NeurIPS 2024 Papers — Page 29
Conference on Neural Information Processing Systems · 4035 papers
Predictor-Corrector Enhanced Transformers with Exponential Moving Average Coefficient Learning
Bei Li (Northeastern University), Xunliang Cai (Meituan Inc.)
TransformerTextOrdinary Differential Equation
🎯 What it does: A PCformer model is proposed, utilizing a prediction-corrector framework combined with exponential moving average coefficient learning to improve the performance of Transformers in various NLP tasks.
Preference Alignment with Flow Matching
Minu Kim (KAIST), Se-Young Yun (KAIST)
GenerationRecommendation SystemReinforcement Learning from Human FeedbackFlow-based ModelImageTextOrdinary Differential Equation
🎯 What it does: Proposes Preference Flow Matching (PFM), which directly learns a vector field from low-preference data to high-preference data through flow matching, achieving preference alignment in black-box models.
Preference Learning Algorithms Do Not Learn Preference Rankings
Angelica Chen (New York University), Kyunghyun Cho (Genentech)
Recommendation SystemReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper investigates whether preference learning algorithms (RLHF and DPO) truly enhance the accuracy of preference ranking when aligning large language models, and through theoretical derivation and experimental validation, it finds that their ranking accuracy on the training set is far below the ideal value, leading to a significant alignment gap.
Preference Learning of Latent Decision Utilities with a Human-like Model of Preferential Choice
Sebastiaan De Peuter (Aalto University), Samuel Kaski (Aalto University)
Recommendation SystemOptimizationTabular
🎯 What it does: Two computable and inferable choice models, Computable Rationality with Cross-Feature Extension (CRCS and LC-CRCS), are proposed to learn latent utility functions from human preference data, and their utility is validated in multi-task cases (structural design, hydraulic networks, chemical synthesis planning).
Preference-based Pure Exploration
Apurv Shukla (University of Michigan), Debabrota Basu (University of Lille)
OptimizationDrug Discovery
🎯 What it does: This paper studies the fixed confidence pure exploration (Preference-based Pure Exploration, PrePEx) problem under a vector reward multi-armed bandit model with a given preference cone, aiming to accurately identify the Pareto optimal arm set that satisfies the preference constraints with the least number of samples.
Preferential Normalizing Flows
Petrus Mikkola (University of Helsinki), Arto Klami (University of Helsinki)
Recommendation SystemOptimizationFlow-based ModelTabular
🎯 What it does: Utilizing preference comparison/ranking information, the high-dimensional probability density of experts is learned using normalized flows.
PrefPaint: Aligning Image Inpainting Diffusion Model with Human Preference
Kendong Liu (City University of Hong Kong), Junhui Hou (City University of Hong Kong)
RestorationGenerationReinforcement Learning from Human FeedbackReinforcement LearningDiffusion modelImage
🎯 What it does: Aligning the image inpainting task based on diffusion models to make the generated results more in line with human aesthetic preferences, while constructing a human preference dataset in the process.
Pretrained Optimization Model for Zero-Shot Black Box Optimization
Xiaobin Li (Xidian University), Jing Liu (Xidian University)
OptimizationTransformerTabular
🎯 What it does: A pre-trained population optimization model (POM) is proposed, which can directly solve new black-box optimization tasks with little or no tuning;
Pretrained Transformer Efficiently Learns Low-Dimensional Target Functions In-Context
Kazusato Oko (University of California), Denny Wu (New York University)
TransformerTabular
🎯 What it does: This study investigates the in-context learning of pre-trained Transformers in high-dimensional Gaussian single-index models, providing a theoretical sample complexity analysis and experimental validation.
Pretraining Codomain Attention Neural Operators for Solving Multiphysics PDEs
Md Ashiqur Rahman (Purdue University), Anima Anandkumar (NVIDIA)
Graph Neural NetworkTransformerSupervised Fine-TuningGraphPhysics Related
🎯 What it does: This paper proposes the Codomain Attention Neural Operator (CoDA‑NO) and conducts self-supervised pre-training to support rapid inference and few-shot fine-tuning of multi-physics coupled PDEs.
Pretraining with Random Noise for Fast and Robust Learning without Weight Transport
Jeonghwan Cheon (Korea Advanced Institute of Science and Technology), Se-Bum Paik (Korea Advanced Institute of Science and Technology)
ClassificationOptimizationMeta LearningConvolutional Neural NetworkReinforcement LearningImage
🎯 What it does: This paper pre-trains the network using random noise inputs and random labels without real data, and subsequently employs the feedback alignment (FA) algorithm during training with real data. It observes that pre-training significantly enhances learning speed and generalization performance.
Preventing Dimensional Collapse in Self-Supervised Learning via Orthogonality Regularization
Junlin He (Hong Kong Polytechnic University), Wei Ma (Hong Kong Polytechnic University)
Object DetectionRepresentation LearningConvolutional Neural NetworkTransformerContrastive LearningImage
🎯 What it does: This paper proposes adding orthogonal regularization to the encoder (including convolutional layers and linear layers) during the pre-training phase of self-supervised learning to prevent the collapse of the weight matrix, hidden features, and representation vectors.
Preventing Model Collapse in Deep Canonical Correlation Analysis by Noise Regularization
Junlin He (Hong Kong Polytechnic University), Wei Ma (Hong Kong Polytechnic University)
Representation LearningAuto EncoderMultimodality
🎯 What it does: This study investigates the model collapse issue of Deep Canonical Correlation Analysis (DCCA) in multi-view representation learning and proposes a Noise Regularization method (NR-DCCA) to prevent a sharp decline in performance during training.
