ICLR 2025 Papers — Page 25
International Conference on Learning Representations · 3704 papers
Online Reinforcement Learning in Non-Stationary Context-Driven Environments
Pouya Hamadanian (Massachusetts Institute of Technology), Mohammad Alizadeh (Massachusetts Institute of Technology)
OptimizationReinforcement LearningTabularTime Series
🎯 What it does: An online reinforcement learning algorithm LCPO is proposed, which suppresses catastrophic forgetting through local constraint optimization using observed non-stationary contexts.
Online Reward-Weighted Fine-Tuning of Flow Matching with Wasserstein Regularization
Jiajun Fan (University of Illinois Urbana-Champaign), Ge Liu (University of Illinois Urbana-Champaign)
GenerationCompressionReinforcement LearningFlow-based ModelImage
🎯 What it does: This paper proposes an online reward-weighted continuous flow matching fine-tuning framework (ORW-CFM-W2), which can directly fine-tune continuous flow models using a reward function without explicit likelihood or KL divergence.
Online-to-Offline RL for Agent Alignment
Xu Liu (Shanghai Jiao Tong University), Shuai Li (Shanghai Jiao Tong University)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerReinforcement LearningContrastive LearningSequential
🎯 What it does: An online-to-offline reinforcement learning framework called ALIGN-GAP is proposed, which first uses a Transformer reward model to extract human preferences from a limited set of offline human trajectories, and then aligns the trained game AI with human preferences through reward calibration and preference curriculum learning.
Open-CK: A Large Multi-Physics Fields Coupling benchmarks in Combustion Kinetics
Zaige Fei (University of Science and Technology of China), Yang Wang (University of Science and Technology of China)
TabularBenchmarkPhysics Related
🎯 What it does: A high-resolution multi-physics coupled fire dynamics dataset Open-CK is constructed based on FDS CFD simulation, and multi-model benchmark evaluation is conducted on this dataset.
Open-Set Graph Anomaly Detection via Normal Structure Regularisation
Qizhou Wang (University of Melbourne), Christopher Leckie (University of Melbourne)
Anomaly DetectionGraph Neural NetworkGraph
🎯 What it does: A new open set graph anomaly detection method called NSReg is proposed, which enhances the model's generalization to unseen anomalies through normal structure regularization.
Open-Vocabulary Customization from CLIP via Data-Free Knowledge Distillation
Yongxian Wei (Tsinghua University), Dacheng Tao (Nanyang Technological University)
GenerationKnowledge DistillationContrastive LearningImageText
🎯 What it does: A data-free distillation framework has been developed, capable of customizing small models from the pre-trained CLIP model, supporting open vocabulary task customization using only text prompts or a few example images.
Open-World Reinforcement Learning over Long Short-Term Imagination
Jiajian Li (Shanghai Jiao Tong University), Xiaokang Yang (Shanghai Jiao Tong University)
Reinforcement LearningWorld ModelMultimodalityBenchmark
🎯 What it does: A vision-based model-based reinforcement learning framework called LS-Imagine is proposed, specifically designed to combine long-term imagination and short-term dynamics in open high-dimensional environments (such as Minecraft) to enhance exploration efficiency.
Open-YOLO 3D: Towards Fast and Accurate Open-Vocabulary 3D Instance Segmentation
Mohamed El Amine Boudjoghra (Technical University of Munich), Fahad Shahbaz Khan (Mohammed Bin Zayed University of Artificial Intelligence)
Object DetectionSegmentationImagePoint Cloud
🎯 What it does: This paper proposes a 3D instance segmentation method based on open-source vocabulary called Open-YOLO 3D, which utilizes 2D open-source object detection from multi-view RGB images to label instances in 3D point clouds, avoiding the use of time-consuming SAM/CLIP semantic features.
OpenHands: An Open Platform for AI Software Developers as Generalist Agents
Xingyao Wang (University of Illinois Urbana-Champaign), Graham Neubig (Carnegie Mellon University)
AI Code AssistantTransformerLarge Language ModelAgentic AITextBenchmark
🎯 What it does: An open-source platform called OpenHands has been built, supporting AI agents to interact with the environment through software interfaces such as code, command line, and browser, and providing a framework for multi-agent collaboration and unified evaluation.
OpenMathInstruct-2: Accelerating AI for Math with Massive Open-Source Instruction Data
Shubham Toshniwal (NVIDIA), Igor Gitman (NVIDIA)
TransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: Created the OpenMathInstruct-2 public mathematical reasoning dataset (14M problem-solution pairs) and fine-tuned it on the Llama3.1 base model, resulting in the high-performance OpenMath2-Llama3.1-8B/70B models.
OpenPRM: Building Open-domain Process-based Reward Models with Preference Trees
Kaiyan Zhang (Tsinghua University), Bowen Zhou (Shanghai Artificial Intelligence Laboratory)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Developed OpenPRM, which extends the process-level reward model to open domains by constructing and utilizing preference trees to achieve unsupervised process-level supervision and enhance sampling performance during inference.
OpenRCA: Can Large Language Models Locate the Root Cause of Software Failures?
Junjielong Xu (Microsoft), Qi Zhang (Microsoft)
Large Language ModelAgentic AIMultimodalityBenchmarkFinance Related
🎯 What it does: The OpenRCA benchmark is proposed to evaluate the performance of large language models in real software system root cause analysis (RCA) and a scalable evaluation framework is designed.
