ICML 2025 Papers — Page 22
International Conference on Machine Learning · 3257 papers
Optimal Information Retention for Time-Series Explanations
Jinghang Yue (Beijing Jiaotong University), Youfang Lin (Beijing Jiaotong University)
OptimizationExplainability and InterpretabilityContrastive LearningTime SeriesElectrocardiogram
🎯 What it does: This paper proposes an optimal information retention principle based on information theory and implements the ORTE framework for generating precise local explanations of temporal models.
Optimal Sensor Scheduling and Selection for Continuous-Discrete Kalman Filtering with Auxiliary Dynamics
Mohamad Al Ahdab (Aalborg University), Zheng-Hua Tan (Aalborg University)
OptimizationRobotic IntelligenceGaussian SplattingSimultaneous Localization and MappingTime SeriesSequential
🎯 What it does: Under the continuous-discrete Kalman filtering framework, this study investigates the optimal sensor scheduling and selection problem coupled with auxiliary dynamics (such as energy, radiation damage, position, etc.) in relation to the measurement rate of multiple sensors. It models the measurement arrival using a Poisson process, derives an upper bound for the mean posterior covariance, constructs a finite time domain optimal control problem, and provides a deterministic measurement timing quantization method based on Wasserstein distance.
Optimal Survey Design for Private Mean Estimation
Yu-Wei Chen (Purdue University), Jordan Awan (Purdue University)
OptimizationSafty and PrivacyTabular
🎯 What it does: A hierarchical sampling scheme is designed to minimize variance under the differential privacy (DP) framework for private mean estimation.
Optimal Task Order for Continual Learning of Multiple Tasks
Ziyan Li (Washington University in St Louis), Naoki Hiratani (Washington University in St Louis)
OptimizationImage
🎯 What it does: This paper studies the impact of task order on continual learning and proposes two actionable task ordering principles: periphery-to-core and max-path.
Optimal Transfer Learning for Missing Not-at-Random Matrix Completion
Akhil Jalan (University of Texas at Austin), Purnamrita Sarkar (University of Texas at Austin)
Domain AdaptationOptimizationBiomedical Data
🎯 What it does: The study utilizes source matrices for transfer learning to complete low-rank target matrices under Missing Not-at-Random (MNAR) conditions.
Optimal Transport Barycenter via Nonconvex-Concave Minimax Optimization
Kaheon Kim (University of Notre Dame), Xiaohui Chen (University of Southern California)
OptimizationImage
🎯 What it does: A new Wasserstein-Descent ˙H¹-Ascent (WDHA) algorithm is proposed for the computation of unregularized Wasserstein barycenters of high-resolution multidimensional probability distributions.
Optimal transport-based conformal prediction
Gauthier Thurin (Ecole Normale Superieure), Claire Boyer (Universite Paris Saclay)
ClassificationOptimizationTabular
🎯 What it does: This paper proposes an optimal transport-based contractible prediction framework (OT-CP) that can construct shape-flexible prediction regions satisfying distribution-independent coverage using multidimensional inconsistent scores, and presents an adaptive version OT-CP+ that achieves approximate conditional coverage.
Optimistic Algorithms for Adaptive Estimation of the Average Treatment Effect
Ojash Neopane (Carnegie Mellon University), Aarti Singh (Carnegie Mellon University)
OptimizationReinforcement LearningTabular
🎯 What it does: This paper proposes an Optimistic Policy Tracking (OPT) algorithm for estimating the Average Treatment Effect (ATE) based on optimistic estimation.
Optimization for Neural Operators can Benefit from Width
Pedro Cisneros-Velarde (VMware Research), Arindam Banerjee (University of Illinois Urbana-Champaign)
Optimization
🎯 What it does: This paper studies the gradient descent optimization convergence of Deep Operator Networks (DON) and Fourier Neural Operators (FNO), providing theoretical guarantees for the convergence of width enhancement.
Optimization over Sparse Support-Preserving Sets: Two-Step Projection with Global Optimality Guarantees
William de Vazelhes (GenBio AI), Bin Gu (Jilin University)
Optimization
🎯 What it does: The paper proposes an iterative hard thresholding algorithm for sparse optimization that can operate under additional support-keeping constraints (such as ℓp constraints, box constraints, etc.) and provides theoretical guarantees for global convergence (convergence of risk values).
Optimization Proxies using Limited Labeled Data and Training Time -- A Semi-Supervised Bayesian Neural Network Approach
Parikshit Pareek (Indian Institute of Technology Roorkee), Deepjyoti Deka (MIT Energy Initiative)
OptimizationTabular
🎯 What it does: The study proposes a semi-supervised Bayesian neural network (Sandwich BNN) as an optimization surrogate, which quickly approaches the solution of constrained optimization problems using a small amount of labeled data and limited training time.
Optimizing Adaptive Attacks against Watermarks for Language Models
Abdulrahman Diaa (University of Waterloo), Nils Lukas (Mohammed Bin Zayed University of Artificial Intelligence)
OptimizationAdversarial AttackReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper addresses the robustness of text watermarking in large language models (LLMs) by proposing the construction of a preference dataset using observable watermark algorithms and proxy models. It employs reinforcement learning (such as DPO) to adaptively tune the rewriting model with open weights, enabling it to efficiently evade watermark detection while maintaining text quality.
