NeurIPS 2024 Papers — Page 28
Conference on Neural Information Processing Systems · 4035 papers
Parallelizing Linear Transformers with the Delta Rule over Sequence Length
Songlin Yang (Massachusetts Institute of Technology), Yoon Kim (Massachusetts Institute of Technology)
TransformerText
🎯 What it does: Parallel training of the DeltaNet linear transformer is conducted, proposing a memory-efficient reparameterization based on the general Householder transformation and WY representation, and implementing chunkwise forward/backward computation.
Parallelizing Model-based Reinforcement Learning Over the Sequence Length
ZiRui Wang (Zhejiang University), Yin Zhang (Zhejiang University)
Reinforcement LearningSequentialBenchmark
🎯 What it does: The PaMoRL framework is proposed, utilizing PWM and PETE to achieve parallel training of MBRL model learning and policy learning on sequence length, enhancing hardware efficiency.
Parameter Competition Balancing for Model Merging
Guodong DU, Min Zhang (Harbin Institute of Technology)
OptimizationTransformerLarge Language ModelText
🎯 What it does: A model merging technique called PCB-MERGING is proposed, which is a training-free and data-free approach that achieves model fusion for multiple tasks, domains, and training configurations by balancing competition among task vectors at the parameter level.
Parameter Disparities Dissection for Backdoor Defense in Heterogeneous Federated Learning
Wenke Huang (Wuhan University), Bo Du (Wuhan University)
Federated LearningImage
🎯 What it does: This paper addresses the issue of backdoor attacks in heterogeneous federated learning by proposing a parameter importance evaluation and rescaling method based on the Fisher Information Matrix (FDCR). It identifies malicious clients through clustering of client gradient differences and weights important parameters during global aggregation to enhance backdoor defense effectiveness.
Parameter Efficient Adaptation for Image Restoration with Heterogeneous Mixture-of-Experts
Hang Guo (Tsinghua University), Shu-Tao Xia (Tsinghua University)
RestorationTransformerMixture of ExpertsImage
🎯 What it does: A parameter-efficient adaptation method for image restoration, AdaptIR, is proposed, which adapts a frozen pre-trained restoration model through a small number of trainable modules.
Parameter Symmetry and Noise Equilibrium of Stochastic Gradient Descent
Liu Ziyin (Massachusetts Institute of Technology), Lei Wu (Peking University)
OptimizationTabularPhysics RelatedStochastic Differential Equation
🎯 What it does: This paper studies the impact of exponential symmetry on stochastic gradient descent (SGD), derives Noether flow and noise equilibrium, and proves that each exponential symmetry guides parameters to converge to a unique equilibrium point along a degenerate direction in SGD, thereby explaining how gradient noise affects model convergence, sharpening/flattening, representation alignment, and warmup phenomena.
Parameter-free Clipped Gradient Descent Meets Polyak
Yuki Takezawa (Kyoto University), Makoto Yamada (Kyoto University)
OptimizationRecurrent Neural NetworkText
🎯 What it does: An Inexact Polyak step size is proposed, which achieves gradient clipping-style gradient descent without the need for any hyperparameter tuning, and its convergence analysis is provided.
Parameter-Inverted Image Pyramid Networks
Xizhou Zhu (Tsinghua University), Jifeng Dai (Tsinghua University)
ClassificationObject DetectionSegmentationComputational EfficiencyTransformerSupervised Fine-TuningImage
🎯 What it does: This paper proposes a Parameter Inverted Image Pyramid Network (PIIP), which uses a multi-branch structure to assign high-resolution images to small models and low-resolution images to large models, achieving multi-scale feature fusion and significantly reducing the computational cost of traditional image pyramids.
Parameterized Approximation Schemes for Fair-Range Clustering
Zhen Zhang (Hunan University of Technology and Business), Qilong Feng (Central South University)
Optimization
🎯 What it does: A (1+ε) approximate fixed-parameter feasible time algorithm is proposed for the fair range k-median and k-means problems with lower/upper bound label constraints in Euclidean space.
Parametric model reduction of mean-field and stochastic systems via higher-order action matching
Jules Berman (Courant Institute of Mathematical Sciences New York University), Benjamin Peherstorfer (Courant Institute of Mathematical Sciences New York University)
GenerationData SynthesisOptimizationComputational EfficiencyTime SeriesSequentialPhysics RelatedStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Learn a minimum energy gradient field based on optimal transport to approximate the dynamics of a population under changes in physical parameters, and use this model to quickly generate complete sample trajectories during inference.
Pard: Permutation-Invariant Autoregressive Diffusion for Graph Generation
Lingxiao Zhao (Carnegie Mellon University), Leman Akoglu (Carnegie Mellon University)
GenerationData SynthesisDrug DiscoveryGraph Neural NetworkTransformerDiffusion modelGraphBiomedical Data
🎯 What it does: The PARD model is proposed, combining autoregressive block generation with discrete diffusion to achieve interchangeability and efficiency in graph generation.
Parseval Regularization for Continual Reinforcement Learning
Wesley Chung (Mila), David Meger (Mila)
Reinforcement LearningSequential
🎯 What it does: This study investigates and verifies the use of Parseval regularization in continual reinforcement learning to maintain the orthogonality of weight matrices, thereby enhancing the network's plasticity and training speed on multi-task sequences.