OpenVid-1M: A Large-Scale High-Quality Dataset for Text-to-video Generation
Kepan Nan (Nanjing University), Ying Tai (Nanjing University)
GenerationData SynthesisTransformerDiffusion modelVideoTextMultimodality
🎯 What it does: This paper constructs the OpenVid 1M dataset, a high-quality text-video dataset with one million entries, and proposes a parallel visual-text structure multimodal video diffusion Transformer (MVDiT) for text-to-video generation;
Operator Deep Smoothing for Implied Volatility
Ruben Wiedemann (Imperial College), Lukas Gonon (University of St. Gallen)
OptimizationGraph Neural NetworkTime SeriesFinance Related
🎯 What it does: A method for smoothing implied volatility based on Graph Neural Operator (GNO) has been designed and implemented, capable of mapping quotes of any scale and spatial layout to an arbitrage-free smooth surface in one go.
OPTAMI: Global Superlinear Convergence of High-order Methods
Dmitry Kamzolov (Mohamed bin Zayed University of Artificial Intelligence), Martin Takáč (Mohamed bin Zayed University of Artificial Intelligence)
OptimizationTabular
🎯 What it does: The OPTAMI high-order optimization library is proposed, with a new NATA (Nesterov Accelerated Tensor Method with A_t Adaptation) designed, along with a global superlinear convergence theory based on high-order methods, which has been validated in practice for effectiveness.
OptiBench Meets ReSocratic: Measure and Improve LLMs for Optimization Modeling
Zhicheng Yang (Hong Kong University of Science and Technology), Jing Tang (Hong Kong University of Science and Technology)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningTextTabularBenchmark
🎯 What it does: An optimization modeling benchmark for LLMs, OPTIBENCH, has been constructed, and a reverse Socratic data synthesis method, ReSocratic, has been proposed to generate a large number of optimization problems.
Optimal Brain Apoptosis
Mingyuan Sun (Northeastern University), Renjing Xu (Hong Kong University of Science and Technology)
OptimizationComputational EfficiencyConvolutional Neural NetworkTransformerImage
🎯 What it does: A novel network pruning method called Optimal Brain Apoptosis (OBA) is proposed, which directly computes the Hessian vector product to achieve parameter importance assessment for efficient pruning.
Optimal Flow Transport and its Entropic Regularization: a GPU-friendly Matrix Iterative Algorithm for Flow Balance Satisfaction
Liangliang Shi (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
OptimizationGraph
🎯 What it does: A framework for solving Optimal Flow Transport (OFT) on general graphs is proposed, and a GPU-friendly Sinkhorn iterative algorithm (OFT-Sinkhorn) is obtained through entropy regularization, along with an EOFT-Sinkhorn with capacity constraints, to solve the minimum cost flow problem.
Optimal Learning of Kernel Logistic Regression for Complex Classification Scenarios
Hongwei Wen (University of Twente), Hanyuan Hang (Contemporary Amperex Technology)
ClassificationDomain AdaptationTabular
🎯 What it does: This paper studies the conditional class probability (CCP) estimation method based on kernel logistic regression (KLR) in complex classification scenarios such as long-tail learning, label transfer, and domain adaptation, and provides its asymptotic optimal convergence rate on cross-entropy loss;
Optimal Non-Asymptotic Rates of Value Iteration for Average-Reward Markov Decision Processes
Jongmin Lee (Seoul National University), Ernest K. Ryu (UCLA)
OptimizationReinforcement Learning
🎯 What it does: This paper presents the first non-asymptotic (sub-linear) convergence rate for value iteration (VI) and its variants in average reward (undiscounted) Markov Decision Processes (MDP) for multichain MDPs, along with corresponding complexity lower bounds, thereby proving the optimality of these algorithms. It also provides an O(1/k) convergence upper bound for weakly communicating and unichain MDPs, which matches the lower bound.
Optimal Protocols for Continual Learning via Statistical Physics and Control Theory
Francesco Mori, Francesca Mignacco (Princeton University)
OptimizationImagePhysics RelatedOrdinary Differential Equation
🎯 What it does: Under the teacher-student framework, the optimal task selection and learning rate scheduling strategy for continuous learning is derived by combining high-dimensional dynamical equations from statistical physics with the Pontryagin maximum principle.
Optimal Strong Regret and Violation in Constrained MDPs via Policy Optimization
Francesco Emanuele Stradi (Politecnico di Milano), Nicola Gatti (Politecnico di Milano)
OptimizationReinforcement Learning
🎯 What it does: An efficient primal-dual strategy optimization algorithm CPD-PO is proposed for online learning constrained Markov decision processes (CMDP), achieving a lower bound of ˜O(√T) for both strong regret and strong violation under strict constraints.
Optimal Transport for Time Series Imputation
Hao Wang (Zhejiang University), Zhichao Chen (Zhejiang University)
OptimizationTime Series
🎯 What it does: An optimization-based framework for filling missing values in time series, called PSW-I, is proposed.
Optimality and Adaptivity of Deep Neural Features for Instrumental Variable Regression
Juno Kim (University of Tokyo), Zhu Li (University College London)
🎯 What it does: This paper studies and proves the convergence properties of two-stage deep feature instrumental variable regression (DFIV) when the structural function belongs to the Besov space, and provides the asymptotically optimal learning rate.