Optimizing Language Models for Inference Time Objectives using Reinforcement Learning
Yunhao Tang (Meta GenAI), Remi Munos (Meta FAIR)
OptimizationLarge Language ModelReinforcement LearningText
🎯 What it does: The study explicitly optimizes inference time objectives (pass@k and majority voting) during the training of language models and implements it through reinforcement learning.
Optimizing Large Language Model Training Using FP4 Quantization
Ruizhe Wang (University of Science and Technology of China), Peng CHENG
OptimizationTransformerLarge Language ModelText
🎯 What it does: The first FP4 low-precision training framework for large language models (LLM) is proposed, demonstrating the feasibility of training a 13B parameter model from scratch with 100B tokens.
Optimizing Noise Distributions for Differential Privacy
Atefeh Gilani (Arizona State University), Lalitha Sankar (Arizona State University)
OptimizationSafty and PrivacyTabularBiomedical Data
🎯 What it does: A unified optimization framework is proposed to design optimal noise that satisfies (ε,δ)-differential privacy under given error constraints and combination counts, targeting both continuous and discrete noise distributions.
Optimizing Robustness and Accuracy in Mixture of Experts: A Dual-Model Approach
Xu Zhang, Ren Wang (Illinois Institute of Technology)
OptimizationAdversarial AttackTransformerMixture of ExpertsImage
🎯 What it does: A method is proposed to simultaneously enhance robustness and natural accuracy in the Mixture of Experts (MoE) architecture, specifically through RT-ER training that strengthens the robustness of only the second highest weight expert, and the JTDMoE dual model scheme that combines standard MoE with robust MoE in a linear manner and trains them jointly.
Optimizing Social Network Interventions via Hypergradient-Based Recommender System Design
Marino Kühne (ETH Zurich), John Lygeros (ETH Zurich)
Recommendation SystemOptimizationGraph
🎯 What it does: In social networks, the balance state of the Friedkin-Johnsen opinion dynamics model is driven by adjusting the edge weights between users to minimize network polarization or inconsistency;
Optimizing Temperature for Language Models with Multi-Sample Inference
Weihua Du (Carnegie Mellon University), Sean Welleck (Carnegie Mellon University)
OptimizationTransformerLarge Language ModelText
🎯 What it does: This paper studies the temperature selection problem under multi-sample aggregation strategies and proposes an automatic temperature optimization method based on entropy inflection points.
Optimizing Test-Time Compute via Meta Reinforcement Finetuning
Yuxiao Qu (Carnegie Mellon University), Aviral Kumar (Carnegie Mellon University)
OptimizationSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This paper proposes viewing the optimization of computational load during inference as a meta-reinforcement learning problem, and designs a fine-tuning framework named MRT, which utilizes cumulative regret and progress rewards to achieve efficient inference.
OptMATH: A Scalable Bidirectional Data Synthesis Framework for Optimization Modeling
Hongliang Lu (Peking University), Zaiwen Wen (Peking University)
Data SynthesisOptimizationTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: A scalable bidirectional data synthesis framework, OptMATH, has been designed and implemented to generate high-quality optimization modeling datasets.
OR-Bench: An Over-Refusal Benchmark for Large Language Models
Justin Cui (University of California Los Angeles), Cho-Jui Hsieh (University of California Los Angeles)
TransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Proposed and implemented an automated generation of a large-scale 'over-rejection' benchmark dataset OR-Bench, including 80K prompts, 1K difficult problems, and 600 toxic prompts.
Oracle-MoE: Locality-preserving Routing in the Oracle Space for Memory-constrained Large Language Model Inference
Jixian Zhou (Fudan University), Li Shang (Fudan University)
TransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: Proposes the Oracle-MoE architecture, which significantly reduces the latency caused by expert switching by implementing routing in the oracle space.
OrcaLoca: An LLM Agent Framework for Software Issue Localization
Zhongming Yu (University of California San Diego), Jishen Zhao (University of California San Diego)
AI Code AssistantTransformerLarge Language ModelAgentic AIText
🎯 What it does: OrcaLoca is proposed, an automated software defect localization framework that combines LLM and code search, improving the navigation and localization accuracy of code repositories.
Organize the Web: Constructing Domains Enhances Pre-Training Data Curation
Alexander Wettig (Princeton University), Luca Soldaini (Allen Institute for Artificial Intelligence)
OptimizationKnowledge DistillationData-Centric LearningTransformerLarge Language ModelText
🎯 What it does: By constructing a two-dimensional domain of topics and formats to annotate and organize large-scale web corpora, and utilizing domain mixing to optimize pre-training data, the performance of downstream tasks is improved.
Orient Anything: Learning Robust Object Orientation Estimation from Rendering 3D Models
Zehan Wang (Zhejiang University), Zhou Zhao (Zhejiang University)
Object DetectionPose EstimationDomain AdaptationVision Language ModelImageMesh
🎯 What it does: A method for object orientation estimation based on 3D model rendering is proposed, and a large-scale rendered dataset with orientation labels is constructed.
Origin Identification for Text-Guided Image-to-Image Diffusion Models
Wenhao Wang (University of Technology Sydney), Yi Yang (Zhejiang University)
GenerationRetrievalDiffusion modelAuto EncoderImageText
🎯 What it does: This paper proposes a new task - identifying the original images from text-guided image-to-image diffusion model generated query images, and constructs the first publicly available dataset OriPID for this task, while also proposing a linear transformation-based VAE embedding method to achieve efficient and generalizable original image retrieval.