Parsimony or Capability? Decomposition Delivers Both in Long-term Time Series Forecasting
Jinliang Deng (Hong Kong Generative AI Research and Development Center), Hui Xiong (Hong Kong University of Science and Technology)
Convolutional Neural NetworkTime Series
🎯 What it does: A long sequence time series prediction model based on a decomposition and selection mechanism (SSCNN) is proposed, which achieves precise modeling of complex patterns by decomposing the time series into long-term, seasonal, short-term, and spatial components.
Partial observation can induce mechanistic mismatches in data-constrained models of neural dynamics
William Qian (Harvard University), Cengiz Pehlevan (Harvard University)
Recurrent Neural NetworkTime Series
🎯 What it does: This study investigates how data-driven models (RNNs) incorrectly infer the dynamical mechanisms of neural networks when only a subset of neurons is observed, focusing on the linear approximations of linear and nonlinear systems, low-rank networks, and feedforward chains.
Partial Structure Discovery is Sufficient for No-regret Learning in Causal Bandits
Muhammad Qasim Elahi (Purdue University), Murat Kocaoglu (Purdue University)
Graph Neural NetworkReinforcement LearningGraph
🎯 What it does: This paper proposes a two-stage causal bandit algorithm: first, it efficiently learns the observable subgraph of the reward node's ancestors and the necessary potential confounding variables using randomized intervention experiments, constructing the Possible Optimal Intervention Set (POMIS); subsequently, standard bandit algorithms such as UCB are run on this set, ultimately achieving sublinear regret.
Partial Transportability for Domain Generalization
Kasra Jalaldoust (Causal Artificial Intelligence Lab Columbia University), Elias Bareinboim (Causal Artificial Intelligence Lab Columbia University)
Domain AdaptationOptimizationTabular
🎯 What it does: This paper proposes a general estimation method to provide a tight upper bound on statistics (such as the generalization error of classifiers) in the target domain under multi-source domains, thereby ensuring the performance of AI models in unknown domains.
Particle Semi-Implicit Variational Inference
Jen Ning Lim (University of Warwick), Adam Michael Johansen
GenerationOptimizationTabularStochastic Differential EquationAudio
🎯 What it does: A Particle Variational Inference (PVI) method based on particle approximation is proposed, which optimizes directly in the mixed distribution space, thus eliminating the need for implicit distributions or explicit parameterization of mixed distributions.
Paths to Equilibrium in Games
Bora Yongacoglu (University of Toronto), Serdar Yuksel
OptimizationReinforcement LearningTabular
🎯 What it does: This paper studies how agents in multi-agent reinforcement learning (MARL) and game theory form a sequence of strategies that satisfy achievement paths during continuous interaction and strategy updates, allowing them to reach Nash equilibrium from any initial strategy.
pcaGAN: Improving Posterior-Sampling cGANs via Principal Component Regularization
Matthew C Bendel, Philip Schniter (Ohio State University)
RestorationGenerationGenerative Adversarial NetworkImageMagnetic Resonance Imaging
🎯 What it does: This paper proposes a new conditional GAN posterior sampler, pcaGAN, which can generate diverse samples that conform to the posterior distribution in image restoration tasks, supporting uncertainty quantification and perceptual/distortion trade-offs.
PCoTTA: Continual Test-Time Adaptation for Multi-Task Point Cloud Understanding
Jincen Jiang (Bournemouth University), Xuequan Lu (La Trobe University)
Object DetectionSegmentationDomain AdaptationContrastive LearningPoint CloudBenchmark
🎯 What it does: The PCoTTA framework is proposed, achieving adaptive continuous testing for multi-task point cloud understanding, significantly enhancing the model's transfer performance in continuously changing target domains.
PCP-MAE: Learning to Predict Centers for Point Masked Autoencoders
Xiangdong Zhang (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
ClassificationSegmentationRepresentation LearningTransformerAuto EncoderPoint Cloud
🎯 What it does: Proposes PCP-MAE, which enhances the semantic representation of the encoder by learning to predict masked point cloud centers instead of directly leaking the centers during self-supervised pre-training.
PEAC: Unsupervised Pre-training for Cross-Embodiment Reinforcement Learning
Chengyang Ying (Tsinghua University), Jun Zhu (Tsinghua University)
Robotic IntelligenceReinforcement LearningImageVideo
🎯 What it does: A cross-body unsupervised pre-training framework CEURL is proposed, and the PEAC algorithm based on information geometric analysis is designed to achieve cross-body knowledge transfer.
Pearls from Pebbles: Improved Confidence Functions for Auto-labeling
Harit Vishwakarma (University of Wisconsin Madison), Ramya Korlakai Vinayak (University of Wisconsin Madison)
ClassificationOptimizationConvolutional Neural NetworkSupervised Fine-TuningImageText
🎯 What it does: A post-processing confidence function learning framework called Colander for threshold-based automatic labeling (TBAL) is proposed, significantly improving the coverage of automatic labeling while keeping the error within a preset threshold.
Pedestrian-Centric 3D Pre-collision Pose and Shape Estimation from Dashcam Perspective
MeiJun Wang, Jian Gao (Northwest University)
Pose EstimationAutonomous DrivingTransformerVideoMesh
🎯 What it does: The first pedestrian-vehicle collision posture dataset PVCP based on dashcam perspective has been constructed, and a dual-stage pedestrian pre-collision posture and shape estimation network PPSENet has been proposed to estimate 3D posture and SMPL mesh from collision videos.
