NeurIPS 2025 Papers — Page 36
Conference on Neural Information Processing Systems · 5275 papers
Prediction-Powered Causal Inferences
Riccardo Cadei (Institute of Science and Technology Austria), Francesco Locatello (Massachusetts Institute of Technology)
Video
🎯 What it does: A prediction-driven causal inference (PPCI) framework is proposed and validated, utilizing pre-trained models to predict outcomes in unlabeled experiments, thereby achieving causal effect estimation without human annotation.
Prediction-Powered Semi-Supervised Learning with Online Power Tuning
Noa Shoham (Technion Institute of Technology), Yaniv Romano (Technion Institute of Technology)
ClassificationOptimizationKnowledge DistillationConvolutional Neural NetworkSupervised Fine-TuningReinforcement LearningImageTabular
🎯 What it does: A semi-supervised learning framework called PP-SSL based on Prediction-Powered Inference (PPI) is proposed, which utilizes a teacher model to generate pseudo-labels and construct unbiased gradient estimates, and adaptively adjusts the pseudo-label weight λ through online learning.
Predictive Coding Enhances Meta-RL To Achieve Interpretable Bayes-Optimal Belief Representation Under Partial Observability
Po-Chen Kuo (University of Washington), Edgar Y. Walker (University of Washington)
Meta LearningRecurrent Neural NetworkReinforcement LearningAuto EncoderSequential
🎯 What it does: In partially observable environments, a self-supervised prediction module (predictive coding) is embedded in a meta reinforcement learning framework to learn interpretable and near-Bayesian optimal belief representations and obtain optimal policies.
Predictive Preference Learning from Human Interventions
Haoyuan Cai (University of California), Bolei Zhou (University of California)
Autonomous DrivingRobotic IntelligenceReinforcement LearningTabular
🎯 What it does: This paper proposes an interactive imitation learning framework called Predictive Preference Learning from Human Interventions (PPL), which accelerates learning and reduces the number of human interventions by utilizing predicted future trajectories and preference labels generated from human interventions.
Preference Distillation via Value based Reinforcement Learning
Minchan Kwon (Korea Advanced Institute of Science and Technology), Junmo Kim (Korea Advanced Institute of Science and Technology)
Knowledge DistillationReinforcement Learning
🎯 What it does: This paper proposes Teacher Value-based Knowledge Distillation (TVKD), which utilizes the value function of the teacher model to perform preference distillation on the student model within the DPO framework.
Preference Learning with Lie Detectors can Induce Honesty or Evasion
Chris Cundy (FAR AI), Adam Gleave (FAR AI)
Recommendation SystemOptimizationTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper studies the process of introducing a lie detector in preference learning (post-training of LLMs), evaluating whether it can effectively encourage the model to be honest or lead the model to evade detection and continue lying.
Preference Learning with Response Time: Robust Losses and Guarantees
Ayush Sawarni (Stanford University), Vasilis Syrgkanis (Stanford University)
Recommendation SystemDiffusion modelTextMultimodality
🎯 What it does: This paper proposes a method that utilizes Neyman orthogonal loss to jointly use binary preference data and response time information for reward model learning.
Preference Optimization by Estimating the Ratio of the Data Distribution
Yeongmin Kim (Korea Advanced Institute of Science and Technology), Il-chul Moon
Recommendation SystemOptimizationLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: A preference optimization framework BPO based on Bregman divergence is proposed, treating DPO as a proportional matching problem, achieving optimality and simplicity without the need for a reward model.
Preference Optimization on Pareto Sets: On a Theory of Multi-Objective Optimization
Abhishek Roy (Texas A&M University), Yian Ma
Optimization
🎯 What it does: This paper studies how to find optimal solutions on the Pareto set given a preference function in multi-objective optimization, and proposes a corresponding theoretical framework and algorithm.
Preference-Based Dynamic Ranking Structure Recognition
Nan Lu (Chinese Academy of Sciences), Xinyu Tian
Recommendation SystemOptimizationTime Series
🎯 What it does: A dynamic framework for simultaneous grouping and ranking is proposed, capable of identifying the group structure of items and its changes in time series.
Preference-based Reinforcement Learning beyond Pairwise Comparisons: Benefits of Multiple Options
Joongkyu Lee (Seoul National University), Min-hwan Oh (Seoul National University)
Reinforcement LearningTabular
🎯 What it does: Proposes the online PbRL algorithm M-AUPO, which utilizes multiple ranking feedback to enhance sample efficiency.
Preference-driven Knowledge Distillation for Few-shot Node Classification
Xing Wei (Tongji University), Wei Ye (Tongji University)
Knowledge DistillationGraph Neural NetworkLarge Language ModelSupervised Fine-TuningReinforcement LearningTextGraph
🎯 What it does: A Preference-driven Knowledge Distillation (PKD) framework has been developed, combining large language models (LLMs) with various graph neural networks (GNNs) for few-shot node classification in Text Attribute Graphs (TAGs).
Preference-Driven Multi-Objective Combinatorial Optimization with Conditional Computation
Mingfeng Fan (National University of Singapore), Guillaume Adrien Sartoretti (National University of Singapore)
OptimizationTransformerReinforcement LearningMixture of ExpertsContrastive LearningTabularBenchmark
🎯 What it does: Proposed and implemented the POCCO framework, which addresses multi-objective combinatorial optimization problems using conditional computation blocks and preference-driven training.