Pricing and Competition for Generative AI
Rafid Mahmood (NVIDIA)
GenerationOptimization
🎯 What it does: Analyzed and constructed a pricing and competition framework for generative AI models, proposing to transform multi-task price competition into piecewise optimization based on performance ratio, and provided a closed-form optimal solution under exponential demand.
Principled Bayesian Optimization in Collaboration with Human Experts
Wenjie Xu (Ecole Polytechnique Federale de Lausanne), Michael A Osborne
OptimizationTabular
🎯 What it does: A collaborative Bayesian optimization framework is designed, utilizing binary labels from experts to guide sampling, and achieving two theoretical guarantees: 'harmless' and 'handover'.
Principled Probabilistic Imaging using Diffusion Models as Plug-and-Play Priors
Zihui Wu (California Institute of Technology), Katherine Bouman
RestorationSuper ResolutionDiffusion modelImageStochastic Differential Equation
🎯 What it does: A posterior sampling method based on diffusion models, PnP-DM, is proposed, which combines split Gibbs sampling with the EDM framework to achieve high-quality reconstruction and uncertainty quantification for both linear and nonlinear inverse problems.
Prior-itizing Privacy: A Bayesian Approach to Setting the Privacy Budget in Differential Privacy
Zeki Kazan (Duke University), Jerome Reiter
Safty and PrivacyTabularBiomedical Data
🎯 What it does: This paper proposes a framework based on Bayesian posterior-prior risk (relative disclosure risk) to determine the value of the privacy budget ε in differential privacy (DP), and provides the corresponding closed-form solution and minimization problem.
Prism: A Framework for Decoupling and Assessing the Capabilities of VLMs
Yuxuan Qiao (Nanjing University), Kai Chen (Shanghai AI Laboratory)
RecognitionOptimizationComputational EfficiencyTransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: This paper proposes the Prism framework, which decouples the visual perception and reasoning processes of visual language models (VLMs). It can be used for systematic evaluation of the perception and reasoning capabilities of VLMs, and it combines lightweight perception modules with powerful LLMs to build efficient visual language processors.
Privacy Backdoors: Enhancing Membership Inference through Poisoning Pre-trained Models
Yuxin Wen (University of Maryland), Nicholas Carlini (Google DeepMind)
Safty and PrivacyAdversarial AttackTransformerSupervised Fine-TuningImageTextBiomedical Data
🎯 What it does: This study investigates a method to implant privacy backdoors in pre-trained models, resulting in a higher leakage of membership inference for the fine-tuned model on training samples.
Privacy without Noisy Gradients: Slicing Mechanism for Generative Model Training
Kristjan Greenewald (MIT-IBM Watson AI Lab), Kai Xu (MIT-IBM Watson AI Lab)
Data SynthesisDomain AdaptationSafty and PrivacyImageTabular
🎯 What it does: By adding noise to data projections and using the gradient-free smoothed-sliced f-divergence to train generative models, differential privacy synthetic data is achieved.
Private Algorithms for Stochastic Saddle Points and Variational Inequalities: Beyond Euclidean Geometry
Raef Bassily (Ohio State University), Michael Menart (Ohio State University)
OptimizationSafty and PrivacyReinforcement Learning
🎯 What it does: This paper studies privacy algorithms for the stochastic saddle point problem (SSP) and stochastic variational inequality (SVI) under the (ε,δ)-differential privacy constraint, and provides nearly optimal convergence rates.
Private and Personalized Frequency Estimation in a Federated Setting
Amrith Setlur (Carnegie Mellon University), Kunal Talwar (Apple)
Federated LearningSafty and PrivacyText
🎯 What it does: In a federated environment, personalized estimation of the term frequency histogram for each user aims to minimize KL divergence error while enhancing estimation accuracy through clustering methods, all under the premise of maintaining user-level joint differential privacy.
Private Attribute Inference from Images with Vision-Language Models
Batuhan Tömekçe (ETH Zurich), Martin Vechev (ETH Zurich)
Safty and PrivacyTransformerPrompt EngineeringVision Language ModelImageChain-of-Thought
🎯 What it does: This study investigates the privacy risks of visual-language models (VLM) in inferring personal attributes from non-human images and constructs the VIP dataset for evaluation.
Private Edge Density Estimation for Random Graphs: Optimal, Efficient and Robust
Hongjie Chen (ETH Zurich), David Steurer (ETH Zurich)
Safty and PrivacyComputational EfficiencyGraph Neural NetworkGraph
🎯 What it does: A polynomial-time, differentially node-private, and robust algorithm is proposed for estimating the edge density of Erdős-Rényi random graphs and their generalization to non-uniform random graphs.
Private Geometric Median
Mahdi Haghifam (Northeastern University), Jonathan Ullman (Northeastern University)
OptimizationSafty and PrivacyTabular
🎯 What it does: This study investigates differentially private algorithms for computing geometric medians and proposes two polynomial-time DP algorithms that ensure the error is related to the effective diameter of the data points.
Private Online Learning via Lazy Algorithms
Hilal Asi (Apple), Kunal Talwar (Apple)
OptimizationSafty and Privacy
🎯 What it does: A new transformation (L2P) is proposed to convert a lazy (low-switching) online learning algorithm into a differentially private online learning algorithm, which is applied to the DP-OPE and DP-OCO problems, achieving a better asymptotic regret upper bound.