PediatricsGPT: Large Language Models as Chinese Medical Assistants for Pediatric Applications
Dingkang Yang (Fudan University), Lihua Zhang (Fudan University)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A pediatric instruction dataset, PedCorpus 300k, was constructed, and PediatricsGPT was trained through steps including continuous pre-training, full-parameter SFT, DFPO preference alignment, and LoRA parameter-efficient SFT, enhancing pediatric diagnosis and consultation capabilities.
Penalty-based Methods for Simple Bilevel Optimization under Hölderian Error Bounds
Pengyu Chen (Fudan University), Jiulin Wang (Fudan University)
Optimization
🎯 What it does: A framework based on penalty functions is proposed to solve simple bilevel optimization problems, along with a non-asymptotic complexity analysis under Hölder error boundary conditions.
Perceiving Longer Sequences With Bi-Directional Cross-Attention Transformers
Markus Hiller (University of Melbourne), Tom Drummond (University of Melbourne)
ClassificationSegmentationRetrievalTransformerImageMultimodalityPoint Cloud
🎯 What it does: A new bidirectional cross-attention Transformer architecture, BiXT, is proposed, which achieves linear time and memory complexity while maintaining the high performance of Transformers, suitable for long sequences and multimodal tasks.
Perception of Knowledge Boundary for Large Language Models through Semi-open-ended Question Answering
Zhihua Wen (National University of Defense Technology), Dongsheng Li (National University of Defense Technology)
TransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes using semi-open-ended QA to probe the knowledge boundaries of large language models (LLMs), constructing a dataset containing multi-answer questions and mining low-probability answers that the target model cannot cover through auxiliary models.
Perceptual Fairness in Image Restoration
Guy Ohayon (Technion Israel Institute of Technology), Tomer Michaeli (Technion Israel Institute of Technology)
RestorationSuper ResolutionAdversarial AttackGenerative Adversarial NetworkImage
🎯 What it does: A new metric for Perceptual Fairness (PF) is proposed to evaluate the fairness of image restoration algorithms across different sensitive attribute groups, and its effectiveness is validated on facial image super-resolution.
PeRFlow: Piecewise Rectified Flow as Universal Plug-and-Play Accelerator
Hanshu Yan (ByteDance), Jiashi Feng (ByteDance)
GenerationData SynthesisOptimizationDiffusion modelFlow-based ModelRectified FlowImageVideoStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Proposes PeRFlow, a stepwise generation framework that segments the sampling trajectory of a pre-trained diffusion model and reconstructs within each segment;
Performative Control for Linear Dynamical Systems
Songfu Cai (Hong Kong University of Science and Technology), Xuanyu Cao (Hong Kong University of Science and Technology)
OptimizationReinforcement LearningTabularTime SeriesFinance Related
🎯 What it does: A control framework is proposed, where the strategy chosen by the controller affects the dynamics of the control system, leading to policy-dependent system state data sequences and their temporal correlations.
Peri-midFormer: Periodic Pyramid Transformer for Time Series Analysis
Qiang Wu (Xidian University), Shuyuan Yang (Xidian University)
ClassificationAnomaly DetectionTransformerTime Series
🎯 What it does: A Transformer model named Peri-midFormer is proposed, which utilizes the multi-periodicity and hierarchical inclusion relationships of time series to decompose the original sequence into a periodic pyramid, and performs self-attention modeling on this structure to accomplish multi-task time series analysis.
PERIA: Perceive, Reason, Imagine, Act via Holistic Language and Vision Planning for Manipulation
Fei Ni (Tianjin University), YAN ZHENG (Huawei Noah's Ark Lab)
Robotic IntelligenceTransformerLarge Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: The PERIA framework is proposed, which combines multimodal large language models and diffusion models to achieve global language planning and visual planning for long-duration robotic arm tasks, enhancing task execution accuracy through consistency alignment.
Periodic agent-state based Q-learning for POMDPs
Amit Sinha (McGill University), Aditya Mahajan (McGill University)
Reinforcement LearningTabular
🎯 What it does: This paper proposes a new variant of offline Q-learning—Periodic Agent State Q-learning (PASQL)—for learning periodic (non-stationary) policies in partially observable Markov decision processes (POMDPs).
Perplexity-aware Correction for Robust Alignment with Noisy Preferences
Keyi Kong (Shandong University), Mohan Kankanhalli (National University of Singapore)
Recommendation SystemOptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: In the presence of noise preferences (NPs), the PerpCorrect method is proposed, which detects and corrects erroneous labels by training a substitute LLM and utilizing the generated perplexity difference (PPLDiff) to obtain clean training data for robust alignment.
Persistence Homology Distillation for Semi-supervised Continual Learning
YanFan, Qinghua Hu (Tianjin University)
Computational EfficiencyKnowledge DistillationImage
🎯 What it does: The paper proposes a knowledge distillation method based on Persistence Homology (PsHD) for semi-supervised continual learning (SSCL) scenarios, aimed at reducing catastrophic forgetting and enhancing robustness to noise.
Persistent Homology for High-dimensional Data Based on Spectral Methods
Sebastian Damrich (Hertie Institute for AI in Brain Health), Dmitry Kobak (Hertie Institute for AI in Brain Health)
Point CloudBiomedical Data
🎯 What it does: The study uses spectral distance (effective resistance, diffusion distance) combined with persistent homology to detect the topological structure of point clouds in high-dimensional noisy environments.
