ICLR 2026 Papers — Page 34
International Conference on Learning Representations · 5356 papers
OpenPros: A Large-Scale Dataset for Limited View Prostate Ultrasound Computed Tomography
Hanchen Wang (University of North Carolina at Chapel Hill), Youzuo Lin (University of North Carolina at Chapel Hill)
SegmentationData SynthesisConvolutional Neural NetworkTransformerBiomedical DataUltrasoundBenchmark
🎯 What it does: Proposed a large-scale, anatomically realistic prostate limited-angle USCT dataset named OPENPROS, and conducted systematic benchmarking based on this dataset.
OpenThoughts: Data Recipes for Reasoning Models
Etash Kumar Guha (Stanford University), Ludwig Schmidt (Technische Universitaet Muenchen)
Data SynthesisKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Constructed and publicly released the OpenThoughts3 dataset with 1.2 million entries, and trained OpenThinker3-7B, achieving state-of-the-art (SOTA) results in the reasoning domain on public data.
Operationalizing Data Minimization for Privacy-Preserving LLM Prompting
Jijie Zhou (Northeastern University), Tianshi Li (Northeastern University)
OptimizationSafty and PrivacyData-Centric LearningTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose a framework based on priority queue tree search to automatically minimize data disclosure in LLM prompts, balancing privacy and functionality
Operator Learning with Domain Decomposition for Geometry Generalization in PDE Solving
Jianing Huang (Bosch (China) Investment Co., Ltd), Ze Cheng (Bosch (China) Investment Co., Ltd)
TransformerMeshPhysics Related
🎯 What it does: Propose a local-global framework based on domain decomposition (SNI) to address bottlenecks in geometric generalization and data efficiency in operator learning.
Operator Theory-Driven Autoformulation of MDPs for Control of Queueing Systems
Victor Baillet (University of Cambridge), Mihaela van der Schaar (University of Cambridge)
TransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Automatically convert natural language described queue system control problems into operator graph-based Markov Decision Process (MDP) Bellman equations, and identify the structural characteristics of the optimal strategy based on this.
OPPO: Accelerating PPO-based RLHF via Pipeline Overlap
Kaizhuo Yan (University of Illinois Urbana Champaign), Fan Lai (University of Illinois Urbana Champaign)
Reinforcement Learning from Human FeedbackTransformerReinforcement LearningText
🎯 What it does: Propose the OPPO framework, enhancing training efficiency by achieving pipeline overlapping during PPO-based RLHF training.
Opponent Shaping in LLM Agents
Marta Emili García Segura, Mirco Musolesi (University College London)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIPrompt Engineering
🎯 What it does: Studied opponent shaping of large language model agents in multi-agent environments and proposed the ShapeLLM method, validated through experiments in various games.
OPRIDE: Efficient Offline Preference-based Reinforcement Learning via In-Dataset Exploration
Yiqin Yang (Institute of Automation, Chinese Academy of Sciences), Chongjie Zhang (Washington University in St. Louis)
Computational EfficiencyRobotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningBenchmark
🎯 What it does: Proposed the OPRIDE algorithm to improve the query efficiency of offline preference reinforcement learning
Optimal Aggregation of LLM and PRM Signals for Efficient Test-Time Scaling
Peng Kuang (Zhejiang University), Haohan Wang (University Of Illinois Urbana Champaign)
OptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper studies how to more effectively integrate signals from large language models (LLM) and process reward models (PRM) during test-time scaling (TTS) to improve reasoning quality and reduce computational costs.
Optimal Brain Restoration for Joint Quantization and Sparsification of LLMs
Hang Guo (ETH Zurich), Yawei Li (ETH Zurich)
CompressionTransformerLarge Language ModelText
🎯 What it does: Proposed the Optimal Brain Restoration (OBR) framework, achieving joint compression of quantization and sparsification, and efficiently compressing large language models without requiring additional training.
Optimal Robust Subsidy Policies for Irrational Agent in Principal-Agent MDPs
Bowen Hu (Chengdu University of Technology), Yixin Tao (Shanghai University of Finance and Economics)
OptimizationReinforcement Learning
🎯 What it does: This paper studies how the principal influences the agent's actions through subsidies in a Markov Decision Process (MDP) to maximize the principal's cumulative expected reward, and designs robust subsidy strategies under the agent's non-rational behavior (bounded rationality).
Optimal Sparsity of Mixture-of-Experts Language Models for Reasoning Tasks
Taishi Nakamura (Institute of Science Tokyo), Rio Yokota (Institute of Science Tokyo)
OptimizationComputational EfficiencyTransformerLarge Language ModelReinforcement LearningMixture of ExpertsText
🎯 What it does: Study the impact of sparsity on memory and inference capabilities of Mixture-of-Experts (MoE) language models under a fixed computational budget, systematically investigating the relationship between total parameters, active parameters, top-k routing, and other factors on pre-training loss and downstream task accuracy.
Optimal transport unlocks end-to-end learning for single-molecule localization
Romain Seailles (Ecole normale superieure), Julien Mairal (Universite Grenoble Alpes)
OptimizationConvolutional Neural NetworkBiomedical Data
🎯 What it does: Proposed an end-to-end deep learning method based on optimal transport for single-molecule localization microscopy (SMLM), leveraging the optical system model through an iterative network to achieve detection and localization without NMS and with a single threshold;
Optimal Transport-Induced Samples against Out-of-Distribution Overconfidence
Keke Tang (Guangzhou University), Zhihong Tian (Guangzhou University)
Data SynthesisAnomaly DetectionAuto EncoderImage
🎯 What it does: The study generates semantically ambiguous OOD samples by leveraging the singular boundary of semi-discrete optimal transport, and reduces the model's overconfidence on OOD inputs during training using a confidence suppression loss.