Private Stochastic Convex Optimization with Heavy Tails: Near-Optimality from Simple Reductions
Hilal Asi (Apple Inc.), Kevin Tian (University of Texas at Austin)
OptimizationSafty and Privacy
🎯 What it does: This paper studies the differential privacy stochastic convex optimization (DP-SCO) problem with heavy-tailed gradient distributions and proposes a novel reduction framework based on population-level localization, achieving an almost optimal high-probability error upper bound.
PrivCirNet: Efficient Private Inference via Block Circulant Transformation
Tianshi Xu (Peking University), Meng Li (Peking University)
Safty and PrivacyComputational EfficiencyConvolutional Neural NetworkTransformerImage
🎯 What it does: The PrivCirNet framework is proposed, which significantly improves the efficiency of DNN encrypted inference through the collaborative optimization of block cyclic matrix transformation and homomorphic encryption protocols.
Probabilistic Conformal Distillation for Enhancing Missing Modality Robustness
mengxi Chen, Yanfeng Wang (Shanghai Jiao Tong University)
RecognitionSegmentationKnowledge DistillationConvolutional Neural NetworkImageMultimodality
🎯 What it does: A probabilistic shape distortion (PCD) method is proposed to enhance the robustness of multimodal models in the absence of certain modalities;
Probabilistic Decomposed Linear Dynamical Systems for Robust Discovery of Latent Neural Dynamics
Yenho Chen (Georgia Institute of Technology), Christopher John Rozell
Anomaly DetectionRepresentation LearningTime SeriesSequentialBiomedical Data
🎯 What it does: This paper proposes a probabilistic decomposition linear dynamical system (p-dLDS) and introduces time-varying bias terms and probabilistic sparsity structures to discover interpretable latent dynamics in high-dimensional neural signals.
Probabilistic Federated Prompt-Tuning with Non-IID and Imbalanced Data
Pei-Yau Weng (Washington State University), Trong Nghia Hoang (Washington State University)
OptimizationFederated LearningTransformerPrompt EngineeringImage
🎯 What it does: In a federated learning environment, prompt tuning of pre-trained models is conducted to address data heterogeneity and imbalance issues, proposing a probabilistic prompt aggregation method.
Probabilistic Graph Rewiring via Virtual Nodes
Chendi Qian (RWTH Aachen University), Mathias Niepert (University of Stuttgart)
Drug DiscoveryGraph Neural NetworkGraph
🎯 What it does: This paper proposes an implicit probability graph re-connection message passing neural network (IPR-MPNN), which enhances the connectivity and information propagation of the graph while maintaining low complexity by adding a small number of virtual nodes and learning the connection probabilities between the original nodes and these virtual nodes, thereby improving the expressive power of MPNN.
Probabilistic size-and-shape functional mixed models
Fangyi Wang (Ohio State University), Sebastian Kurtek (Ohio State University)
Time SeriesBiomedical DataElectrocardiogram
🎯 What it does: A Bayesian functional mixture model based on norm-preserving transformations has been constructed to estimate the average size shape µ of functional data and quantify its uncertainty.
Probabilistic Weather Forecasting with Hierarchical Graph Neural Networks
Joel Oskarsson (Linköping University), Fredrik Lindsten (Linköping University)
Graph Neural NetworkGraphTime Series
🎯 What it does: The Graph-EFM model is proposed, which achieves efficient probabilistic weather forecasting by introducing hierarchical grids and latent variables into graph neural networks.
Probablistic Emulation of a Global Climate Model with Spherical DYffusion
Salva Rühling Cachay (University of California San Diego), Rose Yu (University of California San Diego)
Diffusion modelTime Series
🎯 What it does: A conditional generative model named Spherical DYffusion has been developed to realistically simulate global climate sequences from 10 to 100 years while maintaining low computational costs.
Probing Social Bias in Labor Market Text Generation by ChatGPT: A Masked Language Model Approach
Lei Ding (University of Alberta), Bei Jiang (University of Alberta)
GenerationTransformerLarge Language ModelText
🎯 What it does: Through experimental design, this study utilizes ChatGPT to generate job application materials, combined with the PRISM algorithm to assess gender bias, exploring the impact of AI-generated text on gender inequality in the labor market.
Probing the Decision Boundaries of In-context Learning in Large Language Models
Siyan Zhao (University of California), Aditya Grover (University of California)
ClassificationOptimizationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: By visualizing the decision boundaries of binary classification tasks, this paper analyzes the contextual learning ability of large language models and proposes a smoothing method.
Procedure-Aware Surgical Video-language Pretraining with Hierarchical Knowledge Augmentation
Kun yuan, Nicolas Padoy (University of Strasbourg)
RecognitionRetrievalConvolutional Neural NetworkLarge Language ModelContrastive LearningVideoTextMultimodality
🎯 What it does: A scheme for video-language pretraining in surgical procedures is proposed, which includes hierarchical knowledge enhancement and the Procedure-Encoded Surgical Knowledge-Augmented Video-Language Pretraining (PeskaVLP) framework.
PRODuctive bandits: Importance Weighting No More
Julian Zimmert (Google Research), Teodor Vanislavov Marinov
OptimizationReinforcement Learning from Human FeedbackReinforcement Learning
🎯 What it does: This paper improves the Prod algorithm to achieve optimal asymptotic unbiased returns in multi-armed bandit (MAB) tasks, and proposes a variant that does not require importance weighting, as well as an optimal-both algorithm that can satisfy both stochastic and adversarial environments.