Preference-Guided Diffusion for Multi-Objective Offline Optimization
Yashas Annadani (TU Munich), Barbara E Engelhardt
OptimizationNeural Architecture SearchDiffusion modelTabular
🎯 What it does: A preference-guided diffusion model is proposed for generating Pareto fronts in offline multi-objective optimization (MOO).
PrefixKV: Adaptive Prefix KV Cache is What Vision Instruction-Following Models Need for Efficient Generation
Ao Wang (Tsinghua University), Guiguang Ding (Tsinghua University)
GenerationCompressionComputational EfficiencyTransformerVision Language ModelTextMultimodality
🎯 What it does: Proposes PrefixKV, which achieves efficient generation of visual language models through adaptive prefix KV cache compression.
PreFM: Online Audio-Visual Event Parsing via Predictive Future Modeling
Xiao Yu (Beijing Jiaotong University), Yunchao Wei (Beijing Jiaotong University)
RecognitionComputational EfficiencyKnowledge DistillationVideoMultimodalityAudio
🎯 What it does: This paper proposes an online audio-video event parsing framework called PreFM, which can parse audio, visual, and audio-video events frame by frame in real-time video streams, balancing accuracy and real-time performance.
PRESCRIBE: Predicting Single-Cell Responses with Bayesian Estimation
Jiabei Cheng (Shanghai Jiao Tong University), Jun Xia (Hong Kong University of Science and Technology)
Drug DiscoveryFlow-based ModelBiomedical Data
🎯 What it does: The PRESCRIBE framework is proposed for predicting single-cell gene interference and providing reliable uncertainty estimates.
Preserving LLM Capabilities through Calibration Data Curation: From Analysis to Optimization
Bowei He (City University of Hong Kong), Chen Ma (City University of Hong Kong)
CompressionOptimizationTransformerLarge Language ModelText
🎯 What it does: This paper systematically evaluates the impact of calibration data on various capabilities of LLMs (such as language modeling, common sense reasoning, mathematical reasoning, code generation, and multilingual understanding) in post-training pruning and quantization compression, and proposes a three-stage calibration data planning framework called COLA to maximize the representativeness and diversity of the activation space to enhance the capability retention of the compressed model.
Preserving Task-Relevant Information Under Linear Concept Removal
Floris Holstege (University of Amsterdam), Bram Wouters (New York University)
OptimizationSafty and PrivacyImageText
🎯 What it does: This paper proposes SPLINCE, a linear concept removal method that removes sensitive information while maintaining the covariance with the main task labels.
PRESTO: Preimage-Informed Instruction Optimization for Prompting Black-Box LLMs
Jaewon Chu (Korea University), Hyunwoo J. Kim (KAIST)
OptimizationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: We propose PRESTO, a framework that accelerates black-box LLM instruction optimization using a soft prompt many-to-one mapping (preimage) structure;
Pretraining a Shared Q-Network for Data-Efficient Offline Reinforcement Learning
Jongchan Park (Hyundai Motor Company), Donghwan Lee (Korea Advanced Institute of Science and Technology)
Reinforcement Learning
🎯 What it does: This paper proposes a simple pre-training framework that first pre-trains a shared Q-network's feature extractor using a transition prediction task in offline reinforcement learning, and then uses it as the initialization for the Q-network, directly applied to the training of any offline RL algorithm;
Preventing Shortcuts in Adapter Training via Providing the Shortcuts
Anujraaj Goyal (Snap Inc), Kuan-Chieh Wang (Snap Inc)
GenerationPose EstimationDiffusion modelImage
🎯 What it does: Proposes the Shortcut-Rerouted Adapter Training scheme, which explicitly provides shortcut paths through auxiliary modules such as LoRA and ControlNet during training, suppressing the adapter from learning irrelevant factors like pose, expression, and lighting, thereby achieving more accurate personalized generation of faces and full bodies.
Price of Parsimony: Complexity of Fourier Sparsity Testing
Arijit Ghosh (Indian Statistical Institute), Manmatha Roy (Indian Statistical Institute)
🎯 What it does: This paper proposes an algorithm for tolerance detection of Fourier sparsity for real-valued Boolean functions, and provides an almost optimal query complexity;
PRIMT: Preference-based Reinforcement Learning with Multimodal Feedback and Trajectory Synthesis from Foundation Models
Ruiqi Wang (Purdue University), Byung-Cheol Min (Indiana University)
Robotic IntelligenceLarge Language ModelReinforcement LearningVision Language ModelMultimodality
🎯 What it does: A preference reinforcement learning framework PRIMT based on a fund model is proposed, utilizing multimodal synthetic feedback and trajectory synthesis to achieve zero human annotation in robot learning.