OptimalThinkingBench: Evaluating Over and Underthinking in LLMs
Pranjal Aggarwal (FAIR at Meta), Swarnadeep Saha (FAIR at Meta)
Large Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Proposed a unified evaluation framework called OptimalThinkingBench, which can simultaneously measure over-thinking of LLMs on simple tasks and under-thinking on complex tasks, along with corresponding evaluation metrics;
Optimas: Optimizing Compound AI Systems with Globally Aligned Local Rewards
Shirley Wu (Stanford University), Jure Leskovec (Stanford University)
ClassificationRetrievalRecommendation SystemOptimizationTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBiomedical Data
🎯 What it does: Propose the OPTIMAS framework, achieving separated optimization for various heterogeneous configurations (prompts, hyperparameters, model parameters, etc.) by learning local reward functions aligned with global rewards for each component;
Optimistic Task Inference for Behavior Foundation Models
Thomas Rupf (ETH Zürich), Andreas Krause (ETH Zürich)
Reinforcement LearningBenchmark
🎯 What it does: Propose an online task inference framework OpTI-BFM, which actively collects reward data by interacting with the environment, forms confidence ellipsoids using success features (USFs) in the behavior foundation model (BFM) and least squares task embedding, and achieves rapid inference through optimistic UCB strategy;
Optimizer Choice Matters For The Emergence of Neural Collapse
Jim Zhao (University of Basel), Aurelien Lucchi (University of Basel)
OptimizationHyperparameter SearchConvolutional Neural NetworkImage
🎯 What it does: Studied the impact of optimizer and weight decay coupling methods on the occurrence of neural collapse in deep networks, proposed a new diagnostic metric NC0, and provided theoretical and experimental analysis;
Optimizing Agent Planning for Security and Autonomy
Aashish Kolluri (Microsoft), Santiago Zanella-Beguelin (Microsoft)
OptimizationSafty and PrivacyTransformerLarge Language ModelAgentic AIBenchmark
🎯 What it does: Proposes a secure agent PRUDENTIA based on information flow control to optimize the agent's autonomy and reduce human intervention.
Optimizing Canaries for Privacy Auditing with Metagradient Descent
Matteo Boglioni (ETH Zurich), Steven Wu
OptimizationSafty and PrivacyImage
🎯 What it does: Propose a black-box privacy auditing method that utilizes meta-gradient descent to optimize 'canary' samples;
Optimizing Data Augmentation through Bayesian Model Selection
Madi Matymov (KAUST), Maurizio Filippone (KAUST)
OptimizationData-Centric LearningConvolutional Neural NetworkTransformerImageText
🎯 What it does: This paper proposes the OPTIMA framework, which treats data augmentation (DA) parameters as latent variables in a Bayesian model, jointly optimizing model parameters and DA parameters through variational inference to achieve data-driven learning of DA.
Optimizing ID Consistency in Multimodal Large Models: Facial Restoration via Alignment, Entanglement, and Disentanglement
Yuran Dong (Wuhan University), Mang Ye (Wuhan University)
RestorationDiffusion modelImageMultimodality
🎯 What it does: Propose a training-free, plug-and-play framework called EditedID to achieve facial identity-consistent editing and restoration in multimodal large models (e.g., GPT-4o+, Flux.1 Kontext, etc.).
OptimSyn: Influence-Guided Rubrics Optimization for Synthetic Data Generation
Zhiting Fan (Zhejiang University), Zuozhu Liu (Zhejiang University)
Data SynthesisOptimizationData-Centric LearningLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmark
🎯 What it does: Developed a robust data generation framework based on optimizer-aware influence functions, which automatically learns generation rubrics through reinforcement learning to enhance the supervised fine-tuning of LLMs by leveraging training feedback from the target model.
OptMerge: Unifying Multimodal LLM Capabilities and Modalities via Model Merging
Yongxian Wei (Tsinghua University), Dacheng Tao (Nanyang Technological University)
Computational EfficiencyKnowledge DistillationMixture of ExpertsVision Language ModelImageVideoTextMultimodalityBenchmarkAudio
🎯 What it does: This paper proposes a model merging benchmark for multi-modal large language models (MLLMs) and introduces a novel data-agnostic static merging method called OptMerge, which can merge expert models from different tasks or modalities into a single more powerful and resource-efficient model;
OR-PRM: A Process Reward Model for Algorithmic Problem in Operations Research
Yilin Wang (Zhejiang University), Min Li (Central South University)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This paper proposes the first process reward model (OR-PRM) tailored for operations research (OR), and trains this model using a high-quality OR-ProcessQA dataset with step-by-step annotations;
Oracle-efficient Hybrid Learning with Constrained Adversaries
Princewill Okoroafor (Cornell University), Michael P. Kim (Cornell University)
OptimizationComputational Efficiency
🎯 What it does: This paper proposes a new hybrid online learning algorithm aimed at simultaneously achieving statistical optimality and computational efficiency, particularly in the presence of adversarial labels.
