ICLR 2025 Papers — Page 26
International Conference on Learning Representations · 3704 papers
PhiNets: Brain-inspired Non-contrastive Learning Based on Temporal Prediction Hypothesis
Satoki Ishikawa (Okinawa Institute of Science and Technology), Yuki Takezawa (Kyoto University)
Representation LearningContrastive LearningImage
🎯 What it does: This paper proposes PhiNet, a brain-inspired model based on non-contrastive self-supervised learning, which draws on the hippocampal temporal prediction hypothesis and uses StopGradient to simulate synaptic delay.
PhyloLM: Inferring the Phylogeny of Large Language Models and Predicting their Performances in Benchmarks
Nicolas Yax (INSERM), Stefano Palminteri (INSERM)
Large Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: The PhyloLM method is proposed, which uses evolutionary tree algorithms to infer the phylogenetic relationships between LLMs based on the token sequences output by LLMs and predict their benchmark performance.
PhyloVAE: Unsupervised Learning of Phylogenetic Trees via Variational Autoencoders
Tianyu Xie (Peking University), Cheng Zhang (Peking University)
GenerationData SynthesisOptimizationRecurrent Neural NetworkGraph Neural NetworkAuto EncoderGraphSequential
🎯 What it does: PhyloVAE is proposed, a variational autoencoder framework for unsupervised learning and generating phylogenetic tree topologies.
PhyMPGN: Physics-encoded Message Passing Graph Network for spatiotemporal PDE systems
Bocheng Zeng (Renmin University of China), Hao Sun (Renmin University of China)
Graph Neural NetworkGraphTime SeriesPhysics RelatedOrdinary Differential Equation
🎯 What it does: A physical encoding message passing graph network (PhyMPGN) is designed and implemented for efficiently predicting the evolution of space-time PDE systems under coarse grids with only a small amount of training data.
PhysBench: Benchmarking and Enhancing Vision-Language Models for Physical World Understanding
Wei Chow (University of Southern California), Yue Wang (University of Southern California)
OptimizationRobotic IntelligenceTransformerLarge Language ModelVision Language ModelImageVideoTextMultimodalityBenchmarkPhysics RelatedChain-of-Thought
🎯 What it does: This paper proposes the PhysBench benchmark and the PhysAgent framework to evaluate and enhance the capabilities of visual-language models (VLM) in understanding the physical world.
Physics of Language Models: Part 2.1, Grade-School Math and the Hidden Reasoning Process
Tian Ye (Meta), Zeyuan Allen-Zhu (Meta)
TransformerLarge Language ModelTextPhysics RelatedChain-of-Thought
🎯 What it does: This study investigates how language models solve elementary school math problems, constructs a large-scale, leak-free synthetic dataset called iGSM, and trains and analyzes the reasoning process of GPT-2.
Physics of Language Models: Part 2.2, How to Learn From Mistakes on Grade-School Math Problems
Tian Ye (Meta), Zeyuan Allen-Zhu (Meta)
TransformerLarge Language ModelSupervised Fine-TuningTextPhysics Related
🎯 What it does: In the pre-training phase, adding error + immediate correction (retry) data improves the accuracy of the language model on elementary school math reasoning tasks.
Physics of Language Models: Part 3.2, Knowledge Manipulation
Zeyuan Allen-Zhu (Meta), Yuanzhi Li (Mohamed bin Zayed University of Artificial Intelligence)
ClassificationRetrievalTransformerLarge Language ModelSupervised Fine-TuningTextPhysics RelatedChain-of-Thought
🎯 What it does: Through controlled experiments on pre-trained Transformer models, this study investigates their performance in four types of knowledge operation tasks: retrieval, classification, comparison, and reverse search. It reveals that even when the model has mastered a large amount of factual knowledge, it struggles to effectively operate this knowledge during reasoning. It also points out that using Chain-of-Thought (CoT) prompts or training during reasoning can partially address the issue, while reverse search almost completely fails.
Physics of Language Models: Part 3.3, Knowledge Capacity Scaling Laws
Zeyuan Allen-Zhu (Fair at Meta), Yuanzhi Li (Mohamed bin Zayed University of Artificial Intelligence)
TransformerLarge Language ModelMixture of ExpertsTextBiomedical DataPhysics Related
🎯 What it does: This study investigates the expansion规律 of knowledge storage capacity in language models after sufficient training, proposing and validating the universal law that 'each parameter can store 2 bits of knowledge';
Physics-aligned field reconstruction with diffusion bridge
Zeyu Li (Beihang University), Lijun Yang (Beihang University)
Diffusion modelScore-based ModelTime SeriesPhysics Related
🎯 What it does: This paper proposes the Physics-aligned Schrödinger Bridge (PalSB) framework for reconstructing physical fields from sparse measurements.
Physics-Informed Deep Inverse Operator Networks for Solving PDE Inverse Problems
Sung Woong Cho (Korea Advanced Institute of Science and Technology), Hwijae Son (Konkuk University)
TabularPhysics Related
🎯 What it does: Proposes Physics-Informed Deep Inverse Operator Networks (PI-DIONs), an unsupervised PDE inverse problem operator learning framework;
Physics-Informed Diffusion Models
Jan-Hendrik Bastek (ETH Zurich), Dennis Kochmann
GenerationOptimizationDiffusion modelTabularPhysics Related
🎯 What it does: A physical information diffusion model (PIDM) is proposed, which incorporates PDE residual loss into the training of the diffusion model, ensuring that generated samples maintain diversity while satisfying physical constraints.