Principled Data Augmentation for Learning to Solve Quadratic Programming Problems
Chendi Qian (RWTH Aachen University), Christopher Morris (RWTH Aachen University)
OptimizationGraph Neural NetworkContrastive Learning
🎯 What it does: This paper proposes a principled data augmentation framework based on KKT system affine transformations, aimed at generating linear and quadratic programming instances that maintain optimal solutions, and implements supervised and contrastive learning in the context of learning optimization message passing neural networks (MPNN);
Principled Fine-tuning of LLMs from User-Edits: A Medley of Preference, Supervision, and Reward
Dipendra Misra (Databricks Mosaic Research), Ge Gao (Google DeepMind)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Utilizing user feedback data from deployment logs where they edit LLM responses, a complete learning protocol for offline and online fine-tuning is proposed based on these edits;
Principled Long-Tailed Generative Modeling via Diffusion Models
Pranoy Das (Purdue University), Vijay Gupta (Purdue University)
GenerationData SynthesisOptimizationDiffusion modelImageStochastic Differential Equation
🎯 What it does: This study investigates the learning methods of diffusion generative models under long-tail distributions, proposing to construct a multi-player Nash game through deep mutual learning to balance the generation quality across different categories.
Principled Model Routing for Unknown Mixtures of Source Domains
Christoph Dann (Google Research), Mehryar Mohri (Google Research)
Domain AdaptationOptimizationLarge Language ModelMixture of ExpertsTextBenchmark
🎯 What it does: This paper proposes a model routing algorithm aimed at unknown task mixtures, transforming the routing problem into a multi-source domain adaptation (MSA) framework, and achieving robust routing by minimizing adversarial worst-case regret.
Prior Forgetting and In-Context Overfitting
Sungyoon Lee (Hanyang University)
TransformerTabularOrdinary Differential Equation
🎯 What it does: The study investigates how the two modes of In-Context Learning (ICL) - task recognition and task learning - emerge, disappear, and dynamically evolve during large-scale pre-training.
Prior-Guided Diffusion Planning for Offline Reinforcement Learning
Donghyeon Ki (Korea University), Byung-Jun Lee (Korea University)
Recurrent Neural NetworkReinforcement LearningDiffusion modelTabularBenchmark
🎯 What it does: Proposes Prior Guidance (PG), a learnable prior distribution for training diffusion planners in offline reinforcement learning, allowing for the direct generation of high-value trajectories during inference, thus avoiding multi-trajectory sampling and reward optimization.
Prior-Guided Flow Matching for Target-Aware Molecule Design with Learnable Atom Number
Jingyuan Zhou (Shanghai Jiao Tong University), Lei Xu (Shanghai Jiao Tong University)
Drug DiscoveryGraph Neural NetworkFlow-based ModelGraphOrdinary Differential Equation
🎯 What it does: This paper proposes a target-aware molecular generation framework PAFlow based on flow matching, which can directly generate high-affinity 3D small molecules under the condition of a given protein pocket;
Prioritizing Perception-Guided Self-Supervision: A New Paradigm for Causal Modeling in End-to-End Autonomous Driving
Yi Huang (Chinese University of Hong Kong), Hongbo Zhang (Huawei Noah's Ark Lab)
Autonomous DrivingTransformerVideo
🎯 What it does: A perception-guided self-supervised (PGS) training paradigm is proposed, utilizing outputs from the perception module (lane centerlines, future trajectories of dynamic objects) as positive and negative supervision signals, significantly reducing causal confusion.
Prismatic Synthesis: Gradient-based Data Diversification Boosts Generalization in LLM Reasoning
Jaehun Jung (NVIDIA Research), Yejin Choi (NVIDIA Research)
Data SynthesisOptimizationTransformerLarge Language ModelTextBenchmark
🎯 What it does: This study investigates the importance of data diversity in LLM inference tasks, proposing a diversity metric G-Vendi based on gradient entropy. Based on this, it designs Prismatic Synthesis to generate diversified synthetic data, constructing the PrismMath and PrismNLI datasets, achieving state-of-the-art results on multiple out-of-bag benchmarks.
Privacy amplification by random allocation
Moshe Shenfeld (Hebrew University of Jerusalem), Vitaly Feldman (Apple)
Safty and Privacy
🎯 What it does: This paper studies the privacy amplification properties of the random allocation sampling scheme (k-out-of-t) in differential privacy, providing theoretically computable privacy upper bounds and numerical estimation algorithms.
Privacy Reasoning in Ambiguous Contexts
Ren Yi (Google Research), Marco Gruteser (Google Research)
Safty and PrivacyTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper studies the performance of large language models in privacy decision-making, constructing the PrivacyLens+ and ConfAIde+ datasets that include both positive and negative samples, and proposes the Camber framework to enhance privacy judgment through automated context disambiguation.
Private Continual Counting of Unbounded Streams
Ben Jacobsen (University of Wisconsin - Madison), Kassem Fawaz (University of Wisconsin - Madison)
Safty and PrivacyComputational Efficiency
🎯 What it does: A new algorithm for achieving differentially private continuous counting in unbounded streams with unknown input size is proposed.
Private Evolution Converges
Tomás González, Aaditya Ramdas
OptimizationSafty and PrivacyDiffusion modelImage
🎯 What it does: This paper proposes a new theoretical framework and conducts a rigorous analysis of the convergence of the Private Evolution (PE) algorithm, providing upper bounds for the 1-Wasserstein distance convergence in Euclidean space and general Banach spaces.