ProEdit: Simple Progression is All You Need for High-Quality 3D Scene Editing
Jun-Kun Chen (University of Illinois Urbana-Champaign), Yu-Xiong Wang (University of Illinois Urbana-Champaign)
GenerationOptimizationDiffusion modelGaussian SplattingPoint Cloud
🎯 What it does: This paper proposes a 3D scene editing framework called ProEdit, which addresses the multi-view inconsistency problem caused by a large feasible output space through sub-task decomposition and progressive execution.
Progressive Entropic Optimal Transport Solvers
Parnian Kassraie (ETH Zurich), marco cuturi
OptimizationSupervised Fine-TuningImagePoint Cloud
🎯 What it does: A Progressive Entropic Optimal Transport (PROGOT) solver is proposed, which can simultaneously output the OT plan and Monge mapping without the need for precise parameter tuning, and exhibits good numerical stability and consistency on large-scale point clouds.
Progressive Exploration-Conformal Learning for Sparsely Annotated Object Detection in Aerial Images
Zihan Lu (Nanjing University of Science and Technology), Zhen Cui (Shandong Normal University)
Object DetectionReinforcement LearningImage
🎯 What it does: A Progressive Exploration-Conformal Learning (PECL) framework is proposed for sparse annotation in aerial image object detection.
Promoting Fairness Among Dynamic Agents in Online-Matching Markets under Known Stationary Arrival Distributions
Will Ma (Columbia University), Pan Xu (New Jersey Institute of Technology)
🎯 What it does: This paper proposes the online matching model for long-term fairness (FAIR-L) in online types, named OM-LF, and designs optimal or approximate algorithms for single and multiple offline agents, further extending to group-level fairness and short-term fairness.
Prompt Optimization with EASE? Efficient Ordering-aware Automated Selection of Exemplars
Zhaoxuan Wu (National University of Singapore), Bryan Kian Hsiang Low (National University of Singapore)
OptimizationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: The EASE method is proposed for automatically selecting ordered example sets to enhance the contextual learning effectiveness of large language models.
Prompt Tuning Strikes Back: Customizing Foundation Models with Low-Rank Prompt Adaptation
Abhinav Jain (Rice University), Chris Jermaine (Rice University)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes LoPA, a low-rank prompt adaptation method that constructs prefixes using both task-level and instance-level soft prompts, enabling efficient and user-customizable fine-tuning of large foundational models.
Prompt-Agnostic Adversarial Perturbation for Customized Diffusion Models
Cong Wan (Xi'an Jiaotong University), Yihong Gong (Xi'an Jiaotong University)
GenerationSafty and PrivacyAdversarial AttackDiffusion modelImage
🎯 What it does: In response to the privacy and copyright risks of customized text-to-image diffusion models, a Prompt-agnostic Adversarial Perturbation (PAP) method is proposed, which adds subtle perturbations to images, preventing the model from generating high-quality images of specific individuals or styles under any prompt, thereby protecting personal data and artistic creations.
PromptFix: You Prompt and We Fix the Photo
Yongsheng Yu (University of Rochester), Jiebo Luo (University of Rochester)
Image TranslationRestorationSuper ResolutionLarge Language ModelPrompt EngineeringVision Language ModelDiffusion modelImage
🎯 What it does: A large-scale low-level image processing dataset containing approximately 1.01 million input-target-instruction triples has been constructed, and a unified PromptFix model has been proposed, capable of completing various low-level image processing tasks based on user-defined instructions.
Propensity Score Alignment of Unpaired Multimodal Data
Johnny Xi (University of British Columbia), Jason Hartford (Valence Labs)
ClassificationData SynthesisDomain AdaptationMultimodalityBiomedical Data
🎯 What it does: By training classifiers on each modality to estimate propensity scores, and using these scores to achieve soft matching of unpaired multimodal data within the same experimental label through optimal transport (OT) or shared nearest neighbors (SNN); subsequently, the matching matrix is utilized for cross-modal prediction or downstream tasks.
Proportional Fairness in Clustering: A Social Choice Perspective
Leon Kellerhals (Technische Universitat Clausthal), Jannik Peters (National University of Singapore)
🎯 What it does: This paper combines the proportional fair clustering problem with the concept of proportional representation in multi-winner elections, proving the close relationship between proportional fairness, individual fairness, and the transferable core. Based on this, it proposes two metric-based axioms of proportional representation (mJR and mPJR), further providing algorithms to achieve these axioms and their approximation guarantees. It also improves the approximate upper bound of the fair core (q-core) in the sortition scenario, extending the treatment of infinite candidate sets.
Proportional Fairness in Non-Centroid Clustering
Ioannis Caragiannis (Aarhus University), Nisarg Shah (University of Toronto)
OptimizationTabular
🎯 What it does: This study investigates proportional fairness under non-centroid clustering, proposes core and fully justified representation (FJR) guarantees, and presents greedy algorithms and auditing methods.
Prospective Learning: Learning for a Dynamic Future
Ashwin De Silva (Johns Hopkins University), Pratik Chaudhari (University of Pennsylvania)
ClassificationOptimizationConvolutional Neural NetworkImageSequential
🎯 What it does: A prospective learning framework is proposed, along with the Prospective ERM algorithm; experimental validation is conducted on synthetic data and periodic/hidden Markov tasks from MNIST and CIFAR-10.