Orthogonal Subspace Decomposition for Generalizable AI-Generated Image Detection
Zhiyuan Yan (Peking University), Li Yuan (Peking University)
ClassificationRecognitionSupervised Fine-TuningImage
🎯 What it does: A detection method for AIGI based on orthogonal subspace decomposition is proposed, utilizing high-order features of a pre-trained visual model and retaining its knowledge to enhance generalization ability;
OrthoRank: Token Selection via Sink Token Orthogonality for Efficient LLM inference
Seungjun Shin (Samsung Advanced Institute of Technology), Dokwan Oh (Samsung Advanced Institute of Technology)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper proposes a dynamic token selection method based on the orthogonality of 'sink tokens' (OrthoRank), which only updates the tokens with the minimum orthogonality to the sink token during the LLM inference process, significantly reducing computational load.
Orthus: Autoregressive Interleaved Image-Text Generation with Modality-Specific Heads
Siqi Kou (Shanghai Jiao Tong University), Zhijie Deng (Shanghai Jiao Tong University)
GenerationTransformerVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: Orthus is proposed, a unified multimodal model capable of autoregressively generating images and text.
Oscillation-Reduced MXFP4 Training for Vision Transformers
Yuxiang Chen (Tsinghua University), Jianfei Chen (Tsinghua University)
ClassificationOptimizationTransformerImage
🎯 What it does: This paper proposes a visual Transformer pre-training method called TetraJet based on the low-precision format MXFP4, addressing the weight oscillation problem caused by forward quantization;
OTTER: A Vision-Language-Action Model with Text-Aware Visual Feature Extraction
Huang Huang (University of California), Pieter Abbeel (University of California)
Robotic IntelligenceTransformerVision-Language-Action ModelMultimodality
🎯 What it does: This paper presents OTTER, a visual-language-action model that extracts text-aware visual features and freezes the pre-trained VLM.
Otter: Generating Tests from Issues to Validate SWE Patches
Toufique Ahmed (IBM Research), Martin Hirzel (IBM Research)
GenerationAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: The Otter system is proposed, which utilizes LLM combined with rule-driven analysis and self-reflective action planning to automatically generate fail-to-pass tests based on issue descriptions and old code before receiving a fix patch, supporting TDD and automated repair agents; and has released the TDD-Bench-Verified benchmark.
Outlier Gradient Analysis: Efficiently Identifying Detrimental Training Samples for Deep Learning Models
Anshuman Chhabra (University of South Florida), Hongfu Liu (Brandeis University)
Anomaly DetectionComputational EfficiencyImage
🎯 What it does: This paper proposes a new method for anomaly detection in gradient space, aimed at identifying and removing samples that negatively impact the training of deep models.
Outlier-Aware Post-Training Quantization for Discrete Graph Diffusion Models
Zheng Gong (Hong Kong University of Science and Technology), Ying Sun (Hong Kong University of Science and Technology)
GenerationData SynthesisComputational EfficiencyGraph Neural NetworkDiffusion modelGraphBiomedical Data
🎯 What it does: This paper proposes a post-training quantization framework Bit-DGDM for discrete graph diffusion models (DGDM), aimed at significantly reducing model storage and computational costs while maintaining generation quality.
Outsourced Diffusion Sampling: Efficient Posterior Inference in Latent Spaces of Generative Models
Siddarth Venkatraman (Mila Quebec AI Institute), Nikolay Malkin (University of Edinburgh)
GenerationData SynthesisReinforcement Learning from Human FeedbackProtein Structure PredictionReinforcement LearningDiffusion modelImage
🎯 What it does: This paper proposes a technique for transferring posterior sampling from data space to the noise space of generative models, called Outsourced Diffusion Sampling, which achieves efficient posterior inference by training diffusion models to approximate the posterior distribution in the noise space.
OV-MER: Towards Open-Vocabulary Multimodal Emotion Recognition
Zheng Lian (Institute of Automation, Chinese Academy of Sciences), Jianhua Tao (Tsinghua University)
RecognitionTransformerLarge Language ModelVideoTextMultimodalityAudio
🎯 What it does: Proposed and implemented the Open Vocabulary Multimodal Emotion Recognition (OV-MER) paradigm, constructed the OV-MERD dataset, designed evaluation metrics based on emotion clustering, and conducted benchmark experiments on multimodal large language models.
Over-Tokenized Transformer: Vocabulary is Generally Worth Scaling
Hongzhi Huang (Bytedance), zhou Xun
TransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: This paper proposes the Over-Tokenized Transformer, which decouples the input and output vocabularies and enhances language modeling performance by using large-scale multiple n-gram embeddings at the input layer.
Overcoming Multi-step Complexity in Multimodal Theory-of-Mind Reasoning: A Scalable Bayesian Planner
Chunhui Zhang (Dartmouth College), Shao-Yuan Lo (Honda Research Institute USA)
TransformerLarge Language ModelVideoMultimodalityChain-of-Thought
🎯 What it does: A scalable Bayesian Theory-of-Mind planner is proposed, which employs stepwise Bayesian updates to decompose multi-step reasoning and transfers the ToM behavior of small language models to large models through a weak-to-strong control mechanism, enabling multimodal ToM reasoning.