Persistent Test-time Adaptation in Recurring Testing Scenarios
Trung-Hieu Hoang (University of Illinois at Urbana-Champaign), Minh N. Do (University of Illinois at Urbana-Champaign)
Domain AdaptationImage
🎯 What it does: A recurring test-time adaptation (TTA) scenario is proposed, and the collapse problem of the TTA model is systematically analyzed. Based on theoretical analysis, a persistent test-time adaptation (PeTTA) algorithm is designed.
Personalized Adapter for Large Meteorology Model on Devices: Towards Weather Foundation Models
Shengchao Chen (Australian Artificial Intelligence Institute), Chengqi Zhang (Hong Kong Polytechnic University)
Federated LearningComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTime SeriesSequential
🎯 What it does: A framework named LM-WEATHER has been developed, utilizing pre-trained language models (PLM) combined with lightweight personalized adapters to achieve localized modeling of meteorological variable sequences (prediction and missing value imputation) in a federated learning environment on distributed devices.
Personalized Federated Learning via Feature Distribution Adaptation
Connor Mclaughlin, Lili Su (Northeastern University)
Federated LearningImage
🎯 What it does: This study proposes a personalized federated learning framework named pFedFDA, which models feature distributions as generative (class-conditional Gaussian) to guide the training of shared feature extractors, and uses local-global interpolation to estimate feature distributions on the client side, thereby generating personalized classifiers for each client.
Personalized Federated Learning with Mixture of Models for Adaptive Prediction and Model Fine-Tuning
Pouya M. Ghari (University of California Irvine), Yanning Shen (University of California Irvine)
Federated LearningMixture of ExpertsImageTabular
🎯 What it does: A Fed-POE algorithm is proposed, which achieves personalized training for online prediction and model fine-tuning by dynamically combining local models, global models, and historical models periodically saved by the server within the federated learning framework.
Personalized Steering of Large Language Models: Versatile Steering Vectors Through Bi-directional Preference Optimization
Yuanpu Cao (Pennsylvania State University), Jinghui Chen (Pennsylvania State University)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes an algorithm based on Bidirectional Preference Optimization (BiPO) to learn more effective 'steering vectors' within a single Transformer layer, allowing for precise control of large language model behavior during inference through shifts in the activation space.
Personalizing Reinforcement Learning from Human Feedback with Variational Preference Learning
Sriyash Poddar (Paul G Allen School of Computer Science and Engineering University of Washington), Natasha Jaques (Paul G Allen School of Computer Science and Engineering University of Washington)
Reinforcement Learning from Human FeedbackReinforcement LearningText
🎯 What it does: A personalized reinforcement learning method based on Variational Preference Learning (VPL) is proposed to address the issue that current human feedback reinforcement learning (RLHF) techniques cannot handle individual preference differences.
Pessimistic Backward Policy for GFlowNets
Hyosoon Jang (POSTECH), Sungsoo Ahn (POSTECH)
GenerationOptimizationReinforcement LearningTabularBenchmark
🎯 What it does: Proposes a pessimistic backward policy (PBP-GFN) to address the low sampling issue of GFlowNets when sampling high-reward objects;
pFedClub: Controllable Heterogeneous Model Aggregation for Personalized Federated Learning
Jiaqi Wang (Pennsylvania State University), Fenglong Ma (Pennsylvania State University)
Federated LearningKnowledge DistillationImage
🎯 What it does: This paper proposes pFedClub, which can aggregate heterogeneous models in federated learning and generate personalized, controllable-sized teacher models.
PGN: The RNN's New Successor is Effective for Long-Range Time Series Forecasting
Yuxin Jia (Beijing Jiaotong University), Huaiyu Wan (Beijing Jiaotong University)
Recurrent Neural NetworkTime Series
🎯 What it does: A new parallel gated network (PGN) and its time-based variant TPGN are proposed for long sequence time series forecasting.
Phased Consistency Models
Fu-Yun Wang (Chinese University of Hong Kong), Hongsheng Li (Chinese University of Hong Kong)
GenerationData SynthesisDiffusion modelImageVideoOrdinary Differential Equation
🎯 What it does: This paper proposes a Phase Consistency Model (PCM), which addresses the shortcomings of traditional Latent Consistency Models (LCM) in terms of consistency, controllability, and efficiency by dividing continuous probability flow ODE trajectories into several sub-trajectories and learning self-consistency on each sub-trajectory. It supports deterministic sampling, allows for stronger CFG usage, and improves negative prompt perception.
PhoCoLens: Photorealistic and Consistent Reconstruction in Lensless Imaging
Xin Cai (Chinese University of Hong Kong), Tianfan Xue (Chinese University of Hong Kong)
RestorationDiffusion modelImage
🎯 What it does: This paper proposes a two-stage framework—first using a spatially varying convolutional network to restore low-frequency consistency, and then using a dilated spatial recovery based on a pre-trained diffusion model, achieving a balance between realistic lighting and consistency in lensless optical cameras.
PhyloGen: Language Model-Enhanced Phylogenetic Inference via Graph Structure Generation
ChenRui Duan, Stan Z. Li (Westlake University)
Graph Neural NetworkLarge Language ModelAuto EncoderBiomedical Data
🎯 What it does: The PhyloGen method is proposed, which infers evolutionary trees without relying on evolutionary models or sequence alignment by jointly generating and optimizing tree topology and branch lengths using a pre-trained genomic language model.
PhyRecon: Physically Plausible Neural Scene Reconstruction
Junfeng Ni (Tsinghua University), Siyuan Huang (Peking University)
Neural Radiance FieldPoint CloudMeshPhysics Related
🎯 What it does: By combining differentiable rendering with differentiable physical simulation, we learn a neural implicit surface representation to achieve physically feasible multi-view 3D reconstruction.