Prospective Representation Learning for Non-Exemplar Class-Incremental Learning
Wuxuan Shi (Wuhan University), Mang Ye (Wuhan University)
Knowledge DistillationRepresentation LearningConvolutional Neural NetworkAuto EncoderImage
🎯 What it does: A prospective representation learning framework is proposed, which reserves space through embedding compression constraints in the benchmark phase, and utilizes prototype-guided representation updates in the incremental phase to achieve separation of old and new classes in the latent space, thereby reducing forgetting and conflicts in zero-shot class incremental learning.
ProSST: Protein Language Modeling with Quantized Structure and Disentangled Attention
Mingchen Li (Shanghai Jiao Tong University), Liang Hong (Shanghai Jiao Tong University)
Protein Structure PredictionTransformerLarge Language ModelBiomedical Data
🎯 What it does: ProSST has been developed, a Transformer language model that combines protein sequences and three-dimensional structures, and learns rich contextual representations through pre-training.
Protected Test-Time Adaptation via Online Entropy Matching: A Betting Approach
Yarin Bar (Technion Israel Institute of Technology), Yaniv Romano (Technion Israel Institute of Technology)
Domain AdaptationImage
🎯 What it does: This paper proposes a testing-time adaptive method called POEM based on online entropy matching, which improves inference accuracy under distribution shifts by detecting entropy distribution drift and adaptively updating model parameters.
Protecting Your LLMs with Information Bottleneck
Zichuan Liu (Nanjing University), Jiang Bian (Microsoft Research Asia)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: A lightweight LLM defense framework named IBProtector is designed and implemented, utilizing a trainable prompt extractor to compress and perturb input prompts, retaining only the information crucial to the answers, thereby resisting prompt-level and token-level jailbreak attacks.
Protein-Nucleic Acid Complex Modeling with Frame Averaging Transformer
Tinglin Huang (Yale University), Wengong Jin (Northeastern University)
Protein Structure PredictionTransformerGraphBiomedical Data
🎯 What it does: An unsupervised protein-nucleic acid complex contact map prediction model FAFormer was developed, and the predicted contact maps were used for unsupervised aptamer screening.
ProtGO: Function-Guided Protein Modeling for Unified Representation Learning
Bozhen Hu (Zhejiang University), Stan Z. Li (Westlake University)
Knowledge DistillationRepresentation LearningProtein Structure PredictionGraph Neural NetworkMultimodalityBiomedical Data
🎯 What it does: We propose ProtGO, a multimodal protein representation learning framework that integrates sequence, structure, and functional information into a unified embedding using teacher-student knowledge distillation.
Prototypical Hash Encoding for On-the-Fly Fine-Grained Category Discovery
Haiyang Zheng (University of Trento), Zhun Zhong (Hefei University of Technology)
ClassificationRecognitionTransformerContrastive LearningImage
🎯 What it does: A prototype-based hash coding framework PHE has been developed for online fine-grained category discovery.
ProTransformer: Robustify Transformers via Plug-and-Play Paradigm
Zhichao Hou (North Carolina State University), Xiaorui Liu (North Carolina State University)
OptimizationAdversarial AttackTransformerLarge Language ModelImageTextGraph
🎯 What it does: A pluggable robust attention mechanism called ProAttention is proposed, which enhances the attack resistance of Transformer models without the need for additional training or fine-tuning.
Provable Acceleration of Nesterov's Accelerated Gradient for Asymmetric Matrix Factorization and Linear Neural Networks
Zhenghao Xu (Georgia Institute of Technology), Molei Tao (Georgia Institute of Technology)
OptimizationTabular
🎯 What it does: This paper studies the global convergence rates of gradient descent (GD) and Nesterov accelerated gradient (NAG) in matrix factorization (especially rectangular low-rank factorization) and linear neural networks. It proves that under a certain unbalanced initialization, the iteration complexity for GD to converge to an ϵ-optimal solution is O(d²κ²log(1/ϵ)), while NAG achieves accelerated convergence with the iteration complexity reduced to O(dκlog(1/ϵ)). This analysis is also extended to linear neural networks, demonstrating that NAG can achieve a linear convergence rate when the width only needs to be ≥ output rank (+log(1/δ)).
Provable and Efficient Dataset Distillation for Kernel Ridge Regression
Yilan Chen (University of California San Diego), Tsui-Wei Weng (University of California San Diego)
Computational EfficiencyKnowledge DistillationImage
🎯 What it does: A theoretical analysis and efficient algorithm for dataset distillation of Kernel Ridge Regression (KRR) is proposed, demonstrating that under various settings, only one data point per class is needed to recover the original model performance.
Provable Benefit of Cutout and CutMix for Feature Learning
Junsoo Oh (KAIST), Chulhee Yun (KAIST)
Representation LearningConvolutional Neural NetworkImage
🎯 What it does: This paper explores the feature learning effects of Cutout and CutMix on a two-layer convolutional network through theoretical analysis, demonstrating that Cutout can learn rare features, CutMix can learn extremely rare features, and achieves nearly perfect test accuracy.
Provable Benefits of Complex Parameterizations for Structured State Space Models
Yuval Ran-Milo (Tel Aviv University), Nadav Cohen (Google)
Image
🎯 What it does: This paper discusses the theoretical and practical advantages of using complex parameterization in Structured State Space Models (SSM), demonstrating that complex SSMs outperform real-valued SSMs in expressiveness and learnability.