Overcoming Non-monotonicity in Transducer-based Streaming Generation
Zhengrui Ma (Institute of Computing Technology Chinese Academy of Sciences), Min zhang
GenerationTransformerSupervised Fine-TuningAudio
🎯 What it does: Proposes MonoAttn-Transducer, which introduces learnable monotonic attention into the Transducer architecture, allowing the predictor to pay real-time attention to input history during streaming generation.
Overcoming Spurious Solutions in Semi-Dual Neural Optimal Transport: A Smoothing Approach for Learning the Optimal Transport Plan
Jaemoo Choi (Georgia Institute of Technology), Dohyun Kwon (Korea Institute for Advanced Study)
Image TranslationOptimizationGenerative Adversarial NetworkImage
🎯 What it does: By smoothing the source distribution and gradually denoising, the OTP model is proposed to learn the Optimal Transport Plan, thereby avoiding the pseudo-solution problem that arises in semi-dual neural OT (SNOT), and can correctly solve the case of one-to-many and non-deterministic OT Map;
Overcoming the Curse of Dimensionality in Reinforcement Learning Through Approximate Factorization
Chenbei Lu (Tsinghua University), Adam Wierman (California Institute of Technology)
Reinforcement Learning
🎯 What it does: By using approximate factorization to decompose high-dimensional MDPs into several low-dimensional subcomponents, we design synchronous sampling and model-based/non-model algorithms to address the curse of dimensionality in RL.
Overcoming Vocabulary Mismatch: Vocabulary-agnostic Teacher Guided Language Modeling
Haebin Shin (KAIST AI), Yeyun Gong (Microsoft Research)
Knowledge DistillationTransformerLarge Language ModelText
🎯 What it does: This study investigates how to effectively perform distillation and pre-training when there is a mismatch between the vocabulary of the teacher model and the student model.
Overestimation in LLM Evaluation: A Controlled Large-Scale Study on Data Contamination’s Impact on Machine Translation
Muhammed Yusuf Kocyigit (Boston University), Markus Freitag (Google)
TransformerLarge Language ModelText
🎯 What it does: Systematically evaluated the impact of data pollution (test set unintentionally contaminated by pre-training data) on large language models (1B, 8B) in machine translation tasks, using hierarchical pollution control and branching training strategies for experiments.
Overtrained Language Models Are Harder to Fine-Tune
Jacob Mitchell Springer (Carnegie Mellon University), Aditi Raghunathan (Carnegie Mellon University)
Large Language ModelSupervised Fine-TuningText
🎯 What it does: This paper compares the performance of language models with different numbers of pre-trained tokens after fine-tuning, finding that longer pre-training may actually lead to a decline in performance after fine-tuning, introducing the phenomenon of 'catastrophic overtraining.'
OW-VAP: Visual Attribute Parsing for Open World Object Detection
Xing Xi (South China University of Technology), Ronghua Luo (South China University of Technology)
Object DetectionConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: A framework named OW-VAP is proposed, which utilizes a Visual Attribute Parser (VAP) to detect unknown objects without relying on large language models, and suppresses background noise conflicts through Probabilistic Soft Label Assignment (PSLA).
OWLS: Scaling Laws for Multilingual Speech Recognition and Translation Models
William Chen (Carnegie Mellon University), Shinji Watanabe (Carnegie Mellon University)
RecognitionTransformerPrompt EngineeringAudio
🎯 What it does: A series of OWLS multilingual speech models has been constructed and made public, ranging from 0.25B to 18B parameters, covering 150 languages, and supporting speech recognition and translation.
P(all-atom) Is Unlocking New Path For Protein Design
Wei Qu (Institute of Science and Technology for Brain Inspired Intelligence), Haobo Wang (Levinthal Biotechnology)
Protein Structure PredictionDiffusion modelBiomedical Data
🎯 What it does: The Pallatom model is proposed, which directly learns the joint distribution P(structure, sequence) of all-atom coordinates, enabling the one-time generation of protein sequences and their complete atomic conformations.
PAC Learning with Improvements
Idan Attias (University of Illinois at Chicago), Matthew Walter (Toyota Technological Institute at Chicago)
Tabular
🎯 What it does: A new framework for PAC learning under the condition of feature improvement is proposed, exploring the impact of improvement capability on sample complexity and zero-error learning.
PAC-Bayes Analysis for Recalibration in Classification
Masahiro Fujisawa (Osaka University), Futoshi Futami (Osaka University)
ClassificationOptimizationConvolutional Neural NetworkImage
🎯 What it does: Theoretical analysis of the bias in non-parametric estimation of Expected Calibration Error (ECE) in multi-class scenarios, and providing an optimizable generalization error upper bound within the PAC-Bayes framework, leading to the design of an adaptive recalibration algorithm (PBR) based on this upper bound.
Pairwise Maximum Likelihood For Multi-Class Logistic Regression Model With Multiple Rare Classes
Xuetong Li (Xi'an Jiaotong University), Hansheng Wang (Peking University)
ClassificationOptimizationComputational EfficiencyImage
🎯 What it does: This paper studies the situation in multi-class logistic regression where there is one main class and several rare classes. It proposes estimation methods based on Pairwise Maximum Likelihood Estimation (PMLE) and Subsample Pairwise Maximum Likelihood Estimation (SPMLE), aiming to address the computational difficulties of traditional Global Maximum Likelihood Estimation (GMLE) in high-dimensional, large-class imbalanced data.