Physics-informed Temporal Difference Metric Learning for Robot Motion Planning
Ruiqi Ni (Purdue University), Ahmed H Qureshi
OptimizationRobotic IntelligenceReinforcement LearningPoint CloudPhysics Related
🎯 What it does: This paper proposes a self-supervised time difference metric learning method that generates the value function for robot motion planning by precisely solving the Eikonal equation, and utilizes sampling-based MPC for path inference.
Physiome-ODE: A Benchmark for Irregularly Sampled Multivariate Time-Series Forecasting Based on Biological ODEs
Christian Klötergens (International School of Machine Learning and Learning Systems Volkswagen Financial Services Data Analytics Research Center University of Hildesheim), Lars Schmidt-Thieme (International School of Machine Learning and Learning Systems Volkswagen Financial Services Data Analytics Research Center University of Hildesheim)
Time SeriesBiomedical DataBenchmarkOrdinary Differential Equation
🎯 What it does: A benchmark called Physiome-ODE is proposed and implemented, generating 50 irregular multivariate time series datasets using biological ODE models, and various IMTS prediction models are evaluated on this benchmark.
PhysPDE: Rethinking PDE Discovery and a Physical Hypothesis Selection Benchmark
Mingquan Feng (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
Time SeriesBenchmarkPhysics Related
🎯 What it does: By introducing physical assumptions and prior knowledge, a new task framework for recovering and interpreting PDEs from experimental observation data (PDE interpretation) is proposed, and a mixed-integer programming solution based on decision forests is implemented.
PianoMotion10M: Dataset and Benchmark for Hand Motion Generation in Piano Performance
Qijun Gan (Zhejiang University), Jianke Zhu (Zhejiang University)
GenerationPose EstimationTransformerDiffusion modelVideoBenchmarkAudio
🎯 What it does: A PianoMotion10M dataset containing 116 hours and 10 million frames of hand poses has been constructed, and a two-stage gesture generation baseline model based on position prediction and diffusion generation has been proposed.
PICASO: Permutation-Invariant Context Composition with State Space Models
Tian Yu Liu (University of California Los Angeles), Stefano Soatto (Amazon Web Services AI Labs)
GenerationRetrievalComputational EfficiencyTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes a retrieval-augmented generation framework called PICASO based on the State Space Model (SSM), which utilizes pre-computed contextual state vectors to quickly combine and serve as generation conditions during inference, addressing the computational bottleneck of multi-context concatenation and the uncertainty in ordering.
PiCO: Peer Review in LLMs based on Consistency Optimization
Kun-Peng Ning (Peking University), Li Yuan (Peking University)
OptimizationTransformerLarge Language ModelText
🎯 What it does: A completely unsupervised LLM evaluation framework called PiCO is designed, utilizing a peer review mechanism of LLM mutual evaluation to generate model rankings.
PIED: Physics-Informed Experimental Design for Inverse Problems
Apivich Hemachandra (National University of Singapore), Bryan Kian Hsiang Low (National University of Singapore)
OptimizationMeta LearningReinforcement LearningTime SeriesPhysics Related
🎯 What it does: This paper proposes the PIED framework, which utilizes Physics-Informed Neural Networks (PINN) for both forward simulation and inverse problem solving, optimizing observation configurations (such as sensor locations) through gradient optimization for one-time experimental deployments.
PIG: Physics-Informed Gaussians as Adaptive Parametric Mesh Representations
Namgyu Kang (Yonsei University), Eunbyung Park (Yonsei University)
Gaussian SplattingMeshPhysics Related
🎯 What it does: A Physics-Informed Gaussian (PIG) neural network based on learnable Gaussian functions is proposed, which achieves adaptive grid representation for solving partial differential equations through dynamically adjustable Gaussian positions and covariances.
PIN: Prolate Spheroidal Wave Function-based Implicit Neural Representations
Dhananjaya Jayasundara (Johns Hopkins University), Vishal M. Patel (Johns Hopkins University)
RestorationSuper ResolutionNeural Radiance FieldImagePoint Cloud
🎯 What it does: This study proposes an Implicit Neural Representation (INR) model (PIN) using Prolate Spheroidal Wave Function (PSWF) as the activation function.
PINP: Physics-Informed Neural Predictor with latent estimation of fluid flows
Huaguan Chen (Renmin University of China), Hao Sun (University of Chinese Academy of Sciences)
Explainability and InterpretabilityComputational EfficiencyConvolutional Neural NetworkTime SeriesPhysics Related
🎯 What it does: Developed the Physics-Informed Neural Predictor (PINP) to predict fluid dynamics time series and simultaneously estimate hidden physical quantities (velocity, pressure).
PIORF: Physics-Informed Ollivier-Ricci Flow for Long–Range Interactions in Mesh Graph Neural Networks
Youn-Yeol Yu (Yonsei University), Noseong Park (KAIST)
Graph Neural NetworkMeshGraphPhysics Related
🎯 What it does: A reconnection method that combines physical information with graph topology, called PIORF, is proposed to address the over-squashing problem in grid graph neural networks and significantly improve the accuracy of fluid dynamics simulations.