Provable Editing of Deep Neural Networks using Parametric Linear Relaxation
Zhe Tao (University of California), Aditya Thakur
OptimizationImageBenchmark
🎯 What it does: Designed and implemented PREPARED, a provably correct neural network editing method using parametric linear relaxation;
Provable Partially Observable Reinforcement Learning with Privileged Information
Yang Cai (Yale University), Kaiqing Zhang (University of Maryland)
Knowledge DistillationReinforcement Learning
🎯 What it does: This paper studies the use of privileged information (such as complete state observations) available during training to address the learning problems in partially observable environments (POMDP) and multi-agent POMDP (POSG) in reinforcement learning. Through theoretical analysis, it proves that two mainstream practical paradigms—expert distillation and asymmetric actor-critic—can achieve polynomial sample and computational complexity under appropriate observability assumptions. It also presents new deterministic filtering conditions and an approximate belief learning framework, further extending to multi-agent CTDE scenarios. Numerical experiments show that the new algorithm converges faster and achieves higher rewards than traditional methods.
Provable Posterior Sampling with Denoising Oracles via Tilted Transport
Joan Bruna (New York University), Jiequn Han (Flatiron Institute)
Stochastic Differential Equation
🎯 What it does: This paper proposes a posterior sampling method based on denoising oracles, utilizing tilted transport techniques to transform the original posterior sampling problem into a new problem that is easier to sample from.
Provable Tempered Overfitting of Minimal Nets and Typical Nets
Itamar Harel (Technion), Daniel Soudry (Technion)
🎯 What it does: This study investigates the overfitting behavior of deep threshold networks (binary threshold networks) when label noise is present, proving that under two learning rules (minimum parameter interpolator and random interpolator), it exhibits 'tempered overfitting'.
Provably and Practically Efficient Adversarial Imitation Learning with General Function Approximation
Tian Xu (Nanjing University), Yang Yu (Nanjing University)
Robotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningSequentialBenchmark
🎯 What it does: An online adversarial imitation learning framework OPT‑AIL is designed, utilizing online no-regret reward optimization and optimistic Bellman error minimization to learn rewards and Q-values, thereby achieving theoretically provable efficient imitation learning under general function approximation.
Provably Efficient Interactive-Grounded Learning with Personalized Reward
Mengxiao Zhang (University of Iowa), Paul Mineiro (Microsoft Research)
ClassificationReinforcement Learning from Human FeedbackReinforcement LearningImageText
🎯 What it does: This paper proposes a theoretically optimal algorithm in the context of Interaction-Grounded Learning (IGL) with feedback-dependent contexts, designing a Lipschitz-type reward estimator and implementing two provably efficient offline and online learning algorithms based on this estimator.
Provably Efficient Reinforcement Learning with Multinomial Logit Function Approximation
Long-Fei Li (Nanjing University), Zhi-Hua Zhou (Nanjing University)
OptimizationReinforcement Learning
🎯 What it does: In the Markov Decision Process (MDP) approximated by the Multinomial Logit (MNL) function, two efficient reinforcement learning algorithms are proposed, achieving both statistical and computational optimality;
Provably Faster Algorithms for Bilevel Optimization via Without-Replacement Sampling
Junyi Li (University of Maryland), Heng Huang (University of Maryland)
OptimizationTabular
🎯 What it does: This paper proposes WiOR-BO and WiOR-CBO algorithms based on random reshuffling (shuffle-once) to address non-convex-strongly convex bilevel optimization and conditional bilevel optimization, and extends to minimization and combinatorial optimization.
Provably Mitigating Overoptimization in RLHF: Your SFT Loss is Implicitly an Adversarial Regularizer
Zhihan Liu (Northwestern University), Zhaoran Wang (Northwestern University)
OptimizationReinforcement Learning from Human FeedbackSupervised Fine-TuningReinforcement LearningText
🎯 What it does: A theoretical and practical RLHF algorithm is proposed, which can explicitly alleviate the problem of over-optimization in the reward model.
Provably Optimal Memory Capacity for Modern Hopfield Models: Transformer-Compatible Dense Associative Memories as Spherical Codes
Jerry Yao-Chieh Hu (Northwestern University), Han Liu (Northwestern University)
Tabular
🎯 What it does: This paper studies the optimal memory capacity of the Kernelized Modern Hopfield model (KHM) and proposes a sublinear time algorithm U Hop-+ to achieve this capacity.
Provably Robust Score-Based Diffusion Posterior Sampling for Plug-and-Play Image Reconstruction
Xingyu Xu (Carnegie Mellon University), Yuejie Chi (Carnegie Mellon University)
RestorationSuper ResolutionDiffusion modelScore-based ModelImageStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: A posterior sampling framework DPnP based on the fractional diffusion model is proposed to solve nonlinear inverse problems.
Provably Safe Neural Network Controllers via Differential Dynamic Logic
Samuel Teuber (Karlsruhe Institute of Technology), Andre Platzer (Carnegie Mellon University)
Autonomous DrivingSafty and PrivacyTabularOrdinary Differential Equation
🎯 What it does: A method called VerSAILLE, which combines differential dynamic logic (dL) with neural network verification (NNV), is proposed to provide safety proofs for neural network-based control systems (NNCS) over an infinite time horizon. Additionally, a Mosaic framework is introduced, extending the linear constraint NNV tools to polynomial nonlinear and non-standard queries while maintaining completeness.