PAK-UCB Contextual Bandit: An Online Learning Approach to Prompt-Aware Selection of Generative Models and LLMs
Xiaoyan Hu (Chinese University of Hong Kong), Farzan Farnia (Chinese University of Hong Kong)
GenerationRecommendation SystemTransformerLarge Language ModelReinforcement LearningPrompt EngineeringImageVideoText
🎯 What it does: An online learning framework for dynamically selecting generative models based on prompts, PAK-UCB and RFF-UCB, is proposed, which uses contextual bandits to adaptively select the best text/image/video generative model or LLM.
PANDAS: Improving Many-shot Jailbreaking via Positive Affirmation, Negative Demonstration, and Adaptive Sampling
Avery Ma (University of Toronto), Amir-massoud Farahmand (Polytechnique Montreal)
TransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: The PANDAS method is proposed, which significantly improves the success rate of jailbreaks by incorporating Positive Affirmation (PA), Negative Demonstration (ND), and Adaptive Sampling (AS) targeting prompt themes in multi-shot jailbreaks.
PaperBench: Evaluating AI’s Ability to Replicate AI Research
Giulio Starace (OpenAI), Tejal Patwardhan (OpenAI)
Large Language ModelAgentic AITextBenchmark
🎯 What it does: Create the PaperBench benchmark to evaluate the ability of AI agents to replicate ICML 2024 papers from scratch.
Parallel Simulation for Log-concave Sampling and Score-based Diffusion Models
Huanjian Zhou (University of Tokyo), Masashi Sugiyama (RIKEN)
GenerationData SynthesisComputational EfficiencyDiffusion modelScore-based ModelImageVideoAudio
🎯 What it does: A new parallel sampling method is proposed, aimed at improving the sampling efficiency for high-dimensional probability distributions, particularly for log-concave sampling and score-based diffusion models.
ParallelComp: Parallel Long-Context Compressor for Length Extrapolation
Jing Xiong (University of Hong Kong), Ngai Wong (University of Hong Kong)
RetrievalCompressionTransformerLarge Language ModelText
🎯 What it does: This paper presents PARALLELCOMP, a training-free parallel long-context compression method that enables an 8B language model to extend the context length from 8K to 128K on a single A100 80GB GPU, addressing memory bottlenecks and attention sink issues.
Parameter-Efficient Fine-Tuning of State Space Models
Kevin Galim (FuriosaAI), Kangwook Lee (University of Wisconsin-Madison)
CompressionOptimizationRecurrent Neural NetworkSupervised Fine-TuningImageText
🎯 What it does: This paper studies parameter-efficient fine-tuning (PEFT) methods on state space models (SSM) and proposes a specialized Sparse Dimension Tuning (SDT) technique.
Parameters vs FLOPs: Scaling Laws for Optimal Sparsity for Mixture-of-Experts Language Models
Samira Abnar (Apple), Vimal Thilak (Massachusetts Institute of Technology)
TransformerLarge Language ModelMixture of ExpertsTextChain-of-Thought
🎯 What it does: This study investigates the optimal trade-off between the number of parameters and the FLOPs (computational load) per sample in sparse Mixture-of-Experts (MoE) language models under a fixed training computational budget, and extracts a scaling law suitable for MoE from experimental data.
Parametric Scaling Law of Tuning Bias in Conformal Prediction
Hao Zeng (Southern University of Science and Technology), Hongxin Wei (Southern University of Science and Technology)
ClassificationOptimizationConvolutional Neural NetworkImage
🎯 What it does: This study investigates the tuning bias introduced by calibration in conformal prediction on the same dataset, providing both empirical and theoretical quantitative analysis.
Pareto Merging: Multi-Objective Optimization for Preference-Aware Model Merging
Weiyu Chen (Hong Kong University of Science and Technology), James Kwok
OptimizationTransformerImage
🎯 What it does: This paper proposes a Pareto Merging method based on multi-objective optimization, achieving adaptive model generation that considers different user preferences during multi-model merging.
Pareto-frontier Entropy Search with Variational Lower Bound Maximization
Masanori Ishikura (Nagoya Institute of Technology), Masayuki Karasuyama (Nagoya Institute of Technology)
OptimizationHyperparameter SearchTabular
🎯 What it does: A Pareto Front Entropy Search Method based on Variational Lower Bound Maximization (PFEV) is proposed for multi-objective Bayesian optimization.
Pareto-Optimal Fronts for Benchmarking Symbolic Regression Algorithms
Kei Sen Fong (National University of Singapore), Mehul Motani (National University of Singapore)
OptimizationTabularBenchmark
🎯 What it does: By using Gene Expression Programming (GEP) for exhaustive search on 34 black-box datasets from SRBench, an Absolute Pareto Front (APO front) was constructed, and a systematic comparison of the performance of 8 numerical optimization methods in obtaining the APO front was conducted; at the same time, general standards for benchmark evaluation and visualization were proposed.
Pareto-Optimality, Smoothness, and Stochasticity in Learning-Augmented One-Max-Search
Ziyad Benomar (CREST), Spyros Angelopoulos (CNRS and International Laboratory on Learning Systems)
OptimizationTime SeriesFinance Related
🎯 What it does: A deterministic threshold algorithm is proposed and analyzed for the One-Max search problem with learning enhancement, which simultaneously satisfies Pareto-optimality, smoothness, and robustness.