Private Geometric Median in Nearly-Linear Time
Syamantak Kumar (University of Texas at Austin), Chutong Yang (University of Texas at Austin)
Safty and PrivacyComputational EfficiencyTabular
🎯 What it does: This work proposes a near-linear time differential privacy geometric median estimation algorithm that can achieve α-fold error approximation with the same sample size as the information-theoretic optimal under the premise of (ϵ,δ)-DP.
Private Hyperparameter Tuning with Ex-Post Guarantee
Badih Ghazi (Google Research), Chiyuan Zhang (Google Research)
Safty and PrivacyHyperparameter SearchTabularTime Series
🎯 What it does: This paper proposes a hyperparameter tuning algorithm under ex-post DP and RDP that can handle arbitrary benchmark mechanisms, along with corresponding privacy guarantees.
Private Online Learning against an Adaptive Adversary: Realizable and Agnostic Settings
Bo Li (Guangzhou HKUST Fok Ying Tung Research Institute), Peng Ye (Guangzhou HKUST Fok Ying Tung Research Institute)
OptimizationSafty and PrivacyReinforcement Learning
🎯 What it does: This paper proposes differential privacy online learning algorithms for both realizable and agnostic scenarios, achieving an error upper bound of O(d log T) and a reward lower bound of O~(d√T) under adaptive adversaries.
Private Set Union with Multiple Contributions
Travis Dick, Ananda Theertha Suresh (Google Research)
Safty and Privacy
🎯 What it does: The research achieves differentially private set union under the condition that each user can contribute at most k items, proposing a new metric and providing feasible algorithms and performance upper bounds.
Private Statistical Estimation via Truncation
Manolis Zampetakis (Yale University), Felix Zhou (Yale University)
OptimizationSafty and PrivacyTabular
🎯 What it does: A new framework for differential privacy statistical estimation is proposed, which achieves this by truncating data and can handle unbounded high-dimensional exponential family distributions, along with providing efficient algorithms; under this framework, private estimates of high-dimensional exponential families, Gaussian means, and covariances are realized.
Private Training Large-scale Models with Efficient DP-SGD
Liangyu Wang (King Abdullah University of Science and Technology), Di Wang (King Abdullah University of Science and Technology)
Safty and PrivacyComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: A cache-friendly per-layer DP-SGD algorithm called FlashDP is designed to significantly improve memory utilization and throughput in training large language models while ensuring differential privacy.
Private Zeroth-Order Optimization with Public Data
Xuchen Gong (University of Chicago), Tian Li (University of Chicago)
OptimizationSafty and PrivacyConvolutional Neural NetworkRecurrent Neural NetworkTransformerGaussian SplattingImageText
🎯 What it does: This paper proposes a public data-assisted zero-order differential privacy optimization method (PAZO) to improve the gradient estimation and convergence speed of private training.
Pro3D-Editor: A Progressive Framework for Consistent and Precise 3D Editing
Yang Zheng (University of Science and Technology of China), Zhendong Mao (Institute of Artificial Intelligence, Hefei Comprehensive National Science Center)
GenerationDiffusion modelGaussian SplattingPoint Cloud
🎯 What it does: A text-guided 3D editing framework called Pro3D-Editor based on progressive views is proposed, which enables precise and consistent editing of local areas of 3D objects.
Probabilistic Reasoning with LLMs for Privacy Risk Estimation
Jonathan Zheng (Georgia Institute of Technology), Wei Xu (Georgia Institute of Technology)
Safty and PrivacyTransformerLarge Language ModelText
🎯 What it does: A method is proposed to assess privacy risks in text through probabilistic reasoning, quantified as k-anonymity value, which represents the number of people globally with the same personal attributes as those in the text.
Probabilistic Stability Guarantees for Feature Attributions
Helen Jin (University of Pennsylvania), Eric Wong (University of Pennsylvania)
Explainability and InterpretabilityTransformerImageText
🎯 What it does: This paper introduces the concept of Soft Stability and presents a model-agnostic, sample-efficient stability certification algorithm (SCA) that can provide a probabilistically rigorous assessment of the robustness of feature importance explanations. It also demonstrates through Boolean function analysis that mild smoothing (MuS) can enhance stability without significantly sacrificing accuracy.
Probabilistic Token Alignment for Large Language Model Fusion
Runjia Zeng (Rochester Institute of Technology), Dongfang Liu (Rochester Institute of Technology)
TransformerLarge Language ModelText
🎯 What it does: This paper proposes a probability-based token alignment method (PTA-LLM), which achieves soft mapping between different pre-trained LLMs through optimal transport within a knowledge fusion framework, thereby integrating knowledge from multiple models without the need to retrain them.
Probably Approximately Precision and Recall Learning
Lee Cohen (Stanford University), Han Shao (University of Maryland)
OptimizationSupervised Fine-Tuning
🎯 What it does: A PAC framework for learning set prediction under the condition of only observing positive samples is proposed, focusing on optimizing precision and recall.
Probing Equivariance and Symmetry Breaking in Convolutional Networks
Sharvaree Vadgama (University of Amsterdam), Erik J Bekkers
SegmentationGenerationPose EstimationConvolutional Neural NetworkPoint Cloud
🎯 What it does: This study investigates the impact of equivalence and symmetry breaking in convolutional networks on point cloud tasks, proposing a unified Rapidash framework and conducting theoretical and experimental explorations.