Physical Consistency Bridges Heterogeneous Data in Molecular Multi-Task Learning
Yuxuan Ren (Microsoft Research AI for Science), Tie-Yan Liu (Microsoft Research AI for Science)
Drug DiscoveryDiffusion modelGraphPhysics Related
🎯 What it does: In molecular multi-task learning, by introducing a 'physical consistency' loss based on physical laws, the tasks of energy prediction and equilibrium structure prediction are connected, overcoming data heterogeneity and improving the accuracy of structure prediction.
Physically Compatible 3D Object Modeling from a Single Image
Minghao Guo (Massachusetts Institute of Technology), Wojciech Matusik (Massachusetts Institute of Technology)
GenerationOptimizationImageMeshPhysics Related
🎯 What it does: A framework is proposed for reconstructing physically compatible 3D objects from a single image, separating mechanical properties, external forces, and the object's static shape, and optimizing through static equilibrium hard constraints to obtain models that conform to physical behavior.
Physics-Constrained Comprehensive Optical Neural Networks
Yanbing Liu (Beijing University of Posts and Telecommunications), Wei Li (Beijing Institute of Space Mechanics and Electricity)
ClassificationImagePhysics Related
🎯 What it does: A learning framework for optical neural networks based on physical constraints is proposed, which significantly improves experimental classification accuracy by introducing measurable physical priors such as light source instability and exposure time mismatch into an error compensation network and designing a corresponding loss function.
Physics-informed Neural Networks for Functional Differential Equations: Cylindrical Approximation and Its Convergence Guarantees
Taiki Miyagawa (Independent Researcher), Takeru Yokota (Interdisciplinary Theoretical and Mathematical Sciences Program RIKEN Center for Quantum Computing RIKEN)
Physics RelatedOrdinary Differential Equation
🎯 What it does: A learning framework combining cylindrical approximation and PINN is proposed for solving functional differential equations.
Physics-Informed Regularization for Domain-Agnostic Dynamical System Modeling
Zijie Huang (University of California Los Angeles), Wei Wang (University of California Los Angeles)
Graph Neural NetworkTime SeriesPhysics RelatedOrdinary Differential Equation
🎯 What it does: This paper proposes a dynamic system modeling framework TREAT based on graph neural ODEs, achieving high-precision predictions for various dynamical systems by incorporating a Time-Reversal Symmetry (TRS) regularization term.
Physics-Informed Variational State-Space Gaussian Processes
Oliver Hamelijnck (University of Warwick), Theodoros Damoulas (University of Warwick)
Gaussian SplattingTime SeriesPhysics RelatedStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: A unified physical information state space Gaussian process (PHYSS-GP) is proposed, achieving linear time complexity in the time domain through variational inference and capable of handling spatiotemporal partial differential equations.
Physics-Regularized Multi-Modal Image Assimilation for Brain Tumor Localization
Michal Balcerak (University of Zurich), bjoern menze
SegmentationOptimizationImageBiomedical DataMagnetic Resonance ImagingPositron Emission TomographyPhysics Related
🎯 What it does: By combining physical models with multimodal medical imaging, a flexible constraint physical information network is constructed to learn the distribution of brain tumor cells and the elasticity of brain tissue, thereby improving the radiotherapy plan for glioblastoma.
Piecewise deterministic generative models
Andrea Bertazzi (Ecole Polytechnique), Alain Oliviero Durmus
GenerationData SynthesisDiffusion modelFlow-based ModelImage
🎯 What it does: A class of generative models based on Piecewise Deterministic Markov Processes (PDMP) is proposed, along with explicit analytical and training methods for three typical PDMPs (ZZP, BPS, RHMC) in their time-reversed processes.
Piecewise-Stationary Bandits with Knapsacks
Xilin Zhang (National University of Singapore), Wang Chi Cheung (National University of Singapore)
OptimizationReinforcement LearningTabular
🎯 What it does: An algorithm based on inventory reservation is designed to achieve near-optimal competitive ratios in the piecewise stationary Bandits with Knapsacks problem.
Pin-Tuning: Parameter-Efficient In-Context Tuning for Few-Shot Molecular Property Prediction
Liang Wang (Institute of Automation, Chinese Academy of Sciences)
Meta LearningDrug DiscoveryGraph Neural NetworkSupervised Fine-TuningGraph
🎯 What it does: Proposed Pin-Tuning, a parameter-efficient and context-aware fine-tuning method for few-shot molecular property prediction tasks;
Pipeline Parallelism with Controllable Memory
Penghui Qi (Sea AI Lab), Min Lin (Sea AI Lab)
OptimizationComputational EfficiencyHyperparameter SearchTransformerLarge Language ModelText
🎯 What it does: A framework is proposed to decompose pipeline parallel scheduling into reusable building blocks, and based on the lifecycle of these blocks, peak activation memory is derived, leading to the design of memory-efficient scheduling schemes (V-Min, V-Half, V-ZB) using a V-Shape structure.
PiSSA: Principal Singular Values and Singular Vectors Adaptation of Large Language Models
Fanxu Meng (Peking University), Muhan Zhang (Peking University)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes the PiSSA method, which initializes low-rank adapters using the main singular values/vectors after performing singular value decomposition on the weights of pre-trained models, while freezing the residual part, thereby achieving parameter-efficient fine-tuning and significantly accelerating convergence speed.