PARM: Multi-Objective Test-Time Alignment via Preference-Aware Autoregressive Reward Model
Baijiong Lin (Hong Kong University of Science and Technology), Ying-Cong Chen (Hong Kong University of Science and Technology)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: A single unified preference-aware autoregressive reward model (PARM) is proposed for aligning large language models during multi-objective testing, capable of dynamically controlling the generated results based on user-defined preference vectors.
PARQ: Piecewise-Affine Regularized Quantization
Lisa Jin (Meta), Lin Xiao (Meta)
CompressionOptimizationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: A quantization-aware training method based on convex piecewise affine regularization (PAR) called PARQ is proposed.
Parrot: Multilingual Visual Instruction Tuning
Hai-Long Sun (Nanjing University), Han-Jia Ye (Nanjing University)
Mixture of ExpertsVision Language ModelMultimodalityBenchmark
🎯 What it does: The PARROT model is proposed for multilingual visual instruction tuning, addressing the performance degradation of multimodal models in non-English languages.
Partially Observable Reinforcement Learning with Memory Traces
Onno Eberhard (Max Planck Institute for Intelligent Systems), Claire Vernade (University of Tübingen)
Reinforcement LearningSequential
🎯 What it does: Proposed memory traces as compressible historical features for partially observable reinforcement learning;
Partition First, Embed Later: Laplacian-Based Feature Partitioning for Refined Embedding and Visualization of High-Dimensional Data
Erez Peterfreund (Yale University), Boris Landa (Yale University)
Explainability and InterpretabilityComputational EfficiencyBiomedical Data
🎯 What it does: This paper proposes a framework that first partitions high-dimensional data features and then embeds them separately, which can extract corresponding low-dimensional structures from different feature subsets, enhancing visualization and interpretability.
PASS: Private Attributes Protection with Stochastic Data Substitution
Yizhuo Chen (University of Illinois Urbana-Champaign), Tarek F. Abdelzaher (University of Illinois Urbana-Champaign)
Safty and PrivacyAudio
🎯 What it does: Proposes the PASS (Private Attributes Protection with Stochastic Data Substitution) algorithm, which protects private attributes through random substitution while maintaining the useful attributes and general characteristics of the data.
Patch-wise Structural Loss for Time Series Forecasting
Dilfira Kudrat (Tianjin University), Qinghua Hu (Tianjin University)
Time Series
🎯 What it does: This paper proposes a loss function called PS loss based on a time series block-level structure, aimed at compensating for the traditional MSE's neglect of local structural information and improving multi-step prediction accuracy.
PatchPilot: A Cost-Efficient Software Engineering Agent with Early Attempts on Formal Verification
Hongwei Li (University of California, Santa Barbara), Wenbo Guo (University of California, Santa Barbara)
Computational EfficiencyAI Code AssistantTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: Designed and implemented a rule-based, cost-effective software patch generation agent called PatchPilot, which includes five components: reproduction, localization, generation, verification, and refinement.
PCEvolve: Private Contrastive Evolution for Synthetic Dataset Generation via Few-Shot Private Data and Generative APIs
Jianqing Zhang (Shanghai Jiao Tong University), Qiang Yang (Hong Kong Polytechnic University)
GenerationData SynthesisAnomaly DetectionDiffusion modelContrastive LearningImageBiomedical Data
🎯 What it does: This paper proposes a contrastive learning-based evolutionary algorithm called PCEvolve, which iteratively generates differentially private synthetic data using a small number of private images and a general generative API.
PDE-Controller: LLMs for Autoformalization and Reasoning of PDEs
Mauricio Soroco (Simon Fraser University), Wuyang Chen (Simon Fraser University)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextPhysics Related
🎯 What it does: Proposes the PDE-Controller framework, which utilizes large language models to automatically convert natural language PDE control tasks into STL specifications and generate executable Python code, subsequently enhancing control effectiveness through sub-goal reasoning.
PDE-Transformer: Efficient and Versatile Transformers for Physics Simulations
Benjamin Holzschuh (Technical University of Munich), Nils Thuerey (Technical University of Munich)
TransformerDiffusion modelTime SeriesPhysics Related
🎯 What it does: A Transformer-based PDE-Transformer network is proposed for efficiently approximating physical simulation results on two-dimensional regular grids, which can serve as a foundational model for transfer learning.
PDUDT: Provable Decentralized Unlearning under Dynamic Topologies
Jing Qiao (Shandong University), Dongxiao Yu (Shandong University)
Federated LearningImage
🎯 What it does: A provably decentralized model forgetting algorithm PDUDT under dynamic topology is proposed, allowing clients to completely eliminate the influence of specified clients without the need for additional communication or retraining.
PEAKS: Selecting Key Training Examples Incrementally via Prediction Error Anchored by Kernel Similarity
Mustafa Burak Gurbuz (Georgia Institute of Technology), Constantine Dovrolis (Georgia Institute of Technology)
Computational EfficiencyData-Centric LearningTransformerImage
🎯 What it does: Proposes an Incremental Data Selection (IDS) framework, which uses the PEAKS algorithm to online select the most valuable samples for model training in a streaming data environment, thereby improving the data efficiency of deep learning.
