ICML 2025 Papers — Page 23
International Conference on Machine Learning · 3257 papers
PiD: Generalized AI-Generated Images Detection with Pixelwise Decomposition Residuals
Xinghe Fu (Zhejiang University), Xi Li (Zhejiang University)
ClassificationAnomaly DetectionConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: A pixel-level decomposition residual-based AI-generated image detection method (PiD) is proposed, which quantizes the RGB pixel vector after color space transformation, using the resulting residuals as feature inputs for a lightweight network for binary classification.
PieClam: A Universal Graph Autoencoder Based on Overlapping Inclusive and Exclusive Communities
Daniel Zilberg (Technion Israel Institute of Technology), Ron Levie (Technion Israel Institute of Technology)
Anomaly DetectionRepresentation LearningGraph Neural NetworkAuto EncoderGraph
🎯 What it does: This paper presents PieClam, a general graph autoencoder model that probabilistically generates graphs by embedding nodes into inclusive and exclusive community spaces and learning prior distributions.
PIGDreamer: Privileged Information Guided World Models for Safe Partially Observable Reinforcement Learning
Dongchi Huang (University of Beihang), Kaige Zhang
Safty and PrivacyReinforcement LearningWorld ModelTabularBenchmark
🎯 What it does: PIGDreamer is proposed, a model-based RL method that utilizes privileged information from the training phase to enhance the world model and policy in partially observable safe reinforcement learning.
PILAF: Optimal Human Preference Sampling for Reward Modeling
Yunzhen Feng (New York University), Yaqi Duan (New York University)
OptimizationReinforcement Learning from Human FeedbackReinforcement LearningText
🎯 What it does: This paper presents PILAF, a response sampling strategy for preference labeling in RLHF, aimed at aligning the learning of the reward model with the optimal agent reward objective.
Piloting Structure-Based Drug Design via Modality-Specific Optimal Schedule
Keyue Qiu (Tsinghua University), Wei-Ying Ma
Drug DiscoveryFlow-based ModelMultimodalityBiomedical Data
🎯 What it does: In structural-based drug design, the MolPilot system is proposed, utilizing optimal noise scheduling (VOS) with multimodal (continuous 3D positions and discrete 2D topology) for molecular generation and ligand docking.
PINNsAgent: Automated PDE Surrogation with Large Language Models
Qingpo Wuwu (Peking University), Shanghang Zhang (Peking University)
OptimizationComputational EfficiencyHyperparameter SearchLarge Language ModelAgentic AITabularBenchmarkPhysics Related
🎯 What it does: A multi-agent framework named PINNsAgent has been developed, which utilizes large language models (LLM) to automatically generate, evaluate, and iteratively optimize the network architecture and hyperparameters of Physics-Informed Neural Networks (PINNs), thereby solving various partial differential equations (PDE) without the need for manual tuning.
PIPA: Preference Alignment as Prior-Informed Statistical Estimation
Junbo Li (University of Texas at Austin), qiang liu
Large Language ModelReinforcement LearningText
🎯 What it does: A non-RL probabilistic framework PIPA is proposed, which achieves preference alignment of language models by combining maximum likelihood estimation with prior constraints;
PipeOffload: Improving Scalability of Pipeline Parallelism with Memory Optimization
Xinyi Wan (Sea AI Lab), Jialin Li (National University of Singapore)
TransformerLarge Language ModelText
🎯 What it does: This paper proposes PipeOffload, a scheme to enhance the scalability of Pipeline Parallelism (PP) by activating memory offloading to the host and combining it with new pipeline scheduling.
PISA Experiments: Exploring Physics Post-Training for Video Diffusion Models by Watching Stuff Drop
Chenyu Li (New York University), Saining Xie (New York University)
GenerationOptimizationSupervised Fine-TuningReinforcement LearningDiffusion modelRectified FlowVideoPhysics Related
🎯 What it does: This paper studies post-training on a pre-trained video diffusion model to accurately simulate the physical process of free fall of objects.
Pivoting Factorization: A Compact Meta Low-Rank Representation of Sparsity for Efficient Inference in Large Language Models
Jialin Zhao (Tsinghua University), Carlo Vittorio Cannistraci (Tsinghua University)
CompressionComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: An end-to-end, training-free low-rank compression framework MPIFA is proposed, which combines PIFA meta low-rank representation and online error accumulation minimization reconstruction to compress the linear layers of LLMs.