Optimality of Matrix Mechanism on $\ell_p^p$-metric
Zongrui Zou (Nanjing University), Jalaj Upadhyay (Rutgers University)
OptimizationSafty and Privacy
🎯 What it does: The study addresses the optimality of answering linear queries under differential privacy constraints using ℓₚₚ error metrics; it proposes the ℓₚₚ error and provides lower and upper bounds under (ε, δ)-DP, proving that the matrix mechanism is optimal up to a polylogarithmic factor under this error metric; it further provides the optimal error for prefix sums and even queries.
Optimistic Games for Combinatorial Bayesian Optimization with Application to Protein Design
Melis Ilayda Bal (Max Planck Institute for Intelligent Systems), Andreas Krause (Google DeepMind)
OptimizationDrug DiscoveryProtein Structure PredictionBiomedical Data
🎯 What it does: This paper proposes a game-theory-based combinatorial Bayesian optimization method called GAMEOPT, which treats discrete variables as players in a cooperative game, using the Upper Confidence Bound (UCB) as a reward function to compute the game equilibrium point for iterative selection of evaluation samples.
Optimization by Parallel Quasi-Quantum Annealing with Gradient-Based Sampling
Yuma Ichikawa (Fujitsu Limited), Yamato Arai (Fujitsu Limited)
OptimizationGraph
🎯 What it does: A general combinatorial optimization solver named PQQA is proposed, which combines continuous relaxation, gradient updates, quasi-quantum annealing (QQA), and multi-threaded GPU communication.
Optimized Multi-Token Joint Decoding With Auxiliary Model for LLM Inference
Zongyue Qin (University of California), Yizhou Sun (California Institute of Technology)
OptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper proposes Multi-Word Joint Decoding (MTJD) and its efficient approximate version, Multi-Word Assisted Decoding (MTAD), which achieves the one-time generation of multiple words by sampling on a small model and validating on a large model, while ensuring output quality.
Optimizing $(L_0, L_1)$-Smooth Functions by Gradient Methods
Daniil Vankov (Arizona State University), Sebastian U Stich
Optimization
🎯 What it does: This study investigates gradient optimization methods for (L, L₀₁) smooth functions and provides new convergence theories and step size formulas.
Optimizing 4D Gaussians for Dynamic Scene Video from Single Landscape Images
In-Hwan Jin (Pusan National University), Kyeongbo Kong
GenerationDepth EstimationOptimizationGaussian SplattingOptical FlowImageVideoPoint Cloud
🎯 What it does: This paper proposes a framework for learning a 4D Gaussian distribution from a single landscape image, enabling natural animation of elements such as fluids in three-dimensional space.
Optimizing Backward Policies in GFlowNets via Trajectory Likelihood Maximization
Timofei Gritsaev (Higher School of Economics), Daniil Tiapkin (Ecole Polytechnique)
OptimizationReinforcement LearningSequential
🎯 What it does: This paper proposes a method called Trajectory Likelihood Maximization (TLM) for optimizing the backward policy in Generative Flow Networks (GFlowNet) and integrates it with the soft reinforcement learning (soft RL) based GFlowNet training process.
Optimizing importance weighting in the presence of sub-population shifts
Floris Holstege (University of Amsterdam), Cees Diks (University of Amsterdam)
Domain AdaptationOptimizationSupervised Fine-TuningImageText
🎯 What it does: This paper proposes a framework for finding optimal importance weights through double-layer optimization under the distribution shift of subpopulations, and validates its effectiveness in fine-tuning the last layer of deep networks.
Optimizing Neural Network Representations of Boolean Networks
Joshua Russell (University of Massachusetts Amherst), Hava T Siegelmann
CompressionOptimizationGraphTabularTime SeriesSequentialElectrocardiogram
🎯 What it does: Lossless compression and optimization of the neural network representation of Boolean networks
Optimizing Posterior Samples for Bayesian Optimization via Rootfinding
Taiwo Adebiyi, Ruda Zhang (University of Houston)
Optimization
🎯 What it does: This paper proposes a global optimization strategy named TS-roots for gradient multi-start optimization of posterior sample paths in high-dimensional Bayesian optimization.
OptionZero: Planning with Learned Options
Po-Wei Huang (Academia Sinica), Ti-Rong Wu (Academia Sinica)
Reinforcement LearningVideo
🎯 What it does: Based on MuZero, option networks and an improved dynamics network have been added, allowing the agent to automatically discover and use temporally extended actions (options) during self-play, thereby reducing state iteration costs during search.
Oracle efficient truncated statistics
Konstantinos Karatapanis (Archimedes Athena Research Center), Christos Tzamos (University of Athens)
Optimization
🎯 What it does: The study investigates the problem of learning from truncated samples and proposes a new learning method that can effectively solve optimization problems in polynomial time.
Order-aware Interactive Segmentation
Bin Wang (Northwestern University), Ziyan Wu (University of Central Florida)
SegmentationDepth EstimationComputational EfficiencyTransformerImageVideo
🎯 What it does: This paper proposes an interactive segmentation framework called OIS based on relative depth information, utilizing an order map and order-aware attention mechanism as well as an object-aware attention mechanism, and integrates dense and sparse hints to achieve efficient and accurate segmentation.
ORSO: Accelerating Reward Design via Online Reward Selection and Policy Optimization
Chen Bo Calvin Zhang (Massachusetts Institute of Technology), Pulkit Agrawal
OptimizationRobotic IntelligenceLarge Language ModelReinforcement LearningSequential
🎯 What it does: This study proposes a framework for online reward selection and strategy optimization (ORSO) that automatically selects the optimal reward function from a set of candidate reward functions to accelerate reward design in reinforcement learning.