PEINR: A Physics-enhanced Implicit Neural Representation for High-Fidelity Flow Field Reconstruction
Liming Shen (National University of Defense Technology), Jie Liu (National University of Defense Technology)
TransformerTime SeriesBenchmarkPhysics Related
🎯 What it does: A physics-enhanced implicit neural representation (PEINR) framework is proposed for high-fidelity flow field reconstruction, along with the release of a large benchmark dataset (HFR-Bench) with a scale of up to 5.4 TB, containing 33,600 two-dimensional/three-dimensional vector fields.
Penalizing Infeasible Actions and Reward Scaling in Reinforcement Learning with Offline Data
Jeonghye Kim (Korea Advanced Institute of Science and Technology), Woohyung Lim (LG AI Research)
Reinforcement LearningTabular
🎯 What it does: This paper proposes a method that combines reward scaling and layer normalization (RS-LN) along with a penalty for infeasible actions (PA), resulting in an offline reinforcement learning algorithm named PARS. This algorithm effectively suppresses Q-value extrapolation errors outside of offline data and supports smooth transitions from offline to online learning.
PENCIL: Long Thoughts with Short Memory
Chenxiao Yang (Toyota Technological Institute at Chicago), Zhiyuan Li (Toyota Technological Institute at Chicago)
OptimizationComputational EfficiencyTransformerTextChain-of-Thought
🎯 What it does: PENCIL is proposed, a mechanism that incorporates recursively eliminable intermediate reasoning steps in chain-of-thought (CoT) reasoning, significantly compressing context length while maintaining reasoning depth.
PepTune: De Novo Generation of Therapeutic Peptides with Multi-Objective-Guided Discrete Diffusion
Sophia Tang (University of Pennsylvania), Pranam Chatterjee (University of Pennsylvania)
OptimizationDrug DiscoveryTransformerDiffusion modelTextBiomedical Data
🎯 What it does: PepTune is proposed, a multi-objective discrete diffusion model for generating and optimizing therapeutic peptide SMILES containing non-natural amino acids and cyclic modifications.
Perception in Reflection
Yana Wei (Johns Hopkins University), Vishal M. Patel (Johns Hopkins University)
RecognitionGenerationOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningVision Language ModelImageText
🎯 What it does: This paper proposes a dual-model reflection mechanism called RePer and a reflection-aware learning method (RPL), which iteratively improves visual perception through multiple rounds of strategy-critic interactions, significantly reducing hallucinations and enhancing image description quality.
Perceptual-GS: Scene-adaptive Perceptual Densification for Gaussian Splatting
Hongbi Zhou (Tongji University), Zhangkai Ni (Tongji University)
GenerationData SynthesisComputational EfficiencyGaussian SplattingPoint Cloud
🎯 What it does: This paper proposes the Perceptual-GS framework, which integrates human visual perception sensitivity into 3D Gaussian Splatting, achieving high-quality viewpoint synthesis with a limited number of Gaussians.
Perceptually Constrained Precipitation Nowcasting Model
Wenzhi Feng (Harbin Institute of Technology), Yaowei Wang (Harbin Institute of Technology)
Convolutional Neural NetworkRecurrent Neural NetworkRectified FlowImageTime Series
🎯 What it does: This paper proposes a precipitation nowcasting model based on perceptual constraints (PercpCast), which first estimates the posterior mean sequence using ConvLSTM, and then maps it to the true distribution using a rectified flow with frame sampling, thereby reducing mean square error while preserving details.
Peri-LN: Revisiting Normalization Layer in the Transformer Architecture
Jeonghoon Kim (NAVERCloud), Kang Min Yoo (NAVER)
TransformerLarge Language ModelText
🎯 What it does: This paper proposes and systematically evaluates a new normalization method for Transformer layers—Peri-LN, which uses LayerNorm or RMSNorm on both the input and output of each sub-layer simultaneously.
Peripheral Memory for LLMs: Integration of Sequential Memory Banks with Adaptive Querying
Songlin Zhai (Southeast University), Guilin Qi (Southeast University)
RetrievalOptimizationTransformerLarge Language ModelText
🎯 What it does: A Peripheral Memory framework is proposed, decoupling memory modules from large language models (LLMs) to achieve dynamic querying and writing.
Permutation Equivariant Neural Networks for Symmetric Tensors
Edward Pearce-Crump (Imperial College London)
🎯 What it does: A complete description of linear permutation invariant functions for symmetric tensors is provided, along with two bases (orbital basis and graphical basis) and a mapping labeling method that does not require explicit memory storage for weights. Subsequently, its data efficiency and generalization ability are validated on two synthetic tasks.
Permutation-based Rank Test in the Presence of Discretization and Application in Causal Discovery with Mixed Data
Xinshuai Dong (Carnegie Mellon University), Kun Zhang (Carnegie Mellon University)
Tabular
🎯 What it does: A mixed data permutation rank test (MPRT) is proposed to test the rank of the cross-covariance matrix in the presence of discretized variables, thereby supporting causal discovery.
Permutation-Free High-Order Interaction Tests
Zhaolu Liu (Imperial College London), Mauricio Barahona (Imperial College London)
Anomaly DetectionOptimizationComputational EfficiencyData-Centric LearningTabularTime SeriesFinance Related
🎯 What it does: A series of higher-order interaction test methods (xdHSIC, xLI, xSI) that do not require permutation have been proposed, achieving efficient tests for joint independence, Lancaster, and Streitberg interactions.