PIVOT-R: Primitive-Driven Waypoint-Aware World Model for Robotic Manipulation
Kaidong Zhang (Sun Yat-sen University), Xiaodan Liang (Peng Cheng Laboratory)
Robotic IntelligenceTransformerPrompt EngineeringVision Language ModelMultimodality
🎯 What it does: A primitive-driven keypoint-aware world model (PIVOT-R) is proposed for language-guided robotic operations.
Plan-on-Graph: Self-Correcting Adaptive Planning of Large Language Model on Knowledge Graphs
Liyi Chen (University of Science and Technology of China), Hui Xiong (Hong Kong University of Science and Technology)
TransformerLarge Language ModelPrompt EngineeringTextGraph
🎯 What it does: This paper proposes Plan-on-Graph (PoG), a knowledge graph-based self-correcting adaptive planning framework for LLMs, which combines task decomposition, adaptive exploration of relationships/entities, memory updates, and reflection mechanisms to achieve efficient reasoning for KGQA tasks.
Plant-and-Steal: Truthful Fair Allocations via Predictions
Ilan Reuven Cohen (Bar-Ilan University), Arsen Vasilyan (UC Berkeley)
Tabular
🎯 What it does: This paper proposes a learning-augmented truthful mechanism—Plant-and-Steal framework—for approximately maximizing the minimum share (MMS) allocation among two agents (and generally n≥2 agents), utilizing predictions of agents' rankings of items to achieve both truthfulness and approximate fairness.
PLIP: Language-Image Pre-training for Person Representation Learning
Jialong Zuo (Huazhong University of Science and Technology), Jingdong Wang (Baidu Inc.)
RecognitionRetrievalRepresentation LearningTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: The PLIP framework is proposed, utilizing three pre-training tasks (text-guided image coloring, image-guided attribute prediction, identity-based cross-modal contrast) for language-image cross-modal learning, focusing on character attributes and identity.
Point-PRC: A Prompt Learning Based Regulation Framework for Generalizable Point Cloud Analysis
Hongyu Sun (Renmin University of China), Jianfei Cai (Monash University)
ClassificationRecognitionDomain AdaptationPrompt EngineeringContrastive LearningPoint Cloud
🎯 What it does: This study investigates prompt learning for large-scale 3D models, enhancing downstream point cloud recognition performance while maintaining domain generalization capabilities, and proposes the Point-PRC adjustment framework.
PointAD: Comprehending 3D Anomalies from Points and Pixels for Zero-shot 3D Anomaly Detection
Qihang Zhou (Zhejiang University), Jiming Chen (Zhejiang University)
Anomaly DetectionPrompt EngineeringVision Language ModelMultimodalityPoint Cloud
🎯 What it does: The PointAD framework is proposed, achieving zero-shot 3D anomaly detection and pluggable multimodal detection, directly utilizing the general features of CLIP for anomaly segmentation of unknown point clouds.
PointMamba: A Simple State Space Model for Point Cloud Analysis
Dingkang Liang (Huazhong University of Science and Technology), Xiang Bai (Huazhong University of Science and Technology)
ClassificationSegmentationConvolutional Neural NetworkPoint Cloud
🎯 What it does: A state space model PointMamba based on Mamba is designed, which serializes point clouds using space-filling curves and completes tasks such as point cloud classification and segmentation through a pure Mamba encoder.
Poisson Variational Autoencoder
Hadi Vafaii (University of California Berkeley), Jacob L. Yates (University of California Berkeley)
GenerationRepresentation LearningConvolutional Neural NetworkAuto EncoderImage
🎯 What it does: A Poisson distribution-based variational autoencoder (P-VAE) is proposed, replacing the continuous latent variables in traditional VAEs with discrete neural spike counts, achieving more biologically relevant perceptual inference.
Policy Aggregation
Parand A. Alamdari (University of Toronto), Ariel D. Procaccia (Harvard University)
OptimizationReinforcement Learning
🎯 What it does: This paper proposes that in multi-agent Markov decision processes, the reward functions of each agent are viewed as 'preferences' in social choice problems. By using the occupancy polytope of state-action pairs to aggregate all possible strategies, a interpretable and fair collective strategy is obtained.
Policy Improvement using Language Feedback Models
Victor Zhong (University of Waterloo), Marc-Alexandre Côté (Microsoft Research)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes and trains Language Feedback Models (LFMs) to identify beneficial behaviors in instruction-following tasks from feedback obtained from pre-trained large language models (LLMs), thereby enhancing strategies through imitation learning; it also demonstrates that LFMs can provide interpretable feedback.
Policy Learning from Tutorial Books via Understanding, Rehearsing and Introspecting
Xiong-Hui Chen (Nanjing University), Jun Wang (University College London)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextRetrieval-Augmented Generation
🎯 What it does: Understanding and extracting codified knowledge from tutorial books through large language models (LLM), generating virtual decision trajectories, and introspecting these trajectories using conservative Q-learning to ultimately obtain a policy network that can be executed without real interaction.
Policy Mirror Descent with Lookahead
Kimon Protopapas (ETH Zurich), Anas Barakat (Singapore University of Technology and Design)
OptimizationReinforcement Learning
🎯 What it does: A new Policy Mirror Descent (PMD) algorithm, h-PMD, is proposed, which incorporates multi-step lookahead (h steps) into the traditional PMD to improve policy updates, thereby enhancing convergence speed and reducing sampling complexity.
