimensional function spaces, a rigorous analysis of the Langevin sampling method using the score model (SGM) as a prior is conducted, providing error estimates, convergence proofs, and the form of the optimal preconditioner.
Predictability Enables Parallelization of Nonlinear State Space Models
Xavier Gonzalez (Stanford University), Scott Linderman (Stanford University)
OptimizationRecurrent Neural NetworkTime SeriesStochastic Differential Equation
🎯 What it does: This paper reveals that the predictability of the model (measured by the maximum Lyapunov exponent) determines the condition number of the optimization problem by transforming the trajectory solution of the nonlinear state space model into a fast-converging parallel optimizable problem, thus judging whether efficient parallelization can be achieved on a GPU.
Predictable Scale (Part II) --- Farseer: A Refined Scaling Law in LLMs
Houyi Li (Fudan University), Daxin Jiang (StepFun)
TransformerLarge Language ModelText
🎯 What it does: This paper proposes and validates a new LLM scaling law—Farseer—through large-scale experiments, which can accurately predict the loss of models of different scales;
Predicting Empirical AI Research Outcomes with Language Models
Jiaxin Wen (University of California Berkeley), Shi Feng (George Washington University)
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: A benchmark was constructed to predict which of two research approaches performs better in empirical evaluations, and a system was implemented based on a fine-tuned GPT-4.1 model and a retrieval agent, capable of predicting which approach has a higher probability of success without running experiments.
Predicting Functional Brain Connectivity with Context-Aware Deep Neural Networks
Alexander Ratzan (New York University), Erdem Varol (University of Pennsylvania)
TransformerBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Using deep learning methods, predict the whole-brain functional connectivity matrix based on human Allen gene expression and spatial coordinates.
Predicting partially observable dynamical systems via diffusion models with a multiscale inference scheme
Rudy Morel (Flatiron Institute), Shirley Ho (New York University)
TransformerDiffusion modelVideoMultimodalityPhysics Related
🎯 What it does: A multi-scale inference scheme is proposed for predicting partially observable, long-memory dynamical systems, such as the evolution of solar active regions.
Predicting the Performance of Black-box Language Models with Follow-up Queries
Dylan Sam (Carnegie Mellon University), J Zico Kolter
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: A completely black-box access framework called QueRE is proposed, which extracts the probability of the answer 'yes' as a low-dimensional feature by initiating a set of follow-up questions to the LLM, and uses a linear predictor to determine whether the model's answer is correct, detect if it is influenced by adversarial prompts, and distinguish between different LLMs.
Prediction with expert advice under additive noise
Alankrita Bhatt (Granica Computing), Victoria Kostina (California Institute of Technology)
🎯 What it does: This paper studies the optimal learning performance under the framework of expert opinion prediction when feedback is disturbed by additive noise, providing the limit lower and upper bounds under the influence of noise, and unifying the analysis of Gaussian, uniform, and symmetric log-concave noise.