PivotMesh: Generic 3D Mesh Generation via Pivot Vertices Guidance
Haohan Weng (South China University of Technology), Jun Zhu (Tsinghua University)
GenerationTransformerAuto EncoderMesh
🎯 What it does: A Transformer framework guided by Pivot vertices is proposed to generate compact and sharp 3D meshes.
PixWizard: Versatile Image-to-Image Visual Assistant with Open-Language Instructions
Weifeng Lin (Chinese University of Hong Kong), Hongsheng Li (Chinese University of Hong Kong)
Image TranslationRestorationGenerationTransformerDiffusion modelAuto EncoderImageText
🎯 What it does: PixWizard has been developed, a versatile image-to-image visual assistant that supports various tasks such as image generation, editing, and translation, and can interact through open language instructions.
Planning Anything with Rigor: General-Purpose Zero-Shot Planning with LLM-based Formalized Programming
Yilun Hao (Massachusetts Institute of Technology), Chuchu Fan (Massachusetts Institute of Technology)
OptimizationTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: A general zero-shot planning framework LLMFP is proposed, which utilizes large language models to automatically convert planning tasks from natural language descriptions into constraint optimization problems, and solves them using an SMT solver to obtain optimal plans.
Planning in Natural Language Improves LLM Search for Code Generation
Evan Z Wang, Hugh Zhang (Scale AI)
GenerationAI Code AssistantTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: A code generation method based on natural language plan search called PLANSEARCH is proposed, which utilizes multi-level observation and combination to generate diverse plans that are then transformed into code;
Plastic Learning with Deep Fourier Features
Alex Lewandowski (University of Alberta), Marlos C. Machado (University of Alberta)
ClassificationConvolutional Neural NetworkImage
🎯 What it does: This paper studies the problem of continual plasticity, proving that linear and deep linear networks do not lose trainability in continual learning, and proposes the use of deep Fourier features with sine and cosine at each layer to maintain the plasticity of the network.
pMoE: Prompting Diverse Experts Together Wins More in Visual Adaptation
Shentong Mo (Carnegie Mellon University), Dongsheng Li (Microsoft Research)
ClassificationSegmentationDomain AdaptationTransformerPrompt EngineeringMixture of ExpertsImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A Mixture-of-Experts Prompt Tuning (pMoE) framework is proposed, which dynamically integrates multi-domain expert knowledge for efficient adaptation of visual models through expert-specific prompt tokens and a learnable dispatcher.
PN-GAIL: Leveraging Non-optimal Information from Imperfect Demonstrations
Qiang Liu (Nanjing University), Daoyi Dong (University of Technology Sydney)
Robotic IntelligenceReinforcement LearningGenerative Adversarial NetworkSequential
🎯 What it does: A generative adversarial imitation learning method named PN-GAIL is proposed, which can learn near-optimal policies in the presence of non-optimal examples by utilizing negative risk information and a small amount of demonstration data with confidence labels.
PnP-Flow: Plug-and-Play Image Restoration with Flow Matching
Ségolène Tiffany Martin (Technische Universität Berlin), Gabriele Steidl (Technische Universität Berlin)
RestorationFlow-based ModelImage
🎯 What it does: An algorithm named PnP Flow Matching is proposed to solve inverse problems in image restoration, combining a pre-trained flow matching model with the PnP framework.
POGEMA: A Benchmark Platform for Cooperative Multi-Agent Pathfinding
Alexey Skrynnik (Artificial Intelligence Research Institute), Aleksandr Panov (Artificial Intelligence Research Institute)
OptimizationRobotic IntelligenceReinforcement LearningBenchmark
🎯 What it does: Constructed and released the POGEMA benchmark platform, which includes a fast learning environment, problem instance generator, visualization tools, and automatic evaluation tools for the unified assessment of multi-agent pathfinding (MAPF and LMAPF) learning, planning, and hybrid methods.
Point Cluster: A Compact Message Unit for Communication-Efficient Collaborative Perception
Zihan Ding (Beihang University), Xu Zhou (Shanghai Jiao Tong University)
CompressionAutonomous DrivingSimultaneous Localization and MappingPoint Cloud
🎯 What it does: A new message unit point cluster (Point Cluster) is proposed, and a collaborative perception framework CPPC is constructed based on it.
Point-based Instance Completion with Scene Constraints
Wesley Khademi (Oregon State University), Fuxin Li
RestorationGenerationConvolutional Neural NetworkPoint CloudMesh
🎯 What it does: A framework for instance scene completion based on sparse point cloud constraints is proposed, which can achieve high-quality completion of objects in the scene under arbitrary poses and scales, and a new dataset called ScanWCF is constructed for aligned and collision-free data.
Point-SAM: Promptable 3D Segmentation Model for Point Clouds
Yuchen Zhou (UC San Diego), Hao Su (UC San Diego)
SegmentationTransformerPrompt EngineeringPoint Cloud
🎯 What it does: A Transformer-based 3D point cloud prompt segmentation model called Point-SAM is proposed, and zero-shot point prompt segmentation is achieved through data augmentation with pseudo-labels.