Probing Hidden Knowledge Holes in Unlearned LLMs
Myeongseob Ko, Ruoxi Jia (Virginia Tech)
TransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: This paper proposes a three-step knowledge gap detection framework to assess the knowledge loss of large language models that is not captured by standard benchmarks after executing machine forgetting.
Probing Neural Combinatorial Optimization Models
Zhiqin Zhang (Singapore Management University), Hoong Chuin Lau (Singapore Management University)
OptimizationTransformerTabular
🎯 What it does: This paper is the first to apply probing techniques to Neural Combinatorial Optimization (NCO) models, systematically designing low-order and high-order probing tasks, and proposing a new Coefficient Significance Probing (CS-Probing) method;
Problem-Parameter-Free Decentralized Bilevel Optimization
Zhiwei Zhai (Chinese University of Hong Kong), Ying-Jun Angela Zhang (Chinese University of Hong Kong)
OptimizationMeta LearningImage
🎯 What it does: A completely parameter-free, single-loop decentralized bi-level optimization algorithm AdaSDBO is proposed, suitable for multi-agent collaborative solving of non-convex-strongly convex bi-level problems.
Process vs. Outcome Reward: Which is Better for Agentic RAG Reinforcement Learning
wenlin zhang, Xiangyu Zhao (City University of Hong Kong)
TransformerReinforcement LearningAgentic AITextRetrieval-Augmented Generation
🎯 What it does: A process-supervised reinforcement learning-based agentic RAG framework called ReasonRAG is proposed, enabling LLMs to autonomously perform retrieval, generate queries, extract evidence, and answer questions.
Procurement Auctions with Predictions: Improved Frugality for Facility Location
Eric Balkanski (Columbia University), Xizhi Tan (Stanford University)
Optimization
🎯 What it does: In the strategic location problem of capacity-free facilities, this paper designs a sincere procurement auction and improves cost efficiency through theoretical analysis.
ProDAG: Projected Variational Inference for Directed Acyclic Graphs
Ryan Thompson (University of Technology Sydney), Robert Kohn (University of New South Wales)
Graph Neural NetworkBiomedical Data
🎯 What it does: A new Bayesian variational inference framework called ProDAG is proposed for learning directed acyclic graphs (DAGs) from data and quantifying the uncertainty of the graphs.
Product Distribution Learning with Imperfect Advice
Arnab Bhattacharyya (University of Warwick), Themis Gouleakis (Nanyang Technological University)
🎯 What it does: This paper proposes a method for learning product distributions on the Boolean hypercube with only unlabeled samples and a given imperfect 'suggestion' mean vector q.
ProDyG: Progressive Dynamic Scene Reconstruction via Gaussian Splatting from Monocular Videos
Shi Chen (ETH Zürich), Martin R. Oswald (University of Amsterdam)
Object TrackingPose EstimationGaussian SplattingSimultaneous Localization and MappingOptical FlowVideoPoint Cloud
🎯 What it does: ProDyG combines online SLAM with 3D Gaussian Splatting to simultaneously track camera pose, separate static and dynamic scenes from monocular video, and generate detailed dynamic 3D representations in real-time.
PROFIT: A Specialized Optimizer for Deep Fine Tuning
Anirudh S Chakravarthy (GMCruise LLC), Zhao Chen (GMCruise LLC)
Autonomous DrivingOptimizationSupervised Fine-TuningVision Language ModelImageMultimodality
🎯 What it does: An optimizer called PROFIT has been developed specifically for fine-tuning converged models.
ProfiX: Improving Profile-Guided Optimization in Compilers with Graph Neural Networks
Huiri Tan (Hong Kong University of Science and Technology), Jiasi Shen (Hong Kong University of Science and Technology)
OptimizationGraph Neural NetworkGraph
🎯 What it does: Using a graph neural network model to infer noise and missing execution frequencies collected from sampling-based PGO to enhance compiler optimization effectiveness.
Program Synthesis via Test-Time Transduction
Kang-il Lee (Seoul National University), Kyomin Jung (Seoul National University)
AI Code AssistantTransformerLarge Language ModelReinforcement LearningTextChain-of-Thought
🎯 What it does: Proposed and implemented the SYNTRA framework, which generates candidate programs through LLM and actively performs transductive prediction and hypothesis elimination on visible test inputs to enhance the robustness of program synthesis.
Progress Reward Model for Reinforcement Learning via Large Language Models
Xiuhui Zhang (Beihang University), Yue Deng (Beihang University)
Robotic IntelligenceTransformerLarge Language ModelReinforcement LearningMultimodalityChain-of-Thought
🎯 What it does: The PRM4RL framework is proposed, which utilizes large language models to decompose long tasks into subtasks and construct a progress reward model, integrating high-level planning with low-level rewards.
Progressive Data Dropout: An Embarrassingly Simple Approach to Train Faster
Shriram M S (University of Manchester), Shreyank N Gowda (University of Nottingham)
ClassificationComputational EfficiencyImage
🎯 What it does: A strategy called Progressive Data Dropout (PDD) is proposed, which significantly shortens training time while maintaining or improving accuracy.