Pixel-level Certified Explanations via Randomized Smoothing
Alaa Anani (Max Planck Institute for Informatics), Bernt Schiele (Max Planck Institute for Informatics)
SegmentationExplainability and InterpretabilityConvolutional Neural NetworkTransformerGaussian SplattingImage
🎯 What it does: A pixel-level provably interpretable framework based on randomized smoothing is proposed to generate robust and interpretable pixel importance maps for any black-box attribution method.
Pixel2Feature Attack (P2FA): Rethinking the Perturbed Space to Enhance Adversarial Transferability
Renpu Liu (Nanjing University of Information Science and Technology), Jinwei Wang (Nankai University)
Adversarial AttackConvolutional Neural NetworkTransformerImage
🎯 What it does: Designed and implemented the Pixel2Feature Attack (P2FA), which transfers the perturbation space from pixel space to feature space, perturbs multiple times along the feature importance direction in feature space, and then maps the perturbation back to pixel space using feature inversion, thereby generating more transferable adversarial examples.
Plan-and-Act: Improving Planning of Agents for Long-Horizon Tasks
Lutfi Eren Erdogan (University of California Berkeley), Amir Gholami (University of California Berkeley)
TransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: By separating high-level planning from low-level execution, the PLAN-AND-ACT framework was constructed to enhance the performance of large language models in long-sequence tasks.
Plausible Token Amplification for Improving Accuracy of Differentially Private In-Context Learning Based on Implicit Bayesian Inference
Yusuke Yamasaki (NTT Corporation), Takayuki Miura (NTT Corporation)
ClassificationSafty and PrivacyLarge Language ModelPrompt EngineeringText
🎯 What it does: A method called Plausible Token Amplification (PTA) is proposed to generate more accurate synthetic examples in the context of differential privacy in-context learning (DP-ICL).
Playmate: Flexible Control of Portrait Animation via 3D-Implicit Space Guided Diffusion
Xingpei Ma (Guangzhou Quwan Network Technology), Shunsi Zhang (Guangzhou Quwan Network Technology)
GenerationData SynthesisPose EstimationTransformerDiffusion modelVideoAudio
🎯 What it does: This paper presents Playmate, a two-stage training framework that implements audio-driven facial animation using a diffusion model guided by 3D implicit space, supporting fine-grained control of emotions and poses.
PlaySlot: Learning Inverse Latent Dynamics for Controllable Object-Centric Video Prediction and Planning
Angel Villar-Corrales (University of Bonn), Sven Behnke (University of Bonn)
Robotic IntelligenceTransformerVideo
🎯 What it does: The PlaySlot model is proposed, utilizing unlabeled video self-supervised learning of object slots and latent actions to achieve controllable object-centric video prediction and planning.
Point Cloud Dataset Distillation
Deyu Bo (National University of Singapore), Xinchao Wang (National University of Singapore)
ClassificationSegmentationData SynthesisKnowledge DistillationPoint Cloud
🎯 What it does: The DD3D framework is proposed, which uses a minimal amount of synthetic point clouds to replace large-scale real point clouds while maintaining the performance of classification/segmentation models.
Point-Level Topological Representation Learning on Point Clouds
Vincent Peter Grande, Michael T Schaub
Representation LearningPoint CloudBenchmark
🎯 What it does: A method called TOPF is proposed, which utilizes tools such as persistent homology, α/VR filtration, and the harmonic space of Hodge Laplacian to transform global topological structures into interpretable features for each point, and based on this, clustering and downstream tasks are performed.
Pointwise Information Measures as Confidence Estimators in Deep Neural Networks: A Comparative Study
Shelvia Wongso (National University of Singapore), Mehul Motani (National University of Singapore)
ClassificationAnomaly DetectionConvolutional Neural NetworkImage
🎯 What it does: This study investigates three pointwise information measures (PMI, PSI, PVI) as post-processing methods for confidence assessment in deep neural networks, comparing their performance in tasks such as error prediction detection, selective prediction, and confidence calibration.