PointOBB-v2: Towards Simpler, Faster, and Stronger Single Point Supervised Oriented Object Detection
Botao Ren (Tsinghua University), Zhidong Deng (Wuhan University)
Object DetectionConvolutional Neural NetworkSupervised Fine-TuningPoint Cloud
🎯 What it does: PointOBB-v2 is proposed, which generates pseudo-rotating boxes using only single-point supervision to accomplish directional object detection.
Poison-splat: Computation Cost Attack on 3D Gaussian Splatting
Jiahao Lu (National University of Singapore), Shuicheng YAN
Computational EfficiencyAdversarial AttackGaussian SplattingImage
🎯 What it does: This paper proposes a training data poisoning attack called Poison-splat, which leverages the adjustable complexity characteristics of the 3D Gaussian Splatting system, resulting in extremely high GPU memory usage and time consumption during model training, potentially triggering service interruptions.
Poisson-Dirac Neural Networks for Modeling Coupled Dynamical Systems across Domains
Razmik Arman Khosrovian (Osaka University), Takashi Matsubara (Hokkaido University)
Time SeriesSequentialPhysics Related
🎯 What it does: Proposes Poisson-Dirac Neural Networks (PoDiNNs), a unified model based on the Dirac structure for learning coupled dynamical systems from data.
PolaFormer: Polarity-aware Linear Attention for Vision Transformers
Weikang Meng (Harbin Institute of Technology), Zheng Zhang (UQMMLab University of Queensland)
ClassificationObject DetectionSegmentationTransformerImage
🎯 What it does: A polar-aware linear attention mechanism (Polafomer) is proposed to address the issues of insufficient representation caused by negative value loss and high attention entropy in traditional linear attention, significantly enhancing the performance and efficiency of visual Transformers.
Policy Decorator: Model-Agnostic Online Refinement for Large Policy Model
Xiu Yuan (University of California San Diego), Hao Su (University of California San Diego)
Robotic IntelligenceReinforcement LearningSequential
🎯 What it does: Proposes the Policy Decorator framework, which adaptively refines large-scale offline imitation learning models through online RL learning of residual policies.
Policy Design in Long-run Welfare Dynamics
Jiduan Wu (Max Planck Institute for Intelligent Systems), Ana-Andreea Stoica (Max Planck Institute for Intelligent Systems)
OptimizationTabularStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper establishes a multi-agent stochastic dynamic model to evaluate the long-term social welfare performance of two types of policies: Rawlsian (focusing on the weakest) and utilitarian (focusing on maximum immediate welfare) in the context of long-term welfare dynamics.
Policy Optimization under Imperfect Human Interactions with Agent-Gated Shared Autonomy
Zhenghai Xue (Nanyang Technological University), Shuicheng YAN
Autonomous DrivingOptimizationReinforcement LearningSequential
🎯 What it does: Designed and implemented an Agent-Gated Shared Autonomy framework (AGSA) that achieves reward-free, sample-efficient, and safe RL training by utilizing human evaluations and preference feedback.
PolyhedronNet: Representation Learning for Polyhedra with Surface-attributed Graph
Dazhou Yu (Emory University), Liang Zhao (Emory University)
RetrievalRepresentation LearningGraph Neural NetworkMesh
🎯 What it does: The PolyhedronNet framework is proposed to achieve end-to-end vector representation learning for 3D polyhedra.
PolyNet: Learning Diverse Solution Strategies for Neural Combinatorial Optimization
André Hottung (Bielefeld University), Kevin Tierney (Bielefeld University)
OptimizationTransformerReinforcement Learning
🎯 What it does: A single-decoder neural network model named PolyNet has been developed to learn multiple solution strategies, enabling more efficient exploration and search in the construction of combinatorial optimization problems (TSP, CVRP, CVRPTW, FFSP).
Polynomial Composition Activations: Unleashing the Dynamics of Large Language Models
Zhijian Zhuo (Peking University), Jinwen Ma (Peking University)
TransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: A class of polynomial combination activation functions named PolyCom (PolyReLU, PolyNorm) is proposed and implemented, and its effectiveness is validated in large language models (1B dense models and 1B/7B Mixture of Experts models).
PolyPythias: Stability and Outliers across Fifty Language Model Pre-Training Runs
Oskar van der Wal (University of Amsterdam), Stella Biderman (EleutherAI)
TransformerLarge Language ModelText
🎯 What it does: By releasing PolyPythias, we conducted training with 45 random seeds across five model sizes (14M to 410M parameters), generating approximately 7k checkpoints to systematically analyze the stability of pre-training and its impact on downstream performance, representational features, and training phases.
Polyrating: A Cost-Effective and Bias-Aware Rating System for LLM Evaluation
Jasper Dekoninck (ETH Zurich), Martin Vechev (ETH Zurich)
Large Language ModelText
🎯 What it does: A diversified, bias-aware LLM evaluation scoring system called POLYRATING has been developed.
PooDLe🐩: Pooled and dense self-supervised learning from naturalistic videos
Alex N Wang, Mengye Ren (New York University)
Object DetectionSegmentationAutonomous DrivingRepresentation LearningConvolutional Neural NetworkContrastive LearningOptical FlowVideo
🎯 What it does: A self-supervised learning framework PooDLe is designed, integrating dense optical flow equivalence objectives with a pseudo-label sub-crop pooling objective based on optical flow sampling, and introducing a lightweight spatial decoder to enhance the representation of small targets in natural videos.