Oryx MLLM: On-Demand Spatial-Temporal Understanding at Arbitrary Resolution
Zuyan Liu (Tsinghua University), Yongming Rao (Tencent)
RetrievalCompressionTransformerLarge Language ModelVision Language ModelImageVideoTextMultimodality
🎯 What it does: A unified multimodal large language model Oryx is proposed, supporting visual inputs of any resolution and achieving spatial-temporal understanding of images, videos, and three-dimensional scenes.
OS-ATLAS: Foundation Action Model for Generalist GUI Agents
Zhiyong Wu (Shanghai AI Laboratory), Yu Qiao (Shanghai AI Laboratory)
Data SynthesisRobotic IntelligenceTransformerSupervised Fine-TuningVision Language ModelMultimodality
🎯 What it does: Developed a foundational action model OS-Atlas for multi-platform GUI agents, supporting interaction across desktop, mobile, and web interfaces;
OSCAR: Operating System Control via State-Aware Reasoning and Re-Planning
Xiaoqiang Wang (University of Montreal), Bang Liu (University of Montreal)
Robotic IntelligenceAI Code AssistantLarge Language ModelAgentic AITextMultimodality
🎯 What it does: Developed OS-CAR, a general-purpose agent for desktop and mobile devices that can convert natural language instructions into executable Python code, thereby completing complex GUI tasks through standard OS control methods such as mouse and keyboard.
Oscillatory State-Space Models
T. Konstantin Rusch (Massachusetts Institute of Technology), Daniela Rus (Massachusetts Institute of Technology)
ClassificationComputational EfficiencyRecurrent Neural NetworkTime SeriesSequentialOrdinary Differential Equation
🎯 What it does: A linear oscillation state space model (LinOSS) based on the dynamics of a forced harmonic oscillator is proposed, and it is demonstrated that it can maintain stability, interpretability, and universal approximation capability in long sequence learning.
OSDA Agent: Leveraging Large Language Models for De Novo Design of Organic Structure Directing Agents
Zhaolin Hu (Zhejiang University), Yi Yang (Zhejiang University)
OptimizationDrug DiscoveryTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: An OSDA Agent framework is proposed, utilizing an interactive loop between large language models and computational chemistry tools to achieve the redesign and optimization of organic structure directing agents (OSDAs).
OSTQuant: Refining Large Language Model Quantization with Orthogonal and Scaling Transformations for Better Distribution Fitting
Xing Hu (Houmo AI), Sifan Zhou
TransformerLarge Language ModelText
🎯 What it does: This paper proposes OSTQuant, which utilizes learnable orthogonal and scaling transformations for post-training quantization of LLM weights and activations.
Out-of-distribution Generalization for Total Variation based Invariant Risk Minimization
Yuanchao Wang (Duke University), Tianqi Zhong (Beihang University)
Domain AdaptationOptimizationGenerative Adversarial NetworkTabular
🎯 What it does: A Lagrangian multiplier framework based on Total Variation (TV) called OOD-TV-IRM is proposed, which constructs an adversarial learning process of primal-dual optimization and semi-Nash equilibrium, aimed at enhancing the model's out-of-distribution (OOD) generalization ability.
Outlier Synthesis via Hamiltonian Monte Carlo for Out-of-Distribution Detection
Hengzhuang Li (Huazhong University of Science and Technology), Teng Zhang (Huazhong University of Science and Technology)
Anomaly DetectionContrastive LearningImage
🎯 What it does: The HamOS framework is proposed, which utilizes Hamiltonian Monte Carlo to generate diverse and representative virtual anomaly samples in the unit sphere feature space, and uses them for training to enhance OOD detection performance.
Overcoming False Illusions in Real-World Face Restoration with Multi-Modal Guided Diffusion Model
Keda TAO, Nan Cheng (Xidian University)
RestorationGenerationDiffusion modelImageMultimodality
🎯 What it does: A multi-modal guided real face restoration model MGFR is proposed, which utilizes attribute text prompts, reference high-quality images, and identity information to jointly guide the restoration, significantly reducing artifacts and achieving controllable restoration.
Overcoming Lower-Level Constraints in Bilevel Optimization: A Novel Approach with Regularized Gap Functions
Wei Yao (National Center for Applied Mathematics Shenzhen), Jin Zhang (Southern University of Science and Technology)
OptimizationTabularBiomedical Data
🎯 What it does: This paper proposes a single-loop, Hessian-free solver BiC-GAFFA for solving lower-level constraint-coupled bilevel optimization problems, and extends it to lower-level min-max problems.
Overcoming Slow Decision Frequencies in Continuous Control: Model-Based Sequence Reinforcement Learning for Model-Free Control
Devdhar Patel (University of Massachusetts Amherst), Hava T Siegelmann
Recurrent Neural NetworkReinforcement LearningSequential
🎯 What it does: This paper proposes a model-based Sequence Reinforcement Learning (SRL) algorithm that enables agents to generate action sequences at a lower decision frequency, achieving model-free control.
OvercookedV2: Rethinking Overcooked for Zero-Shot Coordination
Tobias Gessler (FLAIR University of Oxford), Jakob Nicolaus Foerster
Convolutional Neural NetworkRecurrent Neural NetworkReinforcement LearningSequentialBenchmark
🎯 What it does: This study addresses the problem of Zero-Shot Cooperation (ZSC) in the Overcooked environment, proposing a state-enhanced training method and designing a new version, OvercookedV2, as a more challenging ZSC benchmark.