PoisonBench: Assessing Language Model Vulnerability to Poisoned Preference Data
Tingchen Fu (Renmin University of China), Fazl Barez (University of Oxford)
Recommendation SystemAdversarial AttackReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningContrastive LearningTextBenchmark
🎯 What it does: This paper proposes the POISONBENCH benchmark to evaluate the vulnerability of LLMs to data poisoning attacks during the preference learning phase, and experiments are conducted on 22 models.
PoisonedEye: Knowledge Poisoning Attack on Retrieval-Augmented Generation based Large Vision-Language Models
Chenyang Zhang (Xidian University), Xiaofeng Chen (Xidian University)
RetrievalAdversarial AttackTransformerVision Language ModelImageTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: A knowledge poisoning attack targeting the Visual-Language Retrieval-Augmented Generation (VLRAG) system is proposed, where the attacker only needs to inject a single poisoned image-text pair to manipulate the system's response to the target query.
PokéChamp: an Expert-level Minimax Language Agent
Seth Karten (Princeton University), Chi Jin (Princeton University)
TransformerLarge Language ModelReinforcement LearningWorld ModelText
🎯 What it does: Developed PokéChamp, an expert-level Minimax language agent based on LLM for Pokémon battles.
Policy Design for Two-sided Platforms with Participation Dynamics
Haruka Kiyohara (Cornell University), Sarah Dean (Cornell University)
Recommendation SystemOptimizationReinforcement LearningTabularSequential
🎯 What it does: This study investigates the engagement dynamics between audiences and content providers on two-sided platforms (such as video streaming and recruitment) and their impact on recommendation policies, proposing a recommendation strategy that considers long-term demographic effects.
Policy Filtration for RLHF to Mitigate Noise in Reward Models
Chuheng Zhang (Microsoft Research), Jiang Bian (Microsoft Research)
Reinforcement Learning from Human FeedbackReinforcement LearningText
🎯 What it does: The study improves the signal-to-noise ratio of policy learning in RLHF training by filtering unreliable samples from the reward model.
Policy Gradient with Tree Expansion
Gal Dalal (NVIDIA Research), Gal Chechik (Bar-Ilan University)
Reinforcement LearningVideo
🎯 What it does: This paper proposes a SoftTreeMax strategy that combines tree search (Tree Expansion) with policy gradient, utilizing multi-step cumulative rewards and future state logits to form a differentiable softmax alternative.
Policy Guided Tree Search for Enhanced LLM Reasoning
Yang Li
Computational EfficiencyGraph Neural NetworkTransformerLarge Language ModelReinforcement LearningTextChain-of-Thought
🎯 What it does: A policy-guided tree search framework (PGTS) based on reinforcement learning is proposed, which dynamically decides to expand, branch, backtrack, or terminate during the inference process of large language models, thereby achieving efficient exploration of inference paths.
Policy Optimization for CMDPs with Bandit Feedback: Learning Stochastic and Adversarial Constraints
Francesco Emanuele Stradi (Politecnico di Milano), Nicola Gatti (Politecnico di Milano)
OptimizationReinforcement Learning
🎯 What it does: This paper studies how to perform online learning in Constrained Markov Decision Processes (CMDPs) under bandwidth feedback, proposing a new algorithm that can handle both stochastic and adversarial constraints simultaneously.
Policy Regularization on Globally Accessible States in Cross-Dynamics Reinforcement Learning
Zhenghai Xue (Nanyang Technological University), Shuicheng YAN (National University of Singapore)
Reinforcement LearningGenerative Adversarial Network
🎯 What it does: A strategy regularization method based on accessible states, ASOR, is proposed to address the issue of expert state distribution failure in cross-dynamic reinforcement learning.
Policy-labeled Preference Learning: Is Preference Enough for RLHF?
Taehyun Cho (Seoul National University), Jungwoo Lee (Seoul National University)
Robotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningContrastive LearningSequential
🎯 What it does: A new RLHF framework is proposed - Policy-labeled Preference Learning (PPL), which addresses the likelihood mismatch problem in traditional methods by incorporating behavioral policy labels into preference modeling and utilizing regret-based scores.