Orak: A Foundational Benchmark for Training and Evaluating LLM Agents on Diverse Video Games
Dongmin Park (KRAFTON), Jaewoong Cho (KRAFTON)
TransformerLarge Language ModelSupervised Fine-TuningAgentic AIVideoTextBenchmark
🎯 What it does: Proposed the Orak benchmark to evaluate LLM performance in 12 video games, providing a pluggable interface and fine-grained assessment dimensions.
Orbital Transformers for Predicting Wavefunctions in Time-Dependent Density Functional Theory
Xuan Zhang (Texas A&M University), Xiaofeng Qian (Texas A&M University)
Graph Neural NetworkTransformerGraphPhysics Related
🎯 What it does: Propose the OrbEvo model, which utilizes SO(2)-equivariant graph transformers to learn the evolution of electronic wave functions over time in real-time TDDFT, and can predict wave function coefficients, dipole moments, and absorption spectra.
ORCaS: Unsupervised Depth Completion via Occluded Region Completion as Supervision
Hyoungseob Park (Yale University), Alex Wong (Yale University)
Depth EstimationImageBenchmark
🎯 What it does: Proposed an unsupervised depth completion method called ORCaS, which learns an implicit representation of 3D scenes by predicting features in occluded regions, thereby improving dense depth estimation from single-view RGB and sparse depth;
OrchestrationBench: LLM-Driven Agentic Planning and Tool Use in Multi-Domain Scenarios
Aelim Ahn (Kakao Corp), Jihoon kang
TransformerLarge Language ModelAgentic AITextGraphBenchmark
🎯 What it does: Proposed OrchestrationBench, a multilingual (English-Korean) LLM orchestration evaluation benchmark targeting multi-domain, multi-step tasks;
OrderDP: A Theoretically Guaranteed Lossless Dynamic Data Pruning Framework
Chenhan Jin (Chinese University of Hong Kong), Tieyong Zeng (Beijing Normal-Hong Kong Baptist University)
ClassificationComputational EfficiencyConvolutional Neural NetworkTransformerImage
🎯 What it does: Proposed OrderDP, a dynamic data pruning framework based on random candidate batches and loss sorting, which can significantly reduce training costs while maintaining near-lossless performance;
ORION: Decoupling and Alignment for Unified Autoregressive Understanding and Generation
Taihang Hu (Nankai University), Yaxing Wang (Nankai University)
RecognitionGenerationTransformerLarge Language ModelVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: Designed and trained a fully automated regression multimodal large language model, ORION, achieving the unification of image understanding and generation, while alleviating semantic-structural conflicts through decoupling and alignment.
OrthAlign: Orthogonal Subspace Decomposition for Non-Interfering Multi-Objective Alignment
Liang Lin (China Telecom), Kun Wang (Nanyang Technological University)
Reinforcement Learning from Human FeedbackTransformerText
🎯 What it does: Propose OrthAlign, which addresses parameter conflicts in multi-objective preference alignment through orthogonal subspace decomposition, ensuring gradient updates for different objectives occur within mutually non-interfering subspaces.
OrthoRF: Exploring Orthogonality in Object-Centric Representations
Despoina Touska (University of Amsterdam), Pascal Cerfontaine (TH K?ln)
SegmentationExplainability and InterpretabilityRepresentation LearningAuto EncoderImage
🎯 What it does: Propose the OrthoRF autoencoder, which utilizes orthogonal constraints to achieve object segmentation without post-processing and recovers hidden parts in occluded regions.
OrthoSolver: A Neural Proper Orthogonal Decomposition Solver For PDEs
Ying Pang (Beihang University), Fanhao Mu (Beihang University)
BenchmarkPhysics Related
🎯 What it does: Proposed a nonlinear orthogonal decomposition framework called OrthoSolver based on maximizing mutual information for efficiently solving PDEs.
OSCAR: Online Soft Compression for RAG
Maxime Louis (Naver Labs Europe), Stéphane Clinchant (Naver Labs Europe)
RetrievalCompressionComputational EfficiencyKnowledge DistillationTransformerTextRetrieval-Augmented Generation
🎯 What it does: Propose and implement OSCAR — an online soft compression method that dynamically compresses retrieved documents into query-related embedding representations within retrieval-augmented generation (RAG), significantly accelerating inference while maintaining answer quality.
OSIRIS: Bridging Analog Circuit Design and Machine Learning with Scalable Dataset Generation
Giuseppe Chiari (Politecnico di Milano), Davide Zoni (Politecnico di Milano)
GenerationData SynthesisOptimizationLarge Language ModelSupervised Fine-TuningReinforcement LearningImagePhysics Related
🎯 What it does: Built a complete backend workflow OSIRIS for large-scale generation of DRC and LVS compliant analog IC layouts, and implemented a baseline for iterative layout optimization based on reinforcement learning.
OSWorld-MCP: Benchmarking MCP Tool Invocation In Computer-Use Agents
Hongrui Jia (Peking University), Fei Huang (Tongyi Lab, Alibaba Group)
AI Code AssistantLarge Language ModelVision Language ModelVision-Language-Action ModelMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Established the OSWorld-MCP benchmark, integrating 158 high-quality MCP tools with GUI operations to evaluate the tool calling and decision-making capabilities of multimodal models in real computer environments.