Policy Optimization for Robust Average Reward MDPs
Zhongchang Sun (University at Buffalo), Shaofeng Zou (Arizona State University)
OptimizationReinforcement LearningTabular
🎯 What it does: A robust average cost MDP policy gradient algorithm is proposed - Robust Policy Mirror Descent, and its convergence properties are provided;
Policy-shaped prediction: avoiding distractions in model-based reinforcement learning
Miles Richard Hutson (Stanford University), Nick Haber (Stanford University)
Reinforcement LearningWorld ModelVideo
🎯 What it does: A method called Policy-Shaped Prediction (PSP) is proposed, which utilizes policy gradients, a pre-trained segmentation model, and adversarial learning to focus the capacity of the world model, thereby suppressing sensitivity to irrelevant background interference and enhancing the robustness of model-based reinforcement learning.
Polyhedral Complex Derivation from Piecewise Trilinear Networks
Jin-Hwa Kim (NAVER AI Lab & SNU AIIS)
Mesh
🎯 What it does: Theoretical analysis of neural implicit surface networks using trilinear interpolation (such as HashGrid) is conducted, and an analytical method for accurately extracting polygonal meshes from piecewise trilinear networks under the Eikonal constraint is proposed.
Polynomial-Time Computation of Exact $\Phi$-Equilibria in Polyhedral Games
Gabriele Farina (Massachusetts Institute of Technology), Charilaos Pipis (Massachusetts Institute of Technology)
Optimization
🎯 What it does: A general polynomial-time algorithm is proposed for calculating exact Φ-equilibria in polyhedral games, including the first precise computation of Linear Deviation Cooperative Equilibrium (LCE) in generalized extensive-form games.
Poseidon: Efficient Foundation Models for PDEs
Maximilian Herde (ETH Zurich), Siddhartha Mishra (ETH Zurich)
OptimizationComputational EfficiencyData-Centric LearningTransformerSupervised Fine-TuningTime SeriesBenchmarkPhysics Related
🎯 What it does: A PDE-based model named POSEIDON is proposed and trained, capable of learning and generating operators for various physical processes, supporting continuous time evaluation and multi-scale feature extraction.
Position Coupling: Improving Length Generalization of Arithmetic Transformers Using Task Structure
Hanseul Cho (Korea Advanced Institute of Science and Technology), Chulhee Yun (Korea Advanced Institute of Science and Technology)
TransformerSequential
🎯 What it does: This paper proposes a technique called Position Coupling, which directly embeds task structures into the Transformer using position encoding to enhance its performance in length generalization, particularly for algorithmic problems such as integer addition, multiplication, and two-dimensional tasks.
Post-Hoc Reversal: Are We Selecting Models Prematurely?
Rishabh Ranjan (Stanford University), Zachary Chase Lipton (Carnegie Mellon University)
MultimodalityTabular
🎯 What it does: The effects of post-processing transformations such as temperature scaling, ensembling, and stochastic weight averaging on model performance were studied, and the phenomenon of post-processing reversal was discovered for the first time.
Posture-Informed Muscular Force Learning for Robust Hand Pressure Estimation
Kyungjin Seo (Korea Advanced Institute of Science and Technology), Sang Ho Yoon (Korea University)
Pose EstimationConvolutional Neural NetworkMultimodality
🎯 What it does: This paper proposes the PiMForce framework, which achieves real-time pressure estimation of the entire hand during various hand-object interactions by integrating 3D hand posture information with surface electromyography (sEMG) signals from the forearm.
PowerPM: Foundation Model for Power Systems
Shihao Tu (Zhejiang University), Yang Yang (Zhejiang University)
ClassificationAnomaly DetectionGraph Neural NetworkTransformerContrastive LearningTime Series
🎯 What it does: The PowerPM foundational model is proposed for modeling and downstream tasks of electric power time series data (ETS).
PPLNs: Parametric Piecewise Linear Networks for Event-Based Temporal Modeling and Beyond
Chen Song (University of Texas at Austin), Qixing Huang (University of Texas at Austin)
Pose EstimationAutonomous DrivingOptimizationSpiking Neural NetworkTime SeriesSequential
🎯 What it does: This paper proposes a time neural network PPLN based on learnable piecewise linear functions to handle time series data generated by event cameras.
Practical $0.385$-Approximation for Submodular Maximization Subject to a Cardinality Constraint
Murad Tukan (DataHeroes), Moran Feldman (University of Haifa)
Recommendation SystemOptimizationImageTabular
🎯 What it does: A novel combinatorial algorithm is proposed to maximize non-monotone submodular functions under cardinality constraints, balancing theoretical approximation ratios and practical query complexities.
Practical Bayesian Algorithm Execution via Posterior Sampling
Chu Xin Cheng (California Institute of Technology), Yisong Yue (California Institute of Technology)
OptimizationTabular
🎯 What it does: This paper proposes and implements a Bayesian algorithm execution framework based on posterior sampling, called PS-BAX, for efficiently estimating the outputs of base algorithms on expensive black-box functions.
Practical Shuffle Coding
Julius Kunze (University College London), James Townsend (University of Amsterdam)
CompressionGraph
🎯 What it does: A general-purpose entropy coding method for unordered objects is proposed, combining incomplete shuffling coding and autoregressive shuffling coding, capable of compressing large unordered data structures (such as multisets and graphs) in one go and achieving high-speed compression.