Population Transformer: Learning Population-level Representations of Neural Activity
Geeling Chau (California Institute of Technology), Andrei Barbu (Massachusetts Institute of Technology)
Anomaly DetectionRepresentation LearningTransformerContrastive LearningTime SeriesBiomedical Data
🎯 What it does: This paper proposes a self-supervised Population Transformer (PopT) that aggregates multi-channel neural data based on pre-trained time series embeddings, capable of handling arbitrary channel combinations.
Port-Hamiltonian Architectural Bias for Long-Range Propagation in Deep Graph Networks
Simon Heilig (Ruhr University Bochum), Davide Bacciu (University of Pisa)
Graph Neural NetworkGraph
🎯 What it does: A deep graph network based on port-Hamiltonian dynamics (PH-DGN) is proposed, which achieves long-distance information propagation by introducing energy conservation and non-conservative dynamics into graph neural networks, enhancing performance without the need for global encoding or rearrangement.
PortLLM: Personalizing Evolving Large Language Models with Training-Free and Portable Model Patches
Rana Shahroz, Tianlong Chen (University of North Carolina at Chapel Hill)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Proposes the PORTLLM framework, which utilizes untrained model patches to transfer domain knowledge from older versions of LLMs to newer versions, achieving continuous personalization.
Positive-Unlabeled Diffusion Models for Preventing Sensitive Data Generation
Hiroshi Takahashi (NTT Corporation), Tomoya Yamashita (NTT Corporation)
GenerationData SynthesisSafty and PrivacyDiffusion modelImage
🎯 What it does: A positive-negative unlabeled diffusion model is proposed, which trains the diffusion model using a small number of sensitive samples and unlabeled data to avoid generating sensitive content.
Post-hoc Reward Calibration: A Case Study on Length Bias
Zeyu Huang (University of Edinburgh), Ivan Titov (University of Edinburgh)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: A Post-hoc Reward Calibration method is proposed to correct the biases of the Reward Model in features such as length, avoiding reward hijacking and improving the alignment and evaluation quality of LLMs.
PostCast: Generalizable Postprocessing for Precipitation Nowcasting via Unsupervised Blurriness Modeling
Junchao Gong (Shanghai Jiao Tong University), LEI BAI
RestorationGenerationDiffusion modelTime Series
🎯 What it does: This paper proposes a post-processing method named PostCast, aimed at eliminating the ambiguity in precipitation nowcasting and improving the prediction accuracy of extreme precipitation events.
PostEdit: Posterior Sampling for Efficient Zero-Shot Image Editing
Feng Tian (Shanghai Jiao Tong University), Xiaokang Yang (Shanghai Jiao Tong University)
Image TranslationGenerationComputational EfficiencyDiffusion modelImageBenchmarkStochastic Differential Equation
🎯 What it does: A novel non-reverse, non-training image editing method called PostEdit is proposed, which incorporates measurement terms and Langevin dynamics into the diffusion process using posterior sampling theory to achieve efficient zero-shot image editing while maintaining background consistency.
Posterior-Mean Rectified Flow: Towards Minimum MSE Photo-Realistic Image Restoration
Guy Ohayon (Technion Israel Institute of Technology), Michael Elad (Technion Israel Institute of Technology)
RestorationRectified FlowImageOrdinary Differential Equation
🎯 What it does: The Posterior-Mean Rectified Flow (PMRF) algorithm is proposed and implemented, utilizing posterior mean prediction and Rectified Flow for image recovery, directly approximating the optimal estimator that minimizes mean squared error (MSE) under perfect perception constraints.
POTEC: Off-Policy Contextual Bandits for Large Action Spaces via Policy Decomposition
Yuta Saito (Cornell University), Thorsten Joachims (Cornell University)
Recommendation SystemReinforcement LearningTabular
🎯 What it does: A two-stage contextual bandit learning framework called POTEC is proposed, which first selects an action set through clustering and then uses regression to select specific actions within the clusters.
PPT: Patch Order Do Matters In Time Series Pretext Task
Jaeho Kim (Ulsan National Institute of Science and Technology), Seulki Lee (Ulsan National Institute of Science and Technology)
ClassificationRepresentation LearningRecurrent Neural NetworkTransformerContrastive LearningTime SeriesElectrocardiogram
🎯 What it does: This paper proposes the Patch Order-Aware Pretext Task (PPT), which enhances temporal classification performance by learning patch order information through weak/strong permutation of time series patches along the channel dimension and utilizing consistency and contrastive loss.
PQMass: Probabilistic Assessment of the Quality of Generative Models using Probability Mass Estimation
Pablo Lemos (Sandbox Quantum), Yashar Hezaveh (Trottier Space Institute)
GenerationAnomaly DetectionImageMultimodalityTabularTime Series
🎯 What it does: A likelihood-free, model-free training method for probability mass estimation, PQMass, is proposed to determine whether two sets of samples come from the same distribution.
PRDP: Progressively Refined Differentiable Physics
Kanishk Bhatia (Technical University of Munich), Nils Thuerey (Technical University of Munich)
OptimizationComputational EfficiencySequentialPhysics Related
🎯 What it does: This study investigates the use of a differentiable physical solver with incomplete convergence during neural network training and proposes an adaptive progressive refinement method (PRDP) that significantly reduces the number of iterations and training time of the physical solver while maintaining network accuracy.