Otters: An Energy-Efficient Spiking Transformer via Optical Time-to-First-Spike Encoding
Zhanglu Yan (National University of Singapore), Weng-Fai Wong (Westlake University)
ClassificationComputational EfficiencySpiking Neural NetworkTransformerTextBenchmark
🎯 What it does: Developed a Time-to-First-Spike (TTFS) Spiking Transformer model called Otters based on optoelectronic hetero-synapses, achieving significant energy efficiency improvements through hardware-software co-design.
Out of the Memory Barrier: A Highly Memory-Efficient Training System for LLMs with Million-Token Contexts
Wenhao Li (Xiamen University), Rongrong Ji (Xiamen University)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Propose a long-context LLM training framework named OOMB, which adopts block recursive training and activation recomputation to significantly reduce activation memory usage;
Out of the Shadows: Exploring a Latent Space for Neural Network Verification
Lukas Koller (Technical University of Munich), Matthias Althoff (Technical University of Munich)
Safty and PrivacyExplainability and InterpretabilityComputational EfficiencyBenchmark
🎯 What it does: Propose a specification-driven input refinement method that leverages the potential space of projection-based set representations (e.g., zonotopes) to backward map unsafe output sets to the input space, rapidly pruning known safe input regions during branching and bounding processes, significantly reducing the number of subproblems.
Out-of-Distribution Graph Models Merging
Yidi Wang (Great Bay University), Xiao Luo (University of Wisconsin-Madison)
Domain AdaptationGraph Neural NetworkMixture of ExpertsGraph
🎯 What it does: Studied how to merge graph neural network (GNN) models pre-trained on different domains without using source domain data, constructing a unified model capable of generalizing to new domains.
Output Supervision Can Obfuscate the Chain of Thought
jacob drori, Alexander Matt Turner
Reinforcement LearningTextSequentialChain-of-Thought
🎯 What it does: This paper investigates the phenomenon where supervised training on outputs alone can lead to hidden chain-of-thought (CoT) and demonstrates the issue in three reinforcement learning environments.
Overcoming Joint Intractability with Lossless Hierarchical Speculative Decoding
Yuxuan Zhou (Independent Researcher), Zhi-Qi Cheng (University Of Washington)
GenerationComputational EfficiencyText
🎯 What it does: Proposed Hierarchical Speculative Decoding (HSD), significantly increasing the number of acceptable tokens and inference speed while maintaining the integrity of the target distribution.
Overlap-Adaptive Regularization for Conditional Average Treatment Effect Estimation
Valentyn Melnychuk (LMU Munich), Stefan Feuerriegel (LMU Munich)
Meta Learning
🎯 What it does: Proposed an adaptive regularization method called OAR based on overlap weights to improve the CATE estimation performance of meta-learners in low-overlap regions.
Overlap-weighted orthogonal meta-learner for treatment effect estimation over time
Konstantin Hess (LMU Munich), Stefan Feuerriegel (LMU Munich)
Meta LearningRecurrent Neural NetworkTransformerTabularTime SeriesElectronic Health Records
🎯 What it does: Propose an overlapping weighted orthogonal meta-learner (WO-learner) for estimating time-varying heterogeneous treatment effects.
Overparametrization bends the landscape: BBP transitions at initialization in simple Neural Networks
Brandon Livio Annesi (University of Rome La Sapienza), Chiara Cammarota (University of Rome La Sapienza)
OptimizationRepresentation LearningPhysics Related
🎯 What it does: Studied the Hessian spectrum of overparameterized neural networks at random initialization, analyzed the impact of BBP phase transition on information retrieval, and revealed the emergence of continuous and discontinuous phase transitions under different levels of overparameterization.
Overshoot and Shrinkage in Classifier-Free Guidance: From Theory to Practice
Krunoslav Lehman Pavasovic (FAIR at Meta), Marc Mezard (Bocconi University)
GenerationDiffusion modelScore-based ModelFlow-based ModelImageText
🎯 What it does: Investigated the theoretical behavior of classifier-free guidance (CFG) in high-dimensional conditional generation, proving that CFG can generate the correct target distribution in sufficiently high dimensions; analyzed and quantified the issues of mean overshoot and variance shrinkage in low dimensions, and proposed a nonlinear power-law extended CFG to alleviate these defects while maintaining the quality improvement effects of CFG.
Oversmoothing, "Oversquashing'', Heterophily, Long-Range, and more: Demystifying Common Beliefs in Graph Machine Learning
Adrian Arnaiz-Rodriguez (ELLIS Alicante), Federico Errica (NEC Laboratories Europe)
Representation LearningGraph Neural NetworkGraph
🎯 What it does: Systematically question nine common beliefs in graph machine learning (over-smoothing, over-compression, homogeneity/heterogeneity, and long-distance tasks) and refute them using theoretical and experimental counterexamples.
Overthinking Reduction with Decoupled Rewards and Curriculum Data Scheduling
Shuyang Jiang (Fudan University), Yu Wang (Shanghai Jiao Tong University)
Computational EfficiencyLarge Language ModelReinforcement LearningText
🎯 What it does: To address the overthinking problem in large-scale reasoning models during the generation process, this paper proposes the DECS framework, which significantly compresses the inference path length by separating token-level rewards from batch scheduling.
Overtone: Cyclic Patch Modulation for Clean, Efficient, and Flexible Physics Emulators
Payel Mukhopadhyay (University of Cambridge), Miles Cranmer (University of Cambridge)
TransformerPhysics Related
🎯 What it does: This paper proposes the Overtone framework, which improves the long-term prediction accuracy of PDE surrogates and achieves adjustable computational resource allocation by dynamically controlling the patch size during the autoregressive inference process.