OVTR: End-to-End Open-Vocabulary Multiple Object Tracking with Transformer
Jinyang Li (Huazhong University of Science and Technology), Wenbing Tao (Huazhong University of Science and Technology)
Object TrackingTransformerContrastive LearningImageVideo
🎯 What it does: An end-to-end open vocabulary multi-object tracking framework (OVTR) is proposed, capable of achieving continuous tracking and classification without the need for candidate boxes and post-processing.
P-SPIKESSM: HARNESSING PROBABILISTIC SPIKING STATE SPACE MODELS FOR LONG-RANGE DEPENDENCY TASKS
Malyaban Bal (Pennsylvania State University), Abhronil Sengupta (Pennsylvania State University)
Spiking Neural NetworkTextSequentialBenchmarkAudio
🎯 What it does: A scalable spiking network based on a probabilistic state space model (P‑SpikeSSM) is proposed, which utilizes a SpikeSampler layer to randomly generate spikes and achieves multi-layer parallel communication through SpikeMixer and ClampFuse.
PABBO: Preferential Amortized Black-Box Optimization
Xinyu Zhang (Aalto University), Julien Martinelli (Universit' e de Bordeaux)
OptimizationTransformerReinforcement LearningTabular
🎯 What it does: This paper proposes PABBO, an end-to-end amortized preference Bayesian optimization framework that utilizes Transformer neural processes and reinforcement learning to directly output the acquisition values of candidate pairs, achieving black-box function optimization under preference feedback.
PaCA: Partial Connection Adaptation for Efficient Fine-Tuning
Sunghyeon Woo (Seoul National University), Dongsuk Jeon (Seoul National University)
OptimizationComputational EfficiencyTransformerSupervised Fine-TuningText
🎯 What it does: A new parameter-efficient fine-tuning method called PaCA is proposed, which fine-tunes by randomly selecting a portion of connections in the pre-trained weights, avoiding the sequential processing of adapter layers.
Pacmann: Efficient Private Approximate Nearest Neighbor Search
Mingxun Zhou (Carnegie Mellon University), Giulia Fanti (Carnegie Mellon University)
RetrievalSafty and PrivacyComputational EfficiencyGraph Neural Network
🎯 What it does: A scheme for performing approximate nearest neighbor search on a massive vector database without disclosing the query vector is provided;
PAD: Personalized Alignment of LLMs at Decoding-time
Ruizhe Chen (Zhejiang University), Zuozhu Liu (Zhejiang University)
Recommendation SystemReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: A framework for personalized alignment during the inference phase, called PAD, is proposed, which guides the generation of LLMs using a personalized reward model without the need for additional training.
PADRe: A Unifying Polynomial Attention Drop-in Replacement for Efficient Vision Transformer
Pierre-David Letourneau (Qualcomm AI Research), Fatih Porikli (Qualcomm AI Research)
ClassificationObject DetectionComputational EfficiencyTransformerImage
🎯 What it does: A polynomial approximation-based attention replacement framework PADRe is proposed, which can directly replace the self-attention mechanism in Transformers.
Painting with Words: Elevating Detailed Image Captioning with Benchmark and Alignment Learning
Qinghao Ye (ByteDance Research), Haoqi Fan
GenerationOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningVision Language ModelImageTextBenchmark
🎯 What it does: This paper proposes the DCSCORE fine-grained image description evaluation metric, the DECAPBENCH detailed image description benchmark, and designs the FEEDQUILL fine-grained feedback collection and reinforcement learning method to improve the image description quality of VLMs and reduce hallucinations.
Pairwise Elimination with Instance-Dependent Guarantees for Bandits with Cost Subsidy
Ishank Juneja (Carnegie Mellon University), Osman Yagan
Recommendation SystemOptimizationReinforcement LearningTabular
🎯 What it does: A new algorithm called Pairwise Elimination (PE) and its extension PE-CS are proposed and analyzed for the cost-subsidized multi-armed bandit (MAB-CS) problem, specifically addressing two constraint scenarios: known reference arms and optimal subsidy rewards.
PAL: Sample-Efficient Personalized Reward Modeling for Pluralistic Alignment
Daiwei Chen (University of Wisconsin-Madison), Ramya Korlakai Vinayak (University of Wisconsin-Madison)
Recommendation SystemReinforcement LearningTextMultimodality
🎯 What it does: A personalized reward modeling framework named PAL is proposed for paradigm alignment under diverse human preferences.
PaLD: Detection of Text Partially Written by Large Language Models
Eric Lei (University of Pennsylvania), Chun-Fu Chen
ClassificationRecognitionOptimizationTransformerLarge Language ModelText
🎯 What it does: This paper proposes the Partial-LLM Detector (PaLD), which can estimate the proportion of LLM-generated content in mixed text and locate LLM paragraphs.
PALMBENCH: A COMPREHENSIVE BENCHMARK OF COMPRESSED LARGE LANGUAGE MODELS ON MOBILE PLATFORMS
Yilong Li (University of Wisconsin Madison), Suman Banerjee (University of Wisconsin Madison)
CompressionComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextMultimodalityBenchmark
🎯 What it does: This paper presents PalmBench, a lightweight automated LLM benchmark framework for mobile devices, designed to evaluate compressed models on different mobile platforms in terms of memory, power consumption, throughput, as well as accuracy, toxicity, and hallucination metrics.