Progressive Inference-Time Annealing of Diffusion Models for Sampling from Boltzmann Densities
Tara Akhound-Sadegh (Mila - Quebec AI Institute), Alexander Tong (AITHYRA)
Diffusion model
🎯 What it does: A stepwise inference time annealing framework named PITA is designed for training diffusion models to sample Boltzmann densities at low temperatures.
Projecting Assumptions: The Duality Between Sparse Autoencoders and Concept Geometry
Sai Sumedh R. Hindupur (Harvard University), Demba E. Ba (Harvard University)
OptimizationRepresentation LearningAuto EncoderImage
🎯 What it does: This paper reveals the implicit bias of Sparse Autoencoders (SAE) in discovering model concepts through theoretical analysis and experimental validation, and proposes a new geometry-based SAE (SpaDE) to overcome these biases.
Projection-based Lyapunov method for fully heterogeneous weakly-coupled MDPs
XiangCheng Zhang, Weina Wang (Carnegie Mellon University)
OptimizationReinforcement Learning
🎯 What it does: Proposed and analyzed an ID redistribution strategy applicable to completely heterogeneous weakly coupled Markov decision processes (WCMDP), proving its near-optimality with an optimality gap of O(1/√N) in large-scale systems;
Projection-Manifold Regularized Latent Diffusion for Robust General Image Fusion
Lei Cao (Wuhan University), Jiayi Ma (Wuhan University)
RestorationSegmentationDiffusion modelImage
🎯 What it does: This paper proposes a training-free image fusion framework based on a pre-trained latent diffusion model, called PDFuse, which utilizes projection-manifold regularization to achieve multi-source information fusion.
Projective Equivariant Networks via Second-order Fundamental Differential Invariants
Yikang Li (Peking University), Zhouchen Lin (Peking University)
ClassificationDomain AdaptationConvolutional Neural NetworkImage
🎯 What it does: Designed and implemented the first fully realized projective equivariant deep network PDINet, utilizing second-order projective differential invariants to construct pluggable equivariant operators.
Prompt Tuning Decision Transformers with Structured and Scalable Bandits
Finn Rietz (Örebro University), Lele Cao (King AI Labs Microsoft Gaming)
TransformerReinforcement LearningPrompt EngineeringSequential
🎯 What it does: A method for prompt tuning based on multi-armed bandits is proposed to generate optimal trajectory prompts for the pre-trained Prompt Decision Transformer (PDT) in offline multi-task reinforcement learning, enhancing performance and adapting to out-of-vocabulary (OOV) tasks.
Prompt Tuning Transformers for Data Memorization
Haiyu Wang (Chinese University of Hong Kong), Yuanyuan Lin (Chinese University of Hong Kong)
TransformerPrompt EngineeringText
🎯 What it does: This paper studies the theoretical and empirical analysis of prompt tuning in terms of data memorization capabilities, exploring the trade-off between prompt length and the number of autoregressive generation steps.
Prompt-Guided Alignment with Information Bottleneck Makes Image Compression Also a Restorer
Xuelin Shen (Guangdong Laboratory of Artificial Intelligence and Digital Economy), Wenhan Yang (Peng Cheng Laboratory)
RestorationCompressionPrompt EngineeringAuto EncoderImage
🎯 What it does: A lightweight prompt-guided information bottleneck compression scheme is proposed, unifying denoising and compression potential representations.
Prompt-guided Disentangled Representation for Action Recognition
tianci wu, zhang liang
RecognitionRepresentation LearningGraph Neural NetworkPrompt EngineeringVideo
🎯 What it does: This paper proposes a prompt-guided decoupled representation learning framework called ProDA, which can extract and separate subgraph representations of any specified action from spatial-temporal scene graphs in multi-action videos, thereby enhancing action recognition and localization performance.
Promptable 3-D Object Localization with Latent Diffusion Models
Cheng-Yao Hong (Institute of Information Science), Tyng-Luh Liu (Institute of Information Science)
Object DetectionPose EstimationTransformerPrompt EngineeringDiffusion modelMultimodalityPoint Cloud
🎯 What it does: A potential diffusion model framework for 3D object localization through prompts is proposed.
Prompted Policy Search: Reinforcement Learning through Linguistic and Numerical Reasoning in LLMs
Yifan Zhou (Arizona State University), Heni Ben Amor (Procter and Gamble)
OptimizationTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTabular
🎯 What it does: Proposed the Prompted Policy Search (ProPS) framework, which utilizes large language models (LLMs) to directly perform policy search in reinforcement learning, integrating numerical rewards and semantic information through language prompts to form a unified optimization loop;
ProRL: Prolonged Reinforcement Learning Expands Reasoning Boundaries in Large Language Models
Mingjie Liu (NVIDIA), Yi Dong (NVIDIA)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: In this paper, the authors propose a new long-term reinforcement learning training framework called ProRL, which utilizes KL regularization and reference policy resets to perform long-term reinforcement learning on the performance of large language models in reasoning tasks; based on this, they trained a multi-task reasoning model with 1.5 billion parameters named Nemotron-Research-Reasoning-Qwen-1.5B.