OVID: Open-Vocabulary Intrusion Detection
Fujun Han (Shanghai Artificial Intelligence Laboratory), Peng Ye (Shanghai Artificial Intelligence Laboratory)
Object DetectionSegmentationAnomaly DetectionTransformerVision Language ModelImageTextMultimodality
🎯 What it does: Proposed the open-vocabulary intrusion detection (OVID) task, constructed the Cityintrusion-OpenV dataset, and designed a multimodal end-to-end framework OVIDNet along with two enhancement strategies;
OVSeg3R: Learn Open-vocabulary Instance Segmentation from 2D via 3D Reconstruction
Hongyang Li (South China University of Technology), Lei Zhang (South China University of Technology)
SegmentationTransformerVision Language ModelContrastive LearningSimultaneous Localization and MappingVideoTextPoint Cloud
🎯 What it does: Propose the OVSeg3R training scheme, achieving open-vocabulary 3D instance segmentation using 2D videos and 3D reconstruction models.
OWL : Geometry-Aware Spatial Reasoning for Audio Large Language Models
Subrata Biswas (Worcester Polytechnic Institute), Bashima Islam (Worcester Polytechnic Institute)
Representation LearningTransformerLarge Language ModelSupervised Fine-TuningMultimodalityChain-of-ThoughtAudio
🎯 What it does: Designed and implemented a large language model for spatial audio, OWL, along with its geometry-aware encoder, SAGE, and constructed a large-scale dataset, BiDepth, supporting multi-source localization and multi-step spatial reasoning.
OwlEye: Zero-Shot Learner for Cross-Domain Graph Data Anomaly Detection
Lecheng Zheng (Virginia Tech), Jingrui He (University Of Illinois Urbana Champaign)
Anomaly DetectionGraph Neural NetworkGraph
🎯 What it does: Developed a zero-shot cross-domain graph anomaly detection framework named OWLEYE, capable of detecting anomalies on unseen graphs without requiring retraining;
OXtal: An All-Atom Diffusion Model for Organic Crystal Structure Prediction
Emily Jin (University of Oxford), Cheng-Hao Liu (Aithyra)
Drug DiscoveryTransformerDiffusion modelGraphStochastic Differential Equation
🎯 What it does: OXTAL uses an all-atom-level diffusion model to directly predict the 3D crystal structure of molecules from their 2D chemical graphs.
P-GenRM: Personalized Generative Reward Model with Test-time User-based Scaling
Pinyi Zhang (East China Normal University), Kai Zhang (East China Normal University)
Reinforcement Learning from Human FeedbackSupervised Fine-TuningReinforcement LearningTextBenchmark
🎯 What it does: Propose P-GenRM, a personalized generative reward model that converts user mixed preference signals into a structured evaluation chain (including dynamic user persona and scoring rubrics), and improves scoring accuracy through user-level and prototype-level scaling during testing.
P$^2$-DPO: Grounding Hallucination in Perceptual Processing via Calibration Direct Preference Optimization
ruipeng zhang, Tong Zhang (South China University of Technology)
Reinforcement Learning from Human FeedbackVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Propose a self-supervised Perceptual Processing Direct Preference Optimization (P-DPO) framework that addresses visual perception bottlenecks and insufficient robustness by leveraging model-generated aligned preference dialogues.
P2P: Automated Paper-to-Poster Generation and Fine-Grained Benchmark
Tao Sun (ByteDance), Zhoujun Li (Intelligent Strong Technology Co.,Ltd)
GenerationConvolutional Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningAgentic AIPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Designed and implemented a multi-agent framework named P2P for automatically converting research papers into high-quality, HTML-rendered academic posters, along with the corresponding instruction dataset P2PINSTRUCT and a dual evaluation benchmark combining fine-grained (based on manual checklists) and general (XGBoost-simulated subjective aesthetics) criteria named P2PEVAL.
P3D: Highly Scalable 3D Neural Surrogates for Physics Simulations with Global Context
Benjamin Holzschuh (Technical University of Munich), Nils Thuerey (Technical University of Munich)
Convolutional Neural NetworkTransformerDiffusion modelPhysics Related
🎯 What it does: This paper proposes the P3D hybrid CNN-Transformer architecture as a neural agent for high-resolution 3D physical simulations, demonstrating its performance on various PDEs, turbulence, and channel flow tasks.
PA3FF:Learning Part-Aware Dense 3D Feature Field For Generalizable Articulated Object Manipulation
Yue Chen (Peking University), Hao Dong (Peking University)
Representation LearningRobotic IntelligenceTransformerDiffusion modelContrastive LearningTextPoint Cloud
🎯 What it does: Proposed the Part-Aware 3D Feature Field (PA3FF) and Part-Aware Diffusion Policy (PADP) for achieving general joint control of various manipulable objects by robotic arms.
PaAno: Patch-Based Representation Learning for Time-Series Anomaly Detection
Jinju Park (Sungkyunkwan University), Seokho Kang (Sungkyunkwan University)
Anomaly DetectionConvolutional Neural NetworkContrastive LearningTime SeriesBenchmark
🎯 What it does: Developed a lightweight time series anomaly detection method called PaAno based on patch representation learning.