Palu: KV-Cache Compression with Low-Rank Projection
Chi-Chih Chang (National Yang Ming Chiao Tung University), Kai-Chiang Wu (National Yang Ming Chiao Tung University)
CompressionComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: A post-training KV-Cache compression framework called Palu is proposed, which significantly reduces KV-Cache memory usage and accelerates inference while maintaining accuracy by performing low-rank decomposition on Key/Value projection weights and caching low-dimensional latent representations.
Pangea: A Fully Open Multilingual Multimodal LLM for 39 Languages
Xiang Yue (Carnegie Mellon University), Graham Neubig (Carnegie Mellon University)
TransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: A multimodal large language model PANGEA supporting 39 languages has been constructed.
PaPaGei: Open Foundation Models for Optical Physiological Signals
Arvind Pillai (Dartmouth), Mohammad Malekzadeh (Nokia Bell Labs)
ClassificationRepresentation LearningConvolutional Neural NetworkMixture of ExpertsContrastive LearningTime SeriesBiomedical DataElectronic Health Records
🎯 What it does: The first large-scale photoplethysmography (PPG) foundational model, PAPAGEI, has been constructed and made publicly available for multi-task health monitoring.
PaRa: Personalizing Text-to-Image Diffusion via Parameter Rank Reduction
Shangyu Chen (Monash University), Dinh Phung (Monash University)
GenerationData SynthesisConvolutional Neural NetworkDiffusion modelImage
🎯 What it does: This paper achieves personalization of text-to-image models by explicitly controlling the rank of the output from the diffusion model layers (reducing parameter rank);
Param$\Delta$ for Direct Mixing: Post-Train Large Language Model At Zero Cost
Sheng Cao (Meta Platforms Inc), Zechun Liu (Meta Reality Labs)
TransformerLarge Language ModelText
🎯 What it does: This paper proposes a zero-cost parameter differential addition method, Param Δ, which directly adds the weight differences from the post-training of the old model to the updated base model, achieving an approximate post-training effect without the need for additional training.
Parameter and Memory Efficient Pretraining via Low-rank Riemannian Optimization
Zhanfeng Mo (Nanyang Technological University), Sinno Jialin Pan (Chinese University of Hong Kong)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: The researchers propose LORO, a low-rank Riemannian optimizer capable of pre-training low-rank parameterized language models from scratch.
Parameter Expanded Stochastic Gradient Markov Chain Monte Carlo
Hyunsu Kim (Korea Advanced Institute of Science and Technology), Juho Lee (Korea Advanced Institute of Science and Technology)
ClassificationDomain AdaptationOptimizationConvolutional Neural NetworkImageStochastic Differential Equation
🎯 What it does: Parameter extension is applied to traditional Stochastic Gradient Markov Chain Monte Carlo (SGMCMC) by splitting each layer's weight matrix into several matrix products, in order to improve sampling diversity and enhance the predictive uncertainty of Bayesian neural networks and their robustness to out-of-distribution (OOD) data.
ParaSolver: A Hierarchical Parallel Integral Solver for Diffusion Models
Jianrong Lu (City University of Hong Kong), Junhui Hou (City University of Hong Kong)
GenerationComputational EfficiencyDiffusion modelImage
🎯 What it does: Transform the sequential sampling process of diffusion models into solving banded nonlinear equations to achieve hierarchical parallel sampling, proposing the ParaSolver framework.
Pareto Low-Rank Adapters: Efficient Multi-Task Learning with Preferences
Nikolaos Dimitriadis (École Polytechnique Fédérale de Lausanne), François Fleuret (University of Geneva)
SegmentationDepth EstimationAutonomous DrivingConvolutional Neural NetworkSupervised Fine-TuningImageMultimodality
🎯 What it does: This paper proposes PaLoRA, a parameter-efficient method for achieving Pareto front learning in multi-task learning through low-rank adapters, which supports dynamic switching of different task trade-off points based on user preferences during inference.
Pareto Prompt Optimization
Guang Zhao (Brookhaven National Laboratory), Xiaoning Qian (Texas A&M University)
OptimizationTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: A reinforcement learning framework based on multi-objective dominance relations (ParetoPrompt) is proposed for the automatic generation of prompts on the Pareto front.
ParetoFlow: Guided Flows in Multi-Objective Optimization
Ye Yuan (McGill), Xue Liu (MILA - Quebec AI Institute)
OptimizationFlow-based ModelTabularBenchmark
🎯 What it does: A flow matching-based offline multi-objective optimization framework called ParetoFlow is proposed, which utilizes a unified weight vector and neighborhood evolution to guide sampling in approximating the Pareto front.
ParFam -- (Neural Guided) Symbolic Regression via Continuous Global Optimization
Philipp Scholl (LMU Munich), Gitta Kutyniok (LMU Munich)
OptimizationTransformerTabularBenchmarkPhysics Related
🎯 What it does: The ParFam method (and its pre-trained version DL-ParFam) is proposed, which transforms symbolic regression from discrete search to continuous optimization using parameterized rational function networks, and solves it through global optimization.
Partial Gromov-Wasserstein Metric
Yikun Bai (Vanderbilt University), Soheil Kolouri (University of California)
RetrievalOptimizationReinforcement LearningPoint Cloud
🎯 What it does: The Partial Gromov-Wasserstein (PGW) problem is proposed, its metric properties in metric spaces are proven, and two Frank-Wolfe solvers are provided; subsequently, its performance is validated in tasks such as shape matching, retrieval, and interpolation.