Persistent Topological Features in Large Language Models
Yuri Gardinazzi (Area Science Park), Matteo Biagetti (Area Science Park)
TransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Using Zigzag persistence (topological data analysis) to track the point clouds of internal representations of large language models (LLMs) across different layers, and based on this, proposing two statistical descriptors (relative birth frequency and inter-layer persistence) to characterize the dynamic evolution of prompts in the representation space.
PertEval-scFM: Benchmarking Single-Cell Foundation Models for Perturbation Effect Prediction
Aaron Wenteler (Queen Mary University of London), Amaya Gallagher-Syed
Drug DiscoveryBiomedical DataBenchmark
🎯 What it does: This paper proposes the PertEval-scFM framework for evaluating the information content of single-cell foundation models (scFM) in predicting gene perturbation effects under zero-shot settings.
Pessimism Principle Can Be Effective: Towards a Framework for Zero-Shot Transfer Reinforcement Learning
Chi Zhang (University of Central Florida), Yue Wang (University of Central Florida)
Domain AdaptationRobotic IntelligenceReinforcement LearningSequential
🎯 What it does: This study investigates zero-shot transfer reinforcement learning without target domain data and proposes a framework based on the pessimistic principle.
PF3plat: Pose-Free Feed-Forward 3D Gaussian Splatting for Novel View Synthesis
Sunghwan Hong (KAIST AI), Seungryong Kim (KAIST AI)
GenerationData SynthesisDepth EstimationTransformerGaussian SplattingImage
🎯 What it does: A fast feedforward 3D Gaussian scattering framework PF3plat without pose information has been developed for synthesizing new views from sparse pose-free images.
Pfeife: Automatic Pipeline Parallelism for PyTorch
Ho Young Jhoo (Seoul National University), Nuno P. Lopes (INESC-ID Instituto Superior Técnico University of Lisbon)
OptimizationComputational EfficiencyTransformerLarge Language ModelDiffusion modelImageBenchmark
🎯 What it does: Developed Pfeife, an automatic pipeline parallel tool integrated with PyTorch 2, capable of transparently capturing the complete data flow graph of models and automatically partitioning and scheduling pipeline execution across multiple GPUs.
PhantomWiki: On-Demand Datasets for Reasoning and Retrieval Evaluation
Albert Gong (Cornell University), Kilian Q Weinberger
RetrievalTransformerLarge Language ModelTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: A demand-generated virtual Wikipedia-style dataset called PhantomWiki is proposed to evaluate the reasoning and retrieval capabilities of LLMs, avoiding data leakage and overfitting.
Phase and Amplitude-aware Prompting for Enhancing Adversarial Robustness
Yibo Xu (Xidian University), Nannan Wang (Xidian University)
Adversarial AttackConvolutional Neural NetworkTransformerPrompt EngineeringImage
🎯 What it does: A phase and amplitude-aware prompting method (PAP) based on frequency domain phase and amplitude spectra is proposed to enhance the robustness of models under adversarial attacks.
Phase transitions for the existence of unregularized M-estimators in single index models
Takuya Koriyama (University of Chicago), Pierre C Bellec
Tabular
🎯 What it does: This paper studies the phase transition of M-estimators without regularization in single-index models under high-dimensional proportional limits, providing the critical threshold δ∞ and proving its decisive role.
Physics Aware Neural Networks for Unsupervised Binding Energy Prediction
Ke Liu (Zhejiang University), Chunhua Shen (Zhejiang University)
Drug DiscoveryGraph Neural NetworkGraphBiomedical DataPhysics Related
🎯 What it does: An unsupervised protein-ligand binding energy prediction model, CEBind, has been developed. It applies random forces to ligands using physical laws (energy conservation and rigid body dynamics) to generate perturbations, and then predicts the energy difference before and after the perturbation using an energy model, thereby learning the binding energy.
Physics-Informed DeepONets for drift-diffusion on metric graphs: simulation and parameter identification
Jan Blechschmidt (TU Chemnitz), Jan-Frederik Pietschmann
OptimizationComputational EfficiencyData-Centric LearningGraph Neural NetworkSupervised Fine-TuningGraphPhysics Related
🎯 What it does: This paper proposes a physics-informed DeepONet (PI-DeepONet) framework for solving nonlinear drift-diffusion equations on metric graphs, and further applies it to parameter identification (inverse problems).
Physics-Informed Generative Modeling of Wireless Channels
Benedikt Böck (Technical University of Munich), Wolfgang Utschick (Technical University of Munich)
GenerationData SynthesisAuto EncoderGenerative Adversarial NetworkTabularPhysics Related
🎯 What it does: A wireless channel generation model based on physical information is proposed, utilizing Sparse Bayesian Generative Modeling (SBGM) and AmbientGAN to generate channel parameters and samples.
Physics-informed Temporal Alignment for Auto-regressive PDE Foundation Models
Congcong Zhu (University of Science and Technology of China), Jingrun Chen (University of Science and Technology of China)
Time SeriesSequentialPhysics Related
🎯 What it does: A physics-informed temporal alignment framework PITA is proposed to improve the autoregressive PDE-based model, addressing the issues of error accumulation and rapid learning.
Physics-Informed Weakly Supervised Learning For Interatomic Potentials
Makoto Takamoto (NEC Laboratories Europe), Mathias Niepert (University of Stuttgart)
OptimizationData-Centric LearningTabularPhysics Related
🎯 What it does: This paper studies a physical information weak supervision learning framework (PIWSL) aimed at improving the energy and force predictions of machine learning atomic potential models.