PAC-Bayes bounds for cumulative loss in Continual Learning
Lior Friedman (Technion Institute of Technology), Ron Meir (Technion Institute of Technology)
OptimizationMeta LearningImageTabular
🎯 What it does: This paper proposes a PAC-Bayes upper bound on accumulated loss under any task distribution in continual learning, and provides an oracle bound for the Gibbs posterior.
PACE: Pretrained Audio Continual Learning
Chang Li (Tsinghua University), Liyuan Wang (Tsinghua University)
ClassificationRepresentation LearningSupervised Fine-TuningAudio
🎯 What it does: Proposed and implemented the PACE framework to achieve continuous learning in the audio domain;
PACEbench: A Framework for Evaluating Practical AI Cyber-Exploitation Capabilities
Zicheng Liu (Shanghai Artificial Intelligence Laboratory), Jing Shao (Shanghai Artificial Intelligence Laboratory)
Large Language ModelAgentic AITextBenchmark
🎯 What it does: Proposes PACEbench, a benchmark for evaluating autonomous network attacks by large language models (LLMs), and designs PACEagent to simulate the phased workflow of a penetration tester.
PAGE-4D: Disentangled Pose and Geometry Estimation for VGGT-4D Perception
Kaichen Zhou (Harvard University), Mengyu Wang (Harvard University)
Pose EstimationDepth EstimationTransformerSupervised Fine-TuningContrastive LearningVideoPoint Cloud
🎯 What it does: In video sequences, PAGE-4D employs a dynamic perception aggregator and local fine-tuning to transfer the static scene VGGT to dynamic environments, enabling simultaneous prediction of camera pose, depth maps, point clouds, and 3D-2D aligned features within less than a second;
PairFlow: Closed-Form Source-Target Coupling for Few-Step Generation in Discrete Flow Models
Mingue Park (Korea Advanced Institute Of Science And Technology), Minhyuk Sung (Korea Advanced Institute Of Science And Technology)
GenerationData SynthesisComputational EfficiencyFlow-based ModelImageGraph
🎯 What it does: Proposes PAIRFLOW, a lightweight preprocessing step for training discrete flow models (DFM) to achieve few-step sampling without requiring pre-trained teacher models.
Pairwise is Not Enough: Hypergraph Neural Networks for Multi-Agent Pathfinding
Rishabh Jain (University of Cambridge), Amanda Prorok (University of Cambridge)
OptimizationGraph Neural NetworkGraph
🎯 What it does: This paper proposes HMAGAT, a multi-agent path planning method that utilizes hypergraph attention networks;
PALC: Preference Alignment via Logit Calibration
SANGHYUN LEE, Hoh Peter In (Korea University)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Propose the PALC framework, achieving test-time preference alignment for LLMs in the vocabulary (logit) space via calibration vectors, by simply inserting a lightweight Calibration Module outside the frozen model.
Pallatom-Ligand: an All-Atom Diffusion Model for Designing Ligand-Binding Proteins
Haochen Wang (ShanghaiTech University), Jiayi Dou (ShanghaiTech University)
Drug DiscoveryTransformerDiffusion modelBiomedical Data
🎯 What it does: This paper proposes Pallatom-Ligand, a full-atom diffusion model for end-to-end generation of proteins bound to given small-molecule ligands.
PAMDP: Interact to Persona Alignment via a Partially Observable Markov Decision Process
ZHE YANG (Tsinghua University), Junlan Feng (Tsinghua University)
GenerationReinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningText
🎯 What it does: This paper models 'achieving personalized alignment through interaction' as a partially observable Markov decision process (PAMDP) and proposes a dual critic reinforcement learning framework to progressively acquire and utilize user implicit preferences in dialogues.
Panda: A pretrained forecast model for chaotic dynamics
Jeffrey B. Lai (University of Texas at Austin), William Gilpin (University of Texas at Austin)
Representation LearningData-Centric LearningTransformerTime SeriesPhysics Related
🎯 What it does: We propose a pre-trained prediction model called Panda specifically for chaotic dynamics, and construct a large-scale dataset containing approximately 20,000 chaotic ODE trajectories using evolutionary algorithms;
Panoptic Pairwise Distortion Graph
Muhammad Kamran Janjua (Huawei Technologies), Bahador Rashidi (Huawei Technologies)
Explainability and InterpretabilityRepresentation LearningTransformerImageBenchmark
🎯 What it does: Proposed a region-level graph structure called Distortion Graph for fine-grained distortion evaluation of image pairs;
Paper Copilot: Tracking the Evolution of Peer Review in AI Conferences
Jing Yang (University of Southern California), Jiaxin Pei (Stanford University)
Text
🎯 What it does: Built a scalable system and public dataset for tracking and analyzing peer review processes and talent trajectories in AI conferences
Paper2Code: Automating Code Generation from Scientific Papers in Machine Learning
Minju Seo (KAIST), Sung Ju Hwang (KAIST)
AI Code AssistantTransformerLarge Language ModelAgentic AITextBenchmark
🎯 What it does: Propose a multi-stage, multi-agent LLM framework called PaperCoder, which automates the generation of complete executable code repositories from machine learning papers.