Policy-Regret Minimization in Markov Games with Function Approximation
Thanh Nguyen-Tang (Johns Hopkins University), Raman Arora (Johns Hopkins University)
OptimizationReinforcement Learning
🎯 What it does: Proposes the BOVL framework to address the issue of policy degradation in Markov games with function approximation.
Poly2Vec: Polymorphic Fourier-Based Encoding of Geospatial Objects for GeoAI Applications
Maria Despoina Siampou (University of Southern California), Hua Lu (Aalborg University)
Point Cloud
🎯 What it does: A unified geographic object encoding framework called POLY2VEC is proposed, which can simultaneously handle points, polylines, and polygons, and directly maps geometric information into a vector space usable by machine learning models.
polybasic Speculative Decoding Through a Theoretical Perspective
Ruilin Wang (Xiamen University), Rongrong Ji (Xiamen University)
OptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper studies the polybasic framework for multi-model speculative decoding and provides a theoretical derivation of its optimal inference time.
PolyConf: Unlocking Polymer Conformation Generation through Hierarchical Generative Models
Fanmeng Wang (Renmin University of China), Zhifeng Gao (DP Technology)
GenerationData SynthesisTransformerDiffusion modelGraphBenchmark
🎯 What it does: This paper proposes PolyConf, a polymer conformation generation method based on a hierarchical generative model, and creates the PolyBench polymer conformation benchmark dataset.
Polynomial Time Learning Augmented Algorithms for NP-hard Permutation Problems
Evripidis Bampis (Sorbonne Université), Michalis Xefteris (Athens University of Economics and Business)
Optimization
🎯 What it does: This paper proposes a learning-enhanced framework that utilizes the prediction of the order of element pairs (with a prediction accuracy of at least 1/2 + ε) to obtain optimal solutions in polynomial time (with high probability) for NP-hard permutation problems that satisfy decomposability or c-locality (such as maximum acyclic subgraph, minimum linear arrangement, traveling salesman, keyword bidding, etc.). The framework first obtains an approximate order with O(n log n) queries (with an error of O(log n)), and then uses dynamic programming to complete the optimal arrangement, thus achieving a 'learnable' solution.
Polynomial-Delay MAG Listing with Novel Locally Complete Orientation Rules
Tian-Zuo Wang (Nanjing University), Zhi-Hua Zhou (Nanjing University)
Graph
🎯 What it does: Proposes directional rules under single-point background knowledge and designs the first polynomial delay MAG enumeration algorithm based on this;
Polynomial-Time Approximability of Constrained Reinforcement Learning
Jeremy McMahan (University of Wisconsin Madison)
OptimizationReinforcement Learning
🎯 What it does: A polynomial-time (0,ϵ)-additive dual-bound approximation algorithm is designed to find near-optimal policies in constrained Markov decision processes (CMDPs) with any number of constraints.
POQD: Performance-Oriented Query Decomposer for Multi-vector retrieval
Yaoyang Liu (Renmin University), zhen chen
RetrievalOptimizationTransformerLarge Language ModelPrompt EngineeringMultimodalityRetrieval-Augmented Generation
🎯 What it does: A performance-oriented query decomposition framework POQD is proposed, which utilizes LLM to decompose queries and searches for the best prompt through a prompt optimizer to enhance multi-vector retrieval (MVR) performance;
POROver: Improving Safety and Reducing Overrefusal in Large Language Models with Overgeneration and Preference Optimization
Batuhan K. Karaman (Cornell University), Xia Song (Microsoft)
OptimizationSafty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes the POROver (Preference Optimization for Reducing Overrefusal) strategy by using more advanced teacher models (such as GPT-4o) for overgeneration of general and toxic instructions, aiming to reduce the overrefusal rate of large language models while maintaining high safety.
Portable Reward Tuning: Towards Reusable Fine-Tuning across Different Pretrained Models
Daiki Chijiwa (NTT Corporation), Susumu Takeuchi (NTT Corporation)
OptimizationReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningImageText
🎯 What it does: A Portable Reward Tuning (PRT) framework is proposed, viewing fine-tuning as reward maximization under KL regularization, training an explicit reward model, which allows for 'simulated fine-tuning' of new models using only this reward model and any pre-trained model during inference.