Preble: Efficient Distributed Prompt Scheduling for LLM Serving
Vikranth Srivatsa (University of California San Diego), Yiying Zhang (University of California San Diego)
OptimizationComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringVideoText
🎯 What it does: This study investigates large-scale LLM servers for long shared prompts and proposes and implements a distributed scheduling system called Preble, aimed at significantly reducing inference latency for long prompt tasks.
Precedence-Constrained Winter Value for Effective Graph Data Valuation
Hongliang Chi (Rensselaer Polytechnic Institute), Yao Ma (IBM)
Graph Neural NetworkGraph
🎯 What it does: A value assessment method for graph data called PC-Winter is proposed, which quantifies the contribution of nodes (including unlabeled nodes) and edges in the graph to the performance of graph neural networks.
Precise Localization of Memories: A Fine-grained Neuron-level Knowledge Editing Technique for LLMs
Haowen Pan (University of Science and Technology of China), Meng Wang (Hefei University of Technology)
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: A knowledge editing method based on neuron-level fine-grained localization, FiNE, is proposed, which directly fine-tunes specific neurons in the FFN to achieve precise updates of knowledge in large language models.
Precise Parameter Localization for Textual Generation in Diffusion Models
Łukasz Staniszewski (Warsaw University of Technology), Adam Dziedzic (CISPA Helmholtz Center for Information Security)
GenerationTransformerDiffusion modelText
🎯 What it does: By activating patch technology, we locate the few attention layers responsible for text generation in diffusion models, and use the localization results for fine-tuning, text editing, and harmful text protection.
Predicate Hierarchies Improve Few-Shot State Classification
Emily Jin (Stanford University), Jiajun Wu (Stanford University)
ClassificationRobotic IntelligenceLarge Language ModelContrastive LearningImage
🎯 What it does: A model named PHIER has been developed to enhance state classification performance under few-shot conditions using a predicate hierarchical structure.
Predicting the Energy Landscape of Stochastic Dynamical System via Physics-informed Self-supervised Learning
Ruikun Li (Tsinghua University), Yong Li (Tsinghua University)
Graph Neural NetworkGraphPhysics RelatedStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Using self-supervised evolutionary prediction tasks to estimate the energy landscape of stochastic dynamical systems without requiring true energy labels.
Prediction Risk and Estimation Risk of the Ridgeless Least Squares Estimator under General Assumptions on Regression Errors
Sungyoon Lee (Hanyang University), Sokbae Lee (Columbia University)
TabularTime Series
🎯 What it does: This paper studies the prediction risk and estimation risk of the minimum ‖2‖ norm (ridgeless) least squares estimator in over-parameterized linear regression under general error structures, and provides its precise finite sample characterization.
Predictive Inverse Dynamics Models are Scalable Learners for Robotic Manipulation
Yang Tian (Shanghai AI Laboratory), Jiangmiao Pang (Shanghai AI Laboratory)
Robotic IntelligenceTransformerReinforcement LearningMultimodality
🎯 What it does: This paper proposes an end-to-end Predictive Inverse Dynamics Model (PIDM) called Seer, which combines conditional visual foresight with inverse dynamics prediction to learn actions in a closed-loop manner for robotic manipulation tasks. By pre-training on a large-scale robotic dataset and fine-tuning on downstream tasks, it achieves efficient and scalable control strategies.
Predictive Uncertainty Quantification for Bird's Eye View Segmentation: A Benchmark and Novel Loss Function
Linlin Yu (University of Texas at Dallas), Feng Chen (University of Texas at Dallas)
SegmentationAutonomous DrivingImageBenchmark
🎯 What it does: A benchmark for uncertainty quantification was established in the bird's-eye view semantic segmentation task, and the UFCE loss and its related regularization framework were proposed to enhance model calibration and uncertainty prediction.
Preference Diffusion for Recommendation
Shuo Liu (East China Normal University), Tat-Seng Chua (National University of Singapore)
Recommendation SystemDiffusion modelSequential
🎯 What it does: Proposes the PreferDiff optimization objective to improve the training of diffusion models in sequential recommendation, balancing generation and personalized ranking.
Preference Elicitation for Offline Reinforcement Learning
Alizée Pace (ETH Zurich), Giorgia Ramponi (University of Zurich)
Reinforcement LearningTabularBenchmark
🎯 What it does: A completely offline preference-guided reinforcement learning algorithm Sim-OPRL is proposed, which utilizes a learned environment model for trajectory simulation, combining conservative and optimistic strategies for preference sampling, thereby learning the optimal policy in situations where interaction with the environment is not possible.
Preference Optimization for Reasoning with Pseudo Feedback
Fangkai Jiao (Nanyang Technological University), Furu Wei (Microsoft Research)
OptimizationAI Code AssistantTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes a method for preference optimization through the generation of pseudo-feedback (pseudo test cases), thereby enhancing the performance of large language models in mathematical reasoning and code generation tasks.