Provably Transformers Harness Multi-Concept Word Semantics for Efficient In-Context Learning
Dake Bu (City University of Hong Kong), Hau-San Wong (City University of Hong Kong)
TransformerText
🎯 What it does: This paper theoretically proves that during the training process of a single-layer single-head Transformer with softmax attention and ReLU MLP, utilizing the distribution of prompts based on multi-concept semantic linear geometric structures can achieve exponential convergence of 0-1 loss, and demonstrates that this model can efficiently perform contextual learning in cross-concept out-of-distribution (OOD) scenarios.
Proving Theorems Recursively
Haiming Wang (Sun Yat-sen University), Xiaodan Liang (Pengcheng Laboratory)
Large Language ModelSupervised Fine-TuningText
🎯 What it does: By recursively constructing verifiable proof sketches in a hierarchical manner and using sorry placeholders, POETRY advances layer by layer and ultimately achieves a complete proof.
ProvNeRF: Modeling per Point Provenance in NeRFs as a Stochastic Field
Kiyohiro Nakayama (Stanford University), Leonidas Guibas
GenerationData SynthesisNeural Radiance FieldPoint Cloud
🎯 What it does: Introduce the observable position (provenance) distribution of each 3D point in NeRF training, forming a random field;
Proximal Causal Inference With Text Data
Jacob M. Chen (Johns Hopkins University), Katherine A. Keith (Williams College)
TransformerLarge Language ModelTextBiomedical DataElectronic Health RecordsElectrocardiogram
🎯 What it does: This paper proposes a novel method that utilizes a zero-shot model for inferring two proxy variables from preprocessed text and applies it to proximal causal inference (proximal g-formula) to estimate causal effects in the absence of completely unobserved key confounding variables.
ProxyFusion: Face Feature Aggregation Through Sparse Experts
Bhavin Jawade (University at Buffalo), Venu Govindaraju (University at Buffalo)
RecognitionComputational EfficiencyMixture of ExpertsContrastive LearningImageVideo
🎯 What it does: A linear-time feature fusion framework named ProxyFusion is proposed, which uses learnable proxy vectors to select sparse expert networks for weighted aggregation of facial feature sets, thereby generating robust facial templates.
Prune and Repaint: Content-Aware Image Retargeting for any Ratio
Feihong Shen (Southeast University), Hao Chen (Southeast University)
Image TranslationGenerationDiffusion modelImage
🎯 What it does: A content-aware image relocalization method named PruneRepaint is proposed, which combines hierarchical semantic-guided seam carving and an adaptive repainting module to achieve image retargeting at arbitrary ratios.
Pruning neural network models for gene regulatory dynamics using data and domain knowledge
Intekhab Hossain (Harvard T.H. Chan School of Public Health), John Quackenbush (Harvard T.H. Chan School of Public Health)
OptimizationExplainability and InterpretabilityBiomedical DataOrdinary Differential Equation
🎯 What it does: A domain knowledge-based network pruning framework called DASH is proposed for inferring sparse neural network models of gene regulatory networks.
Pseudo-Private Data Guided Model Inversion Attacks
Xiong Peng (Hong Kong Baptist University), Mingyuan Zhou (University of Texas at Austin)
GenerationAdversarial AttackConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: The research proposes a model inversion attack guided by pseudo-private data (PPDG-MI), which enhances the attack effectiveness by dynamically fine-tuning the generator's prior distribution.
Pseudo-Siamese Blind-spot Transformers for Self-Supervised Real-World Denoising
Yuhui Quan (South China University of Technology), Hui Ji (National University of Singapore)
RestorationTransformerImage
🎯 What it does: In this paper, the authors propose a Transformer-based self-supervised denoising method for real scene images, utilizing a blind spot mechanism to train the network using only noisy images.
PSL: Rethinking and Improving Softmax Loss from Pairwise Perspective for Recommendation
Weiqin Yang (Zhejiang University), Can Wang (Zhejiang University)
Recommendation SystemTabular
🎯 What it does: Proposed Pairwise Softmax Loss (PSL) to improve traditional Softmax Loss and achieve a more compact DCG approximation in recommendation systems.
PTQ4DiT: Post-training Quantization for Diffusion Transformers
Junyi Wu (University of Illinois Chicago), Yan Yan (University of Illinois Chicago)
GenerationCompressionTransformerDiffusion modelImage
🎯 What it does: A post-training quantization (PTQ) method called PTQ4DiT is proposed for Diffusion Transformers (DiTs), which can compress the model to 8-bit weights/activations while maintaining generation quality, and further supports 4-bit weight quantization.
Public-data Assisted Private Stochastic Optimization: Power and Limitations
Enayat Ullah (Meta), Raman Arora (Johns Hopkins University)
OptimizationSafty and Privacy
🎯 What it does: This study investigates the theoretical limits and feasible methods of Public Data-Assisted Differential Privacy (PA-DP) in Stochastic Convex Optimization (SCO) and supervised learning, clarifying the performance bottlenecks and breakthrough points in the cases of complete/labelled public data and unlabelled public data.
PuLID: Pure and Lightning ID Customization via Contrastive Alignment
Zinan Guo (ByteDance Inc), Qian HE
RecognitionGenerationData SynthesisDiffusion modelContrastive LearningImageText
🎯 What it does: This paper proposes PuLID, a tuning-free method for identity (ID) customization in text-to-image (T2I) diffusion models.