Positional Attention: Expressivity and Learnability of Algorithmic Computation
Artur Back de Luca (University of Waterloo), Kimon Fountoulakis (University of Waterloo)
TransformerTabularSequential
🎯 What it does: This paper proposes and theoretically analyzes the expressiveness and learnability of an attention mechanism that uses only positional encoding (Positional Transformer) in algorithm execution tasks.
Positional Encoding meets Persistent Homology on Graphs
Yogesh Verma (Aalto University), Vikas K Garg
Drug DiscoveryGraph Neural NetworkGraph
🎯 What it does: This paper proposes the PiPE method, which combines graph positional encoding with persistent homology to enhance the expressive power of graph neural networks.
Positive-unlabeled AUC Maximization under Covariate Shift
Atsutoshi Kumagai (NTT Corporation), Yasuhiro Fujiwara (Yokohama National University)
ClassificationDomain AdaptationAnomaly DetectionConvolutional Neural NetworkContrastive LearningImageTabular
🎯 What it does: Proposes a method to maximize AUC in positive-unlabeled (PU) learning under covariate shift.
Posterior Inference with Diffusion Models for High-dimensional Black-box Optimization
Taeyoung Yun (Korea Advanced Institute of Science and Technology), Jinkyoo Park (Korea Advanced Institute of Science and Technology)
OptimizationDiffusion modelTabularOrdinary Differential Equation
🎯 What it does: Using diffusion models and deep ensemble agents for high-dimensional black-box function optimization, achieving a balance between exploration and exploitation through posterior sampling.
Potemkin Understanding in Large Language Models
Marina Mancoridis (Massachusetts Institute of Technology), Sendhil Mullainathan (Massachusetts Institute of Technology)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper introduces the concept of 'Potemkin Understanding' and explores the phenomenon where large language models perform excellently on standard benchmarks yet lack genuine conceptual understanding. It quantifies the prevalence of this phenomenon through two evaluation methods.
Power Mean Estimation in Stochastic Continuous Monte-Carlo Tree Search
Tuan Quang Dam (University of Lille)
OptimizationReinforcement LearningTabular
🎯 What it does: This paper studies a new MCTS algorithm called Stochastic-Power-UCT, which utilizes power mean estimators and polynomial exploration rewards for planning in stochastic MDPs, and proves the convergence of its root node value estimation.
PPDiff: Diffusing in Hybrid Sequence-Structure Space for Protein-Protein Complex Design
Zhenqiao Song (Carnegie Mellon University), Martin Renqiang Min (NEC Laboratories America)
Drug DiscoveryProtein Structure PredictionGraph Neural NetworkTransformerDiffusion modelBiomedical DataBenchmark
🎯 What it does: A diffusion model called PPDiff is proposed for the joint design of protein ligand sequences and structures, aimed at generating high-affinity binding proteins for any protein target.
Pre-training Auto-regressive Robotic Models with 4D Representations
Dantong Niu (Berkeley AI Research), Roei Herzig (Berkeley AI Research)
Robotic IntelligenceTransformerReinforcement LearningVideo
🎯 What it does: ARM4R is proposed, an autoregressive robot model that utilizes low-level 4D representations (3D point tracking) learned from human videos for pre-training to enhance robot control performance.
Pre-Training Graph Contrastive Masked Autoencoders are Strong Distillers for EEG
Xinxu Wei (Lehigh University), Yu Zhang (Lehigh University)
Knowledge DistillationRepresentation LearningGraph Neural NetworkAuto EncoderContrastive LearningTime SeriesBiomedical Data
🎯 What it does: A unified graph contrastive masked autoencoder pre-training framework (EEG-DisGCMAE) is proposed, serving as a knowledge distiller from high-density EEG to low-density EEG, enhancing diagnostic performance in scarce label environments.
Preconditioned Riemannian Gradient Descent Algorithm for Low-Multilinear-Rank Tensor Completion
Yuanwei Zhang (Shanghai Jiao Tong University), Jian-Feng Cai (Hong Kong University of Science and Technology)
OptimizationVideo
🎯 What it does: This paper addresses the low multilinear rank tensor completion problem and proposes a Preconditioned Riemannian Gradient Descent (PRGD) algorithm, which utilizes a data-driven Riemannian metric to accelerate convergence while maintaining approximately optimal sampling complexity.