ProSpero: Active Learning for Robust Protein Design Beyond Wild-Type Neighborhoods
Michal Kmicikiewicz (Helmholtz Munich), Ewa Szczurek (Helmholtz Munich)
OptimizationProtein Structure PredictionConvolutional Neural NetworkReinforcement LearningBiomedical Data
🎯 What it does: A proactive learning framework PROSPERO based on frozen pre-trained generative models is proposed, utilizing target masking and biologically constrained sequence Monte Carlo sampling to iteratively update the surrogate model and guide the generation of high-fitness and novel protein sequences.
Prot2Text-V2: Protein Function Prediction with Multimodal Contrastive Alignment
Xiao Fei (École Polytechnique), Michalis Vazirgiannis (Mohamed bin Zayed University of Artificial Intelligence)
GenerationProtein Structure PredictionTransformerLarge Language ModelContrastive LearningTextMultimodality
🎯 What it does: This paper proposes the Prot2Text-V2 model, which can directly generate free-text protein function descriptions from amino acid sequences.
Protein Design with Dynamic Protein Vocabulary
Nuowei Liu (East China Normal University), Yuanbin Wu
GenerationProtein Structure PredictionTransformerLarge Language ModelTextBiomedical DataRetrieval-Augmented Generation
🎯 What it does: A functional-driven protein design framework named PRODVA is proposed, which can dynamically retrieve and integrate text function descriptions with natural protein fragments into the generation process, outputting structurally feasible and functionally aligned protein sequences.
Protein Inverse Folding From Structure Feedback
Junde Xu (Chinese University of Hong Kong), Jiezhong Qiu
OptimizationProtein Structure PredictionSupervised Fine-TuningGraph
🎯 What it does: By combining the structural feedback of protein folding models with Direct Preference Optimization (DPO), the inverse folding (protein sequence design) model is optimized to generate sequences that are closer to the target three-dimensional structure.
ProtInvTree: Deliberate Protein Inverse Folding with Reward-guided Tree Search
Mengdi Liu (Institute of Computing Technology, Chinese Academy of Sciences), Xilin Chen (Institute of Computing Technology, Chinese Academy of Sciences)
Protein Structure PredictionReinforcement LearningDiffusion modelBiomedical Data
🎯 What it does: This paper proposes ProtInvTree, a protein inverse folding framework based on reward-guided tree search, which achieves stepwise sequence generation through phased focus and baselining actions.
Protocols for Verifying Smooth Strategies in Bandits and Games
Miranda Christ, Jonathan Shafer
OptimizationReinforcement Learning from Human Feedback
🎯 What it does: This paper proposes an interactive proof protocol for verifying the approximate optimality of smooth strategies in multi-armed bandits and multi-player regularized games.
ProtoPairNet: Interpretable Regression through Prototypical Pair Reasoning
Rose Gurung (University of Maine), Chaofan Chen (University of Maine)
Explainability and InterpretabilityImage
🎯 What it does: This paper proposes ProtoPairNet, which combines prototype pairs for interpretable regression prediction.
Provable Gradient Editing of Deep Neural Networks
Zhe Tao (University of California), Aditya V. Thakur (University of California)
OptimizationConvolutional Neural NetworkSupervised Fine-TuningImageText
🎯 What it does: A provable gradient editing method called ProGrad is proposed, which minimizes parameter changes while keeping model predictions unchanged and strictly satisfies gradient linear constraints.
Provable Meta-Learning with Low-Rank Adaptations
Jacob L. Block (University of Texas at Austin), Sanjay Shakkottai (University of Texas at Austin)
Meta LearningTransformerSupervised Fine-TuningImageText
🎯 What it does: This paper proposes a meta-learning framework based on Low-Rank Adaptors (LoRA) (PEFT-ML) to make pre-trained models more amenable to subsequent few-shot fine-tuning during the multi-task training phase.
Provable Ordering and Continuity in Vision-Language Pretraining for Generalizable Embodied Agents
Zhizhen Zhang (University of Queensland), Yadan Luo (University of Queensland)
Robotic IntelligenceVision Language ModelContrastive LearningVideoText
🎯 What it does: The AcTOL method is proposed, utilizing target-free constraint temporal consistency learning (VLO loss and Brownian Bridge continuity constraint) for visual-language pre-training on human action videos, enhancing the temporal order and continuity of embeddings.
Provable Sample-Efficient Transfer Learning Conditional Diffusion Models via Representation Learning
Ziheng Cheng (University of California), Cheng Zhang (Peking University)
RestorationData SynthesisRepresentation LearningMeta LearningDiffusion modelScore-based ModelImage
🎯 What it does: A theoretical and experimental framework is proposed to enhance the sample efficiency of Conditional Diffusion Models (CDM) in few-shot environments using conditional representation learning.
Provable Scaling Laws for the Test-Time Compute of Large Language Models
Yanxi Chen (Alibaba Group), Jingren Zhou (Alibaba Group)
TransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: Two types of black-box LLM-based two-stage elimination and league-style aggregation algorithms are proposed, along with a provable extension law for computational performance during testing.
Provable Watermarking for Data Poisoning Attacks
Yifan Zhu (Chinese Academy of Sciences), Xiao-Shan Gao (Chinese Academy of Sciences)
Adversarial AttackData-Centric LearningConvolutional Neural NetworkImage
🎯 What it does: Two provable watermarking schemes are proposed, namely post-poisoning and poisoning-concurrent, for declaring and detecting poisoned data in data poisoning attacks.