Pure Message Passing Can Estimate Common Neighbor for Link Prediction
Kaiwen Dong (University of Notre Dame), Nitesh V Chawla
Graph Neural NetworkGraph
🎯 What it does: This study investigates whether Message Passing Neural Networks (MPNN) can capture link structural features, and based on this, proposes the MPLP (Message Passing Link Predictor) model to estimate and predict potential edges in a graph.
PURE: Prompt Evolution with Graph ODE for Out-of-distribution Fluid Dynamics Modeling
Hao Wu (University of Science and Technology of China), Xiao Luo (University of California)
Domain AdaptationOptimizationGraph Neural NetworkTransformerPrompt EngineeringTime SeriesPhysics RelatedOrdinary Differential Equation
🎯 What it does: This study investigates the out-of-distribution fluid dynamics modeling problem and proposes a framework called PURE (Prompt Evolution with Graph ODE) to enhance the model's predictive performance under parameter and temporal distribution shifts.
PureGen: Universal Data Purification for Train-Time Poison Defense via Generative Model Dynamics
Omead Pooladzandi (California Institute of Technology), Gregory Pottie (University of California)
Adversarial AttackData-Centric LearningDiffusion modelImage
🎯 What it does: This paper proposes PUREGEN, a dynamic random preprocessing method that utilizes Energy-Based Models (EBM) and Diffusion Models (DDPM) for unsupervised purification of data during the training phase, thereby defending against backdoor and triggerless training-time data poisoning attacks.
Putting Gale & Shapley to Work: Guaranteeing Stability Through Learning
Hadi Hosseini (Penn State University), Duohan Zhang (Penn State University)
Reinforcement LearningAgentic AI
🎯 What it does: In a two-sided matching market, we learn unknown preferences and propose algorithms that can guarantee the stability of the final matching.
PV-Tuning: Beyond Straight-Through Estimation for Extreme LLM Compression
Vladimir Malinovskii (Yandex), Peter Richtárik
CompressionOptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes PV-Tuning, a fine-tuning framework for large language models with extremely low bit quantization (1-2 bits), which can simultaneously update continuous parameters (such as scaling factors and codebooks) and discrete parameters (weight quantization codes), significantly improving the accuracy of the quantized model.
Q-Distribution guided Q-learning for offline reinforcement learning: Uncertainty penalized Q-value via consistency model
Jing Zhang (Hong Kong University of Science and Technology), Bingyi Jing
Reinforcement LearningTabularBenchmark
🎯 What it does: This paper proposes an offline reinforcement learning method called Q-Distribution Guided Q-Learning (QDQ), which estimates the uncertainty of the Q value distribution using a consistency model and applies a lazy penalty to out-of-distribution (OOD) actions to balance excessive pessimism and exploration.
Q-VLM: Post-training Quantization for Large Vision-Language Models
Changyuan Wang (Tsinghua University), Jiwen Lu (Tsinghua University)
CompressionComputational EfficiencyTransformerMixture of ExpertsVision Language ModelMultimodality
🎯 What it does: A post-training quantization framework Q-VLM is proposed for efficient multimodal inference of large visual language models, reducing model size and inference speed.
QBB: Quantization with Binary Bases for LLMs
Adrian Bulat (Samsung AI), Georgios Tzimiropoulos (Queen Mary University of London)
CompressionOptimizationKnowledge DistillationTransformerLarge Language ModelText
🎯 What it does: This paper proposes a new low-bit quantization method called Quantization with Binary Bases (QBB), which decomposes weights into several 1-bit binary matrices and scaling vectors, nearly eliminating multiplication operations and retaining only summation calculations.
QGFN: Controllable Greediness with Action Values
Elaine Lau (McGill University), Emmanuel Bengio (Mila - Quebec AI Institute)
GenerationDrug DiscoveryReinforcement LearningSequential
🎯 What it does: This paper proposes a QGFN method that combines GFlowNet with the action value function Q, achieving a controllable greedy strategy during sampling through an adjustable mixing parameter p, thereby improving the generation rate of high-reward samples while maintaining diversity.
QKFormer: Hierarchical Spiking Transformer using Q-K Attention
Chenlin Zhou (Pengcheng Laboratory), Yonghong Tian (Pengcheng Laboratory)
ClassificationSpiking Neural NetworkTransformerImage
🎯 What it does: A directly trained hierarchical spiking Transformer called QKFormer is designed, and a linear complexity Q-K attention mechanism along with a cross-scale SPEDS embedding module is proposed, aiming to enhance the performance and energy efficiency of Spiking Neural Networks (SNNs);
QT-ViT: Improving Linear Attention in ViT with Quadratic Taylor Expansion
Yixing Xu (Advanced Micro Devices), Emad Barsoum (Advanced Micro Devices)
ClassificationObject DetectionSegmentationTransformerImage
🎯 What it does: A linear attention mechanism utilizing second-order Taylor expansion and Kronecker product is proposed to construct the QT-ViT model, aiming to reduce the time complexity of ViT's self-attention.
QTIP: Quantization with Trellises and Incoherence Processing
Albert Tseng (Cornell University), Christopher De Sa (Cornell University)
CompressionTransformerLarge Language ModelText
🎯 What it does: This paper proposes QTIP, a post-training weight quantization method based on Tree Code Quantization (TCQ), which can achieve ultra-high dimensional (greater than 100) low-precision (2–4 bit) compression while maintaining high-speed inference.