Predicting High-precision Depth on Low-Precision Devices Using 2D Hilbert Curves
Mykhail Uss, Jaeyun Jeong (Konkuk University)
Depth EstimationPoint Cloud
🎯 What it does: High dynamic range depth is encoded into two-dimensional low dynamic range values using a two-dimensional Hilbert curve on low-precision devices, and high-precision depth is restored in the post-processing stage.
Predicting mutational effects on protein binding from folding energy
Arthur Deng (Stanford University), Brian L. Trippe
Protein Structure PredictionSupervised Fine-TuningBiomedical Data
🎯 What it does: This paper proposes a deep learning model STAB-DDG that uses folding energy to predict changes in protein-protein binding energy.
Predicting the Susceptibility of Examples to Catastrophic Forgetting
Guy Hacohen (KU Leuven), Tinne Tuytelaars (KU Leuven)
Image
🎯 What it does: This paper proposes a learning speed-based buffer sampling strategy by observing the relationship between the learning speed of neural networks and catastrophic forgetting.
Prediction models that learn to avoid missing values
Lena Stempfle (Chalmers University of Technology), Fredrik D. Johansson (Chalmers University of Technology)
TabularAlzheimer's Disease
🎯 What it does: A missing value avoidance (MA) machine learning framework is proposed, which actively encourages decision trees, sparse linear models, and tree ensemble models to minimize reliance on missing features during prediction through a penalty term, while maintaining or improving predictive performance.
Prediction via Shapley Value Regression
Amr Alkhatib (Orebro University), Michalis Vazirgiannis (Ecole Polytechnique)
Explainability and InterpretabilityComputational EfficiencyConvolutional Neural NetworkImageTabular
🎯 What it does: A method called ViaSHAP is proposed, which embeds the calculation of Shapley values into model training, allowing predictions to be directly obtained by summing Shapley values, while no additional post-processing is required during inference.
Prediction-Aware Learning in Multi-Agent Systems
Aymeric Capitaine (Centre de Mathématiques Appliquées - CNRS - École Polytechnique), Alain Oliviero Durmus
OptimizationReinforcement Learning from Human FeedbackReinforcement LearningAgentic AITabularTime Series
🎯 What it does: This paper proposes a predictive-aware learning framework for enabling players to anticipate natural states in time-varying multi-agent games and adjust their strategies accordingly, along with the corresponding POMWU algorithm.
Prediction-Powered Adaptive Shrinkage Estimation
Sida Li (University of Chicago), Nikolaos Ignatiadis (University of Chicago)
Text
🎯 What it does: A multiple mean estimation method (PAS) that combines Predictive Power Improvement (PPI) with empirical Bayesian shrinkage has been proposed and validated.
Prediction-Powered E-Values
Daniel Csillag (Getulio Vargas Foundation), Guilherme Tegoni Goedert (Getulio Vargas Foundation)
TabularTime SeriesElectronic Health Records
🎯 What it does: By introducing prediction-driven e-values, a general prediction-powered inference framework is proposed, and its application is demonstrated in four practical cases (estimation of diabetes prevalence, online risk monitoring, change point detection, and causal structure learning).
Predictive Data Selection: The Data That Predicts Is the Data That Teaches
KaShun SHUM, Junxian He (Hong Kong University of Science and Technology)
Computational EfficiencyData-Centric LearningTransformerLarge Language ModelText
🎯 What it does: This paper proposes a data selection method called PRESELECT, which selects the most helpful data for learning by evaluating the correlation between the loss of text on different pre-trained models and the ranking of downstream task performance.
Predictive Performance of Deep Quantum Data Re-uploading Models
Xin Wang (Tsinghua University), Rebing Wu
Image
🎯 What it does: This paper reveals that under high-dimensional data, the predictive performance of deep data re-uploading models degrades to near random guessing through theoretical analysis and experimental research.
Preference Adaptive and Sequential Text-to-Image Generation
Ofir Nabati (Google Research), Craig Boutilier (Google Research)
GenerationRecommendation SystemOptimizationLarge Language ModelReinforcement LearningPrompt EngineeringDiffusion modelImageTextMultimodality
🎯 What it does: This