Paradigm Shift of GNN Explainer from Label Space to Prototypical Representation Space
Jun Yin (Central South University), Chengqi Zhang (Beijing University of Posts and Telecommunications)
Explainability and InterpretabilityRepresentation LearningGraph Neural NetworkGraph
🎯 What it does: Propose to migrate GNN interpreter optimization from the graph label space to the graph representation space, and design a general IDEA framework that utilizes prototype graph representation space for discovering and aligning interpretable subgraphs;
Parallel Sampling from Masked Diffusion Models via Conditional Independence Testing
Iskander Azangulov (Oxford University), Sushrut Karmalkar (Microsoft)
GenerationTransformerDiffusion modelTextBiomedical DataBenchmark
🎯 What it does: This paper proposes a training-agnostic sampler called PUNT, which efficiently identifies approximately conditionally independent token sets that can be parallel decoded in Masked Diffusion Models (MDMs), significantly reducing the number of inference steps and improving generation speed.
Parallel Token Prediction for Language Models
Felix Draxler (University of California Irvine), Stephan Mandt (University of California Irvine)
Computational EfficiencyKnowledge DistillationTransformerTextTime SeriesBenchmark
🎯 What it does: Propose Parallel Token Prediction (PTP), enabling language models to generate multiple tokens at once, eliminating the sequential bottleneck in autoregressive inference.
Parallel-R1: Towards Parallel Thinking via Reinforcement Learning
Tong Zheng (Tencent AI Lab), Dong Yu (Tencent AI Lab)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextChain-of-Thought
🎯 What it does: This paper proposes the Parallel-R1 framework, which utilizes reinforcement learning (RL) to enable large language models to learn and apply parallel thinking from scratch to solve real-world mathematical reasoning tasks;
ParallelBench: Understanding the Trade-offs of Parallel Decoding in Diffusion LLMs
Wonjun Kang (FuriosaAI), Kangwook Lee (UW-Madison)
Computational EfficiencyTransformerLarge Language ModelDiffusion modelTextBenchmark
🎯 What it does: Proposed the PARALLELBENCH evaluation framework for parallel decoding, and empirically validated through information-theoretic analysis the quality degradation of diffusion LLMs (dLLMs) during parallel decoding;
Parameter-Efficient Reinforcement Learning using Prefix Optimization
Itamar Rocha Filho (Harvard University), Samy Jelassi (Harvard University)
Computational EfficiencyTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextPhysics Related
🎯 What it does: Studied methods that perform reinforcement learning optimization only on the top k tokens before the answer, proposing Prefix-Clustering and Prefix-RL, and verified that they can significantly improve the accuracy of large language models on mathematical reasoning tasks.
Parameterization-Based Dataset Distillation of 3D Point Clouds through Learnable Shape Morphing
Dongwook Kim (Ulsan National Institute of Science and Technology), Jae-Young Sim (Ulsan National Institute of Science and Technology)
ClassificationData SynthesisKnowledge DistillationPoint Cloud
🎯 What it does: For dataset distillation on 3D point cloud data, a parameterized framework is proposed, generating diverse synthetic samples by learning shape deformation.
Parameterized Hardness of Zonotope Containment and Neural Network Verification
Vincent Froese (Technische Universitat Berlin), Moritz Stargalla (University of Technology Nuremberg)
Explainability and Interpretability
🎯 What it does: This paper proves that the positivity, surjectivity, computation of L^p-Lipschitz constants, and the corresponding polytope/zonotope containment problem for 2D ReLU networks are W[1]-hard in input dimension d via parameterized complexity analysis, and provides runtime lower bounds based on the Exponential Time Hypothesis (ETH).
Parameters vs. Context: Fine-Grained Control of Knowledge Reliance in Language Models
Baolong Bi (State Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences), Xueqi Cheng (State Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences)
Explainability and InterpretabilityTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Proposed the CK-PLUG plugin, which dynamically controls large language models' dependence on parameter knowledge and retrieved context during retrieval-augmented generation (RAG) through token-level probability modulation.
ParaRNN: Unlocking Parallel Training of Nonlinear RNNs for Large Language Models
Federico Danieli (Apple), Luca Zappella (Apple)
Computational EfficiencyRecurrent Neural NetworkLarge Language ModelText
🎯 What it does: The ParaRNN framework achieves the first large-scale language model-level training of nonlinear RNNs by reconstructing the recurrence relations of nonlinear RNNs as a system of nonlinear equations and implementing sequence parallelization using Newton iteration combined with parallel prefix scan.
ParaS2S: Benchmarking and Aligning Spoken Language Models for Paralinguistic-aware Speech-to-Speech Interaction
Shu-wen Yang (National Taiwan University), Yonghui Wu (ByteDance Seed)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmarkAudio
🎯 What it does: Proposed the ParaS2S framework, including the Paralinguistic-aware S2S benchmark ParaS2SBench, a multi-stage automatic evaluator based on PolyTone, and the ParaS2SAlign method using reinforcement learning (RL), achieving end-to-end voice-to-voice interaction.
PARD: Accelerating LLM Inference with Low‑Cost PARallel Draft Model Adaptation
Zihao An (Advanced Micro Devices Inc), Emad Barsoum (Advanced Micro Devices Inc)
Computational EfficiencyText
🎯 What it does: Propose PARD, a target-agnostic parallel draft model that significantly reduces the adaptation cost of LLM inference.
Pareto Variational Autoencoder
Mincheol Cho (Seoul National University), Joong-Ho Won (Seoul National University)
RestorationGenerationData SynthesisRepresentation LearningAuto EncoderImageTextGraph
🎯 What it does: Proposed a new multivariate symmetric Pareto distribution (symPareto) and constructed the ParetoVAE model, using symPareto as the prior and encoder, supporting t-distribution or symPareto decoders; achieved a closed-form training objective through γ-power divergence, overcoming the challenges of traditional KL divergence in heavy-tail scenarios.