Partially Observed Trajectory Inference using Optimal Transport and a Dynamics Prior
Anming Gu (Boston University), Kristjan Greenewald (IBM Research)
Time SeriesStochastic Differential Equation
🎯 What it does: The paper proposes a method for inferring partially observed trajectories based on optimal transport and dynamic priors, aimed at recovering the temporal dynamics of a population from unpaired temporal marginal data.
PARTNR: A Benchmark for Planning and Reasoning in Embodied Multi-agent Tasks
Matthew Chang (FAIR Meta), Tsung-Yen Yang (FAIR Meta)
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: In this work, we constructed a large-scale collaborative execution task benchmark named PARTNR, which includes 100,000 natural language instructions, covering 60 multi-room simulated residences and 5,819 unique objects, to evaluate the planning, perception, and execution performance of human-machine collaboration, and compared various LLM planning methods with human cooperation.
PathGen-1.6M: 1.6 Million Pathology Image-text Pairs Generation through Multi-agent Collaboration
Yuxuan Sun (Zhejiang University), Lin Yang (Center for Interdisciplinary Research and Innovation)
GenerationData SynthesisRetrievalLarge Language ModelVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: A multi-agent collaborative system called PathGen-1.6M is proposed, which automatically generates 1.6 million high-quality pathology image-text pairs.
PEAR: Primitive Enabled Adaptive Relabeling for Boosting Hierarchical Reinforcement Learning
Utsav Singh (Indian Institute of Technology Kanpur), Vinay P. Namboodiri
Robotic IntelligenceReinforcement LearningAgentic AI
🎯 What it does: Based on a small number of expert demonstrations, adaptive relabeling of demonstration trajectories is performed using the value function of the current lower-level policy to generate achievable sub-goals. This is combined with reinforcement learning and imitation learning to jointly optimize the hierarchical policy, thereby improving the learning efficiency and stability of hierarchical reinforcement learning in sparse reward long-horizon tasks.
PEARL: Parallel Speculative Decoding with Adaptive Draft Length
Tianyu Liu (University of Science and Technology of China), Xiao Sun (Shanghai AI Laboratory)
GenerationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: The PEARL framework is proposed, utilizing pre-validation and post-validation strategies to achieve parallel and adaptive draft lengths for Speculative Decoding, accelerating LLM inference.
PEARL: Towards Permutation-Resilient LLMs
Liang CHEN, Kam-Fai Wong (Chinese University of Hong Kong)
GenerationOptimizationAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: A framework named PEARL is proposed, which utilizes distributed robust optimization and a learnable permutation generation network (P-Net) to enhance the robustness of large language models in few-shot scenarios regarding demonstration order arrangements, and achieves adaptive learning of adversarial arrangements through adversarial training.
Pedestrian Motion Reconstruction: A Large-scale Benchmark via Mixed Reality Rendering with Multiple Perspectives and Modalities
Yichen Wang (Shanghai Jiao Tong University), Liqing Zhang (Shanghai Jiao Tong University)
Object DetectionObject TrackingPose EstimationTransformerDiffusion modelSimultaneous Localization and MappingVideoMultimodalityPoint CloudBenchmark
🎯 What it does: A large-scale pedestrian motion reconstruction dataset PMR based on a mixed reality platform has been proposed and released, and various SOTA methods have been benchmarked from third-person, first-person, and LiDAR perspectives.
Periodic Materials Generation using Text-Guided Joint Diffusion Model
KISHALAY DAS, Niloy Ganguly (Indian Institute of Technology Kharagpur)
GenerationData SynthesisGraph Neural NetworkDiffusion modelText
🎯 What it does: This paper proposes a text-guided periodic material generation model TGDMat, which integrates diffusion atomic types, coordinates, and lattices, and introduces text context during the denoising process.
PeriodWave: Multi-Period Flow Matching for High-Fidelity Waveform Generation
Sang-Hoon Lee (Ajou University), Seong-Whan Lee (Korea University)
GenerationData SynthesisConvolutional Neural NetworkFlow-based ModelOrdinary Differential EquationAudio
🎯 What it does: A full-scene waveform generation model called PeriodWave is proposed, based on periodic-aware flow matching, achieving multi-period estimation, discrete wavelet multi-band modeling, and FreeU denoising technology, enhancing high-frequency information and periodic reproduction.
Perm: A Parametric Representation for Multi-Style 3D Hair Modeling
Chengan He (Yale University), Yi Zhou (Adobe Research)
GenerationData SynthesisAuto EncoderGenerative Adversarial NetworkMesh
🎯 What it does: This paper presents PERM, a lightweight and separable parameter 3D hair representation model that can be used for single-view reconstruction, hair editing, and as a general prior for conditional generation.
Permute-and-Flip: An optimally stable and watermarkable decoder for LLMs
Xuandong Zhao (University of California Berkeley), Yu-Xiang Wang (University of California San Diego)
GenerationTransformerLarge Language ModelText
🎯 What it does: A new decoding method called Permute-and-Flip (PF) is proposed for generating LLM text, along with a corresponding watermark scheme PF watermark.
Perplexed by Perplexity: Perplexity-Based Data Pruning With Small Reference Models
Zachary Ankner (MIT), Mansheej Paul (Databricks)
Large Language ModelText
🎯 What it does: Using a small language model to compute perplexity for trimming large pre-trained corpora, and studying the impact of trimming on downstream task performance in scenarios of different domain compositions, overfitting, and data limitations.