Pre-trained Large Language Models Use Fourier Features to Compute Addition
Tianyi Zhou (University of Southern California), Robin Jia (University of Southern California)
TransformerLarge Language ModelText
🎯 What it does: Analyzes how pre-trained large language models utilize Fourier features to perform integer addition tasks, revealing the mechanisms by which the model approximates and performs modulo operations at different levels;
Pre-Trained Multi-Goal Transformers with Prompt Optimization for Efficient Online Adaptation
Haoqi Yuan (Peking University), Zongqing Lu (Peking University)
OptimizationExplainability and InterpretabilityTransformerReinforcement LearningPrompt EngineeringSequential
🎯 What it does: Pre-trained Transformer multi-objective policy, optimized through online prompts (goal sequence) for efficient adaptation to long-horizon tasks; no additional online RL fine-tuning required.
Pre-trained Text-to-Image Diffusion Models Are Versatile Representation Learners for Control
Gunshi Gupta (University of Oxford), Tim G. J. Rudner (New York University)
Representation LearningRobotic IntelligenceConvolutional Neural NetworkRecurrent Neural NetworkReinforcement LearningDiffusion modelImageText
🎯 What it does: The research utilizes a pre-trained text-image diffusion model (Stable Diffusion) to extract visual-language representations for policy learning in robot control tasks.
Pre-training Differentially Private Models with Limited Public Data
Zhiqi Bu (Amazon), George Karypis (Amazon)
OptimizationSafty and PrivacyTransformerSupervised Fine-TuningImage
🎯 What it does: This paper proposes a continuous pre-training method that utilizes a small amount of public data, maintaining differential privacy while having almost no impact on the pre-training effectiveness of large models, addressing the performance degradation issue of traditional DP pre-training.
Precipitation Downscaling with Spatiotemporal Video Diffusion
Prakhar Srivastava (University of California), Stephan Mandt (University of California)
GenerationData SynthesisDiffusion modelVideoTime Series
🎯 What it does: Using video diffusion models to perform spatiotemporal downscaling of low-resolution precipitation sequences, refining them into high-resolution sequences.
Precise asymptotics of reweighted least-squares algorithms for linear diagonal networks
Chiraag Kaushik (Georgia Institute of Technology), Vidya Muthukumar (Georgia Institute of Technology)
OptimizationTabular
🎯 What it does: This paper provides a unified asymptotic analysis for a class of algorithms, including the Iteratively Reweighted Least Squares (IRLS) algorithm, the recently proposed lin-RFM algorithm, and the alternating minimization algorithm on linear diagonal neural networks, aimed at recovering unknown signals from linear measurements.
Predicting Future Actions of Reinforcement Learning Agents
Stephen Chung (University of Cambridge), David Krueger (Mila)
Convolutional Neural NetworkRecurrent Neural NetworkTransformerReinforcement LearningWorld ModelSequential
🎯 What it does: This paper studies the prediction of future actions and events of deployed reinforcement learning agents, systematically comparing explicit planning, implicit planning, and non-planning agents, and proposes two prediction methods based on internal states and simulations.
Predicting Ground State Properties: Constant Sample Complexity and Deep Learning Algorithms
Marc Wanner (Chalmers University of Technology and University of Gothenburg), Alexandru Gheorghiu (Chalmers University of Technology and University of Gothenburg)
TabularPhysics Related
🎯 What it does: Two machine learning algorithms are proposed to predict the ground state properties of quantum many-body systems with constant sample complexity.
Predicting Label Distribution from Ternary Labels
Yunan Lu (Nanjing University of Science and Technology), Xiuyi Jia (Nanjing University of Science and Technology)
ClassificationData-Centric LearningImage
🎯 What it does: This paper proposes a method for predicting label distributions based on three-value labels (-1, 0, 1) and applies it to the Label Enhancement framework.
Predicting the Performance of Foundation Models via Agreement-on-the-Line
Rahul Saxena (Carnegie Mellon University), Aditi Raghunathan (Carnegie Mellon University)
ClassificationDomain AdaptationTransformerSupervised Fine-TuningImageText
🎯 What it does: This study investigates how to predict the OOD performance of models using the agreement-on-the-line (AGL) method on unlabeled data with a foundation model, and proposes techniques for constructing diverse ensembles.
Prediction with Action: Visual Policy Learning via Joint Denoising Process
Yanjiang Guo (Tsinghua University), Jianyu Chen
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerDiffusion modelImageVideoMultimodality
🎯 What it does: This paper proposes PAD (Prediction with Action Diffuser), a visual strategy learning framework that simultaneously predicts future images and robot actions during joint denoising.
Prediction-Powered Ranking of Large Language Models
Ivi Chatzi (Max Planck Institute for Software Systems), Manuel Gomez Rodriguez (Max Planck Institute for Software Systems)
Recommendation SystemTransformerLarge Language ModelText
🎯 What it does: A statistical framework based on prediction-driven inference is proposed, which constructs a confidence set (rank-set) for model ranking using a small amount of human comparisons and a large amount of powerful LLM comparison data, thereby quantifying ranking uncertainty.
Predictive Attractor Models
Ramy Mounir (University of South Florida), Sudeep Sarkar (University of South Florida)
GenerationData SynthesisRepresentation LearningSequential
🎯 What it does: This paper proposes Predictive Attractor Models (PAM) — an online sparse distributed representation sequence memory architecture that can learn after a single observation, avoid catastrophic forgetting, and randomly generate future events in various possible scenarios, while being robust to noise.