Preserving Deep Representations in One-Shot Pruning: A Hessian-Free Second-Order Optimization Framework
Ryan Lucas (Massachusetts Institute of Technology), Rahul Mazumder (Massachusetts Institute of Technology)
OptimizationRepresentation LearningConvolutional Neural NetworkTransformerImage
🎯 What it does: A one-shot post-training pruning framework called SNOWS is proposed, aiming to significantly improve the performance of pruned models by adjusting the weights of existing sparse masks without retraining, thereby preserving deep network representations.
Preserving Diversity in Supervised Fine-Tuning of Large Language Models
Ziniu Li (Chinese University of Hong Kong), Ruoyu Sun (Chinese University of Hong Kong)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: The GEM algorithm is proposed, which is a distribution matching framework based on game theory. It achieves sparse updates and adaptive termination in the supervised fine-tuning of LLMs through auxiliary variables to maintain output diversity.
Presto! Distilling Steps and Layers for Accelerating Music Generation
Zachary Novack (University of California San Diego), Nicholas J. Bryan (Adobe Research)
GenerationComputational EfficiencyKnowledge DistillationTransformerDiffusion modelGenerative Adversarial NetworkAudio
🎯 What it does: A dual distillation method (Presto) is proposed to accelerate Transformer-based continuous-time diffusion models, reducing both sampling steps and computational costs per step, thereby achieving faster and higher-quality music generation.
Prevalence of Negative Transfer in Continual Reinforcement Learning: Analyses and a Simple Baseline
Hongjoon Ahn (Seoul National University), Taesup Moon (Seoul National University)
Knowledge DistillationReinforcement LearningSequential
🎯 What it does: This study focuses on the problem of negative transfer in continual reinforcement learning and proposes a simple baseline method called R&D based on reset and distillation.
Prioritized Generative Replay
Renhao Wang (University of California), Alexei A Efros
GenerationReinforcement LearningDiffusion modelImageSequential
🎯 What it does: A priority generative replay framework based on conditional diffusion models is proposed to more effectively replay experiences in online reinforcement learning.
PRISM: Privacy-Preserving Improved Stochastic Masking for Federated Generative Models
Kyeongkook Seo (Ulsan National Institute of Science and Technology), Jaejun Yoo (Ulsan National Institute of Science and Technology)
GenerationFederated LearningSafty and PrivacyGenerative Adversarial NetworkImage
🎯 What it does: A PRISM framework is proposed in the federated learning environment, achieving high-quality image generation by finding sparse sub-networks, using MMD loss, and dynamic moving average aggregation, significantly reducing communication overhead.
Privacy Auditing of Large Language Models
Ashwinee Panda (Princeton University), Prateek Mittal (Princeton University)
Safty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper designs more memorable 'watermark' texts (canaries) to conduct more effective audits of privacy leakage in large language models (LLMs), achieving practical black-box privacy auditing especially in differential privacy (DP) training scenarios.
Privacy-Aware Lifelong Learning
Ozan Ozdenizci (Montanuniversität Leoben), Robert Legenstein (Graz University of Technology)
Safty and PrivacyKnowledge DistillationConvolutional Neural NetworkTransformerImage
🎯 What it does: A privacy-aware lifelong learning (PALL) framework is proposed, achieving non-catastrophic forgetting, selective task forgetting (complete task forgetting) in task incremental learning while maintaining knowledge transfer and keeping low model memory usage in fixed capacity networks.
Privacy-Preserving Personalized Federated Prompt Learning for Multimodal Large Language Models
Linh Tran (Rensselaer Polytechnic Institute), Ana Milanova (Rensselaer Polytechnic Institute)
Federated LearningSafty and PrivacyTransformerLarge Language ModelPrompt EngineeringImageMultimodality
🎯 What it does: Prompt learning for multimodal large language models is conducted within the federated learning framework, introducing differential privacy to protect user data.
Private Mechanism Design via Quantile Estimation
Yuanyuan Yang (University of Washington), Jamie Heather Morgenstern
OptimizationSafty and PrivacyTabular
🎯 What it does: An approximately optimal revenue mechanism that guarantees pure differential privacy (ε, 0-DP) is designed for single-item auctions, supporting independent but non-identically distributed value distributions that can be bounded or unbounded.
Privately Counting Partially Ordered Data
Matthew Joseph (Google Research), Alexander Yu (Google Research)
Safty and PrivacyComputational EfficiencyGraph
🎯 What it does: An efficient pure differential privacy K-norm mechanism for counting partially ordered data has been developed, capable of sampling its induced norm ball in O(d²) time.
Proactive Agent: Shifting LLM Agents from Reactive Responses to Active Assistance
Yaxi Lu (Tsinghua University), Maosong Sun (Tsinghua University)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AITextBenchmark
🎯 What it does: A ProactiveBench dataset was constructed and an LLM agent was trained to proactively predict and provide tasks that users may need without explicit instructions.
Proactive Privacy Amnesia for Large Language Models: Safeguarding PII with Negligible Impact on Model Utility
Martin Kuo (Duke University), Hai Li (Duke University)
Safty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A Proactive Privacy Amnesia (PPA) method is designed, which first locates key tokens in PII sequences through sensitivity analysis, and then selectively forgets only these key tokens while compensating for the performance loss of the model through memory implantation.
ProAdvPrompter: A Two-Stage Journey to Effective Adversarial Prompting for LLMs
Hao Di (Xi'an Jiaotong University), Ivor Tsang
Adversarial AttackTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: ProAdvPrompter, a two-stage method, is proposed to construct aligned LLM attacker prompters and significantly improve the attack success rate.