Perplexity Trap: PLM-Based Retrievers Overrate Low Perplexity Documents
Haoyu Wang (Renmin University of China), Ji-Rong Wen (Renmin University of China)
RetrievalTransformerLarge Language ModelText
🎯 What it does: This paper explains and verifies the source bias caused by the over-scoring of low perplexity (PPL) documents by pre-trained language model (PLM) retrievers through causal graphs, and proposes a causal diagnosis and correction-based debiasing method (CDC) to mitigate this bias during inference.
Persistent Pre-training Poisoning of LLMs
Yiming Zhang (Carnegie Mellon University), Daphne Ippolito (Google DeepMind)
Adversarial AttackTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This study investigates whether attacks can persist after the subsequent alignment training (SFT + DPO) when a small amount of poisoned data is injected during the LLM pre-training phase.
Personality Alignment of Large Language Models
Minjun Zhu (Zhejiang University), Yue Zhang (Westlake University)
TransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: This paper proposes the concept of 'personalized alignment', constructs a large-scale real user personality questionnaire dataset PAPI, and develops a parameter-free activation search (PAS) method, enabling large language models to quickly and cost-effectively align with individual personality traits such as the Big Five and the Dark Triad while maintaining their original capabilities.
Personalized Representation from Personalized Generation
Shobhita Sundaram (Massachusetts Institute of Technology), Phillip Isola (Massachusetts Institute of Technology)
ClassificationObject DetectionSegmentationGenerationRetrievalDiffusion modelContrastive LearningImageBenchmark
🎯 What it does: By fine-tuning Stable Diffusion with only a small number of real images to generate synthetic data, and then using contrastive learning (LoRA + InfoNCE) to learn personalized representations on a pre-trained visual encoder, the performance of multi-tasking (classification, retrieval, detection, segmentation) is improved.
Personalized Visual Instruction Tuning
Renjie Pi (Hong Kong University of Science and Technology), Tong Zhang (University of Illinois Urbana-Champaign)
GenerationData SynthesisTransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Designed and implemented the Personalized Visual Instruction Tuning (PVIT) framework, which enhances the performance of multimodal large language models (MLLMs) in conversations tailored to specific individuals through an automated data generation pipeline.
PersonalLLM: Tailoring LLMs to Individual Preferences
Thomas P Zollo, Hongseok Namkoong (Columbia University)
Meta LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: A public benchmark PersonalLLM is proposed to evaluate the personalized alignment performance of LLMs;
Perturbation-Restrained Sequential Model Editing
Jun-Yu Ma (University of Science and Technology of China), Jia-Chen Gu (University of California, Los Angeles)
Large Language ModelText
🎯 What it does: The PRUNE framework is proposed, which reduces the disturbance to existing knowledge by constraining the condition number of the edit matrix during the continuous model editing process, thereby maintaining the general capabilities of large language models after multiple edits.
PerturboLLaVA: Reducing Multimodal Hallucinations with Perturbative Visual Training
Cong Chen (Zhejiang University), Chunhua Shen (Zhejiang University of Technology)
RecognitionGenerationTransformerLarge Language ModelSupervised Fine-TuningImageTextMultimodalityBenchmark
🎯 What it does: This paper proposes a fine-grained multimodal text error evaluation metric called HalFscore, and significantly reduces hallucinations in dense image captions by incorporating adversarial text (Perturbative Visual Training) during the training phase to lessen the model's reliance on language priors.
PETRA: Parallel End-to-end Training with Reversible Architectures
Stephane Rivaud (Sorbonne Université), Edouard Oyallon (Flatiron Institute)
OptimizationComputational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: This paper proposes PETRA, a framework that utilizes reversible networks to achieve model parallel training, allowing for the decoupling of forward and backward propagation while maintaining single-version parameters, thus enabling the parallelization of forward and backward gradient computations.
PFDiff: Training-Free Acceleration of Diffusion Models Combining Past and Future Scores
Guangyi Wang (Xiamen University), Song-Zhi Su
GenerationComputational EfficiencyDiffusion modelImageOrdinary Differential Equation
🎯 What it does: This paper proposes a training-free sampling acceleration method called PFDiff, which can be combined with any ODE solver. It utilizes past and future score information to skip time steps, thereby reducing the number of function evaluations and correcting discretization errors.
PFGuard: A Generative Framework with Privacy and Fairness Safeguards
Soyeon Kim (Korea Advanced Institute of Science and Technology), Steven Euijong Whang (Korea Advanced Institute of Science and Technology)
GenerationData SynthesisSafty and PrivacyGenerative Adversarial NetworkGaussian SplattingImage
🎯 What it does: This paper proposes and implements PFGuard, a generative framework that simultaneously meets the requirements of differential privacy and fairness, suitable for high-dimensional image data.
PharmacoMatch: Efficient 3D Pharmacophore Screening via Neural Subgraph Matching
Daniel Rose (University of Vienna), Thierry Langer (University of Vienna)
Drug DiscoveryGraph Neural NetworkContrastive LearningGraph
🎯 What it does: A neural subgraph matching framework called PharmacoMatch based on contrastive learning has been developed for rapid 3D pharmacophore screening.
Phidias: A Generative Model for Creating 3D Content from Text, Image, and 3D Conditions with Reference-Augmented Diffusion
Zhenwei Wang (City University of Hong Kong), Rynson W. H. Lau (City University of Hong Kong)
GenerationData SynthesisDiffusion modelImageTextPoint Cloud
🎯 What it does: Phidias is proposed, a diffusion model based on 3D references for generating high-quality 3D content from text, images, or 3D conditions, supporting a unified framework and various applications;