Pareto-Conditioned Diffusion Models for Offline Multi-Objective Optimization
Jatan Shrestha, Joni Pajarinen (Aalto University)
OptimizationDiffusion modelBenchmark
🎯 What it does: This paper proposes a framework called Pareto-Conditioned Diffusion (PCD), which converts offline multi-objective optimization problems into conditional sampling tasks, capable of directly generating high-quality solution sets under given objective trade-offs.
ParoQuant: Pairwise Rotation Quantization for Efficient Reasoning LLM Inference
Yesheng Liang (University Of California San Diego), Zhijian Liu (University Of California San Diego)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Proposing a PTQ method (ParoQuant) based on sparse independent Givens rotations and channel scaling for weight quantization in large language models, achieving low-bitwidth (INT4) weight quantization while significantly reducing quantization error.
Part-level Semantic-guided Contrastive Learning for Fine-grained Visual Classification
Zhijian Lin (Xidian University), Hong Han (Xidian University)
ClassificationConvolutional Neural NetworkTransformerVision Language ModelContrastive LearningImageText
🎯 What it does: Proposes the Part-level Semantic-guided Contrastive Learning (PSCL) framework for fine-grained visual classification, combining part localization, cross-scale multi-branch reasoning, and visual-language contrastive learning.
Part-X-MLLM: Part-aware 3D Multimodal Large Language Model
Chunshi Wang (Zhejiang University), Chunchao Guo (Tencent Hunyuan)
GenerationTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningMultimodalityPoint Cloud
🎯 What it does: Designed Part-X-MLLM, a localized 3D multimodal large language model capable of generating, editing, and answering structured plans (e.g., BBox and edit instructions) from RGB point clouds and natural language prompts, and achieving fine-grained 3D shape generation and editing through an executable syntax-driven backend geometry engine.
Partial Soft-Matching Distance For Neural Representational Comparison With Partial Unit Correspondence
Chaitanya Kapoor (University of California San Diego), Meenakshi Khosla (University of California San Diego)
Representation LearningConvolutional Neural NetworkImageBiomedical Data
🎯 What it does: Proposed a 'partial soft-matching' distance based on partial optimal transport to compare unit correspondences in neural networks or neural recordings, allowing partial unit mismatches to enhance robustness and interpretability.
Partially Equivariant Reinforcement Learning in Symmetry-Breaking Environments
Junwoo Chang (Yonsei University), Jongeun Choi (Yonsei University)
Robotic IntelligenceConvolutional Neural NetworkReinforcement Learning
🎯 What it does: This paper proposes a 'Partial Symmetry Reinforcement Learning' framework that enhances sample efficiency and robustness in symmetric-breaking environments by locally exploiting symmetry.
Partition Generative Modeling: Masked Modeling Without Masks
Justin Deschenaux (EPFL), Caglar Gulcehre (EPFL)
GenerationTransformerImageText
🎯 What it does: Propose Partition Generative Modeling (PGM) for text and image generation tasks, which splits sequences into two groups using Partition Transformer, eliminates [MASK] tokens, improving inference speed while maintaining quality.
PartSAM: A Scalable Promptable Part Segmentation Model Trained on Native 3D Data
Zhe Zhu (Nanjing University of Aeronautics and Astronautics), Mingqiang Wei (Nanjing University of Aeronautics and Astronautics)
SegmentationTransformerPrompt EngineeringNeural Radiance FieldContrastive LearningPoint CloudMesh
🎯 What it does: Proposes PartSAM, an interactive 3D part segmentation model that enables prompt-based segmentation and is directly trained on large-scale 3D data.
PAS: Estimating the target accuracy before domain adaptation
Raphaella Diniz (Simon Fraser University), Martin Ester (Simon Fraser University)
Domain AdaptationImageBenchmark
🎯 What it does: Propose a score called PAS to evaluate the potential adaptability of the source domain and pre-trained model to the target task before domain adaptation, and select the optimal source domain or pre-trained model based on this score;
PASER: Post-Training Data Selection for Efficient Pruned Large Language Model Recovery
Bowei He (City University of Hong Kong), Chen Ma (City University of Hong Kong)
Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
🎯 What it does: Designed and implemented a post-training data selection method called PASER to efficiently restore the capabilities of pruned large language models.
PAT3D: Physics-Augmented Text-to-3D Scene Generation
Guying Lin (Carnegie Mellon University), Minchen Li (Carnegie Mellon University)
SegmentationGenerationVision Language ModelDiffusion modelWorld ModelTextMeshPhysics Related
🎯 What it does: This paper proposes the PAT3D framework, which generates interactive three-dimensional scenes directly usable for physical simulation from text;
Patch-as-Decodable-Token: Towards Unified Multi-Modal Vision Tasks in MLLMs
Yongyi Su (South China University of Technology), Xun Xu (Institute for Infocomm Research, A*STAR)
RecognitionObject DetectionSegmentationTransformerLarge Language ModelVision Language ModelImageText
🎯 What it does: A unified framework PaDT was constructed by introducing Visual Reference Tokens (VRT) and a lightweight decoder, enabling multimodal large language models to directly generate text and visual outputs;