Probabilistic Conformal Prediction with Approximate Conditional Validity
Vincent Plassier (Lagrange Mathematics and Computing Research Center), Eric Moulines (École Polytechnique)
TabularTime Series
🎯 What it does: The CP2 framework is proposed, which combines split conformal prediction with conditional distribution estimation to generate an approximately conditionally valid prediction set.
Probabilistic Geometric Principal Component Analysis with application to neural data
Han-Lin Hsieh (University of Southern California), Maryam Shanechi
Biomedical Data
🎯 What it does: A probabilistic dimensionality reduction model called PGPCA has been developed for a given nonlinear manifold, and its EM learning algorithm has been provided.
Probabilistic Language-Image Pre-Training
Sanghyuk Chun (NAVER AI Lab), Sangdoo Yun (NAVER AI Lab)
ClassificationRecognitionTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: A full-probability visual-language pre-training model ProLIP is proposed, which maps images and text to Gaussian distributions and effectively estimates variance through uncertainty tokens to address the ambiguity of many-to-many matching.
Probabilistic Learning to Defer: Handling Missing Expert Annotations and Controlling Workload Distribution
Cuong C. Nguyen (University of Surrey), Gustavo Carneiro (University of Surrey)
ClassificationOptimizationMixture of ExpertsBiomedical DataMagnetic Resonance Imaging
🎯 What it does: This paper proposes a probabilistic Learning to Delay (L2D) model that can use only a portion of expert annotations during training and infer missing annotations through the EM algorithm, while incorporating constraints in the E-step to control the workload of human experts and AI classifiers.
Probabilistic Neural Pruning via Sparsity Evolutionary Fokker-Planck-Kolmogorov Equation
Zhanfeng Mo (Nanyang Technological University), Sinno Jialin Pan (Chinese University of Hong Kong)
OptimizationImageStochastic Differential Equation
🎯 What it does: This paper proposes a probability neural pruning framework based on the Sparse Evolving Fokker-Planck-Kolmogorov equation (SFPK) and implements the corresponding particle simulation pruning algorithm SFPK-pruner.
Probe before You Talk: Towards Black-box Defense against Backdoor Unalignment for Large Language Models
Biao Yi (Nankai University), Yiming Li (Nanyang Technological University)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes a black-box input-level defense method named BEAT, which can detect and suppress security failures caused by backdoor unalignment during LLM inference.
Probe Pruning: Accelerating LLMs through Dynamic Pruning via Model-Probing
Qi Le (University of Minnesota), Ali Anwar (University of Minnesota)
GenerationOptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: A novel online dynamic structured pruning framework called Probe Pruning is proposed, which guides weight channel pruning through small-scale probing for each batch while maintaining efficiency during the full inference phase.
Probing the Latent Hierarchical Structure of Data via Diffusion Models
Antonio Sclocchi (Institute of Physics, École Polytechnique Fédérale de Lausanne), Matthieu Wyart (Johns Hopkins University)
GenerationData SynthesisDiffusion modelImageText
🎯 What it does: The forward-backward experiments of diffusion models were studied, utilizing this process to quantify the hierarchical structure of data, and validating the prediction that the relevant block length diverges with noise changes and at phase transitions on synthetic, text, and image data.
Problem-Parameter-Free Federated Learning
Wenjing Yan (Chinese University of Hong Kong), Xuanyu Cao (Washington State University)
OptimizationFederated LearningImage
🎯 What it does: A problem-free parameter Federated Learning algorithm PAdaMFed is designed, which can automatically adjust the step size under non-convex objectives and handle arbitrary heterogeneous data.
Procedural Knowledge in Pretraining Drives Reasoning in Large Language Models
Laura Ruis (University College London), Max Bartolo (Cohere)
TransformerLarge Language ModelText
🎯 What it does: This study investigates how large language models utilize pre-training data in reasoning tasks, using influence functions to measure the impact of each document on model inference and fact retrieval outputs.
Procedural Synthesis of Synthesizable Molecules
Michael Sun (Massachusetts Institute of Technology), Wojciech Matusik (Massachusetts Institute of Technology)
OptimizationDrug DiscoveryGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: A dual-layer program synthesis framework is designed to generate synthesizable molecules and their analogs, optimizing molecular structures under given property targets.
Process Reward Model with Q-value Rankings
Wendi Li (Huazhong University of Science and Technology), Yixuan Li (University of Wisconsin Madison)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This paper proposes the Process Q-value Model (PQM), which transforms the modeling of process rewards into a Q-value ranking problem and designs a margin-based comparison loss to train the model.
Programming Refusal with Conditional Activation Steering
Bruce W. Lee (University of Pennsylvania), Amit Dhurandhar (IBM Research)
Large Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: This paper proposes Conditional Activation Steering (CAST), a method that utilizes hidden layer activation vectors to achieve context-aware, controllable rejection, and directed behavior without updating model weights.
Progress or Regress? Self-Improvement Reversal in Post-training
Ting Wu (Fudan University), Pengfei Liu (Shanghai Jiao Tong University)
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: This study investigates the self-improvement effects of large language models through iterative post-training and proposes a more comprehensive evaluation framework to detect potential degradation phenomena beyond accuracy improvement.