Provably Efficient Multi-Task Meta Bandit Learning via Shared Representations
Jiabin Lin (Iowa State University), Shana Moothedath (Iowa State University)
Recommendation SystemMeta LearningTabular
🎯 What it does: This paper proposes a linear Bandit meta-learning framework based on shared low-dimensional representations, addressing two core issues of multi-task learning and transfer learning.
Provably Efficient Online RLHF with One-Pass Reward Modeling
Long-Fei Li (Nanjing University), Zhi-Hua Zhou (Nanjing University)
OptimizationReinforcement Learning from Human FeedbackTransformerReinforcement LearningText
🎯 What it does: Proposes a one-pass reward modeling algorithm for online RLHF that eliminates the need for historical data storage and achieves constant time updates per round;
Provably Efficient RL under Episode-Wise Safety in Constrained MDPs with Linear Function Approximation
Toshinori Kitamura (University of Tokyo), Yutaka Matsuo (University of Tokyo)
Safty and PrivacyComputational EfficiencyReinforcement Learning
🎯 What it does: This paper proposes a reinforcement learning algorithm for constrained Markov decision processes (CMDP) based on linear function approximation, which ensures safety constraints in each episode while achieving sub-linear regret loss.
Proximalized Preference Optimization for Diverse Feedback Types: A Decomposed Perspective on DPO
Kaiyang Guo (Huawei Noah's Ark Lab), Zhitang Chen (Huawei Noah's Ark Lab)
Recommendation SystemOptimizationReinforcement LearningText
🎯 What it does: By decomposing the DPO loss, the PRO (Proximalized Preference Optimization) method is proposed to uniformly handle various types of feedback (pairwise, binary, scalar) and eliminate the likelihood uncertainty (reward-hacking) problem of DPO.
Proxy Target: Bridging the Gap Between Discrete Spiking Neural Networks and Continuous Control
Zijie Xu (Peking University), Zhaofei Yu (Peking University)
Spiking Neural NetworkReinforcement LearningSequential
🎯 What it does: A proxy target framework is proposed to address the mismatch between the discrete SNN and the soft update mechanism of the continuous target network, stabilizing the reinforcement learning training of SNN in continuous control.
Proxy-SPEX: Sample-Efficient Interpretability via Sparse Feature Interactions in LLMs
Landon Butler (University of California Berkeley), Kannan Ramchandran (University of California Berkeley)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: PROXYSPEX is proposed, an efficient explanation method for discovering sparse hierarchical interactions in LLMs through Gradient Boosting Trees (GBT).
PRSformer: Disease Prediction from Million-Scale Individual Genotypes
Payam Dibaeinia (23andMe), Aly A Khan
ClassificationOptimizationTransformerBiomedical Data
🎯 What it does: A multi-task Transformer model named PRSformer has been constructed to directly predict the risk of 18 autoimmune and inflammatory diseases from millions of individual genotypes.
Pruning Spurious Subgraphs for Graph Out-of-Distribution Generalization
Tianjun Yao (Mohamed bin Zayed University of Artificial Intelligence), Zhiqiang Shen (Mohamed bin Zayed University of Artificial Intelligence)
Graph Neural NetworkGraph
🎯 What it does: A pruning-based graph OOD method called PrunE is proposed, which enhances the generalization ability of graph neural networks on out-of-distribution data by removing spurious edges to maintain invariant subgraphs.
Pruning-Robust Mamba with Asymmetric Multi-Scale Scanning Paths
Jindi Lv (Sichuan University), Kai Wang
ClassificationSegmentationConvolutional Neural NetworkImage
🎯 What it does: A visual Mamba architecture named AMVim is proposed, which enhances the robustness of token pruning by utilizing asymmetric multi-scale scanning paths.
Pseudo-Riemannian Graph Transformer
Le Viet Quan, Viet Cuong Ta (VNU University of Engineering and Technology)
ClassificationRepresentation LearningGraph Neural NetworkTransformerGraph
🎯 What it does: A differential isometric framework for graph embedding is proposed, which decomposes pseudo-Riemannian manifolds into a product of spheres and hyperbolic spaces, and based on this, a pseudo-Riemannian graph Transformer (Q-GT) is constructed to effectively represent complex graph structures.
PseuZO: Pseudo-Zeroth-Order Algorithm for Training Deep Neural Networks
Pengyun Yue (Peking University), Zhouchen Lin (Peking University)
OptimizationLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A Pseudo-Zeroth-Order (PseuZO) optimization framework is proposed, which first estimates the Jacobian matrix of the model output, then combines it with the gradient of the external loss, and employs exponential moving average and sliding window techniques to achieve efficient fine-tuning of large-scale LLMs.
PT-MoE: An Efficient Finetuning Framework for Integrating Mixture-of-Experts into Prompt Tuning
Zongqian Li (University of Cambridge), Nigel Collier (University of Cambridge)
TransformerPrompt EngineeringMixture of ExpertsText
🎯 What it does: PT-MoE is proposed, an efficient prompt tuning framework that combines matrix factorization with mixture of experts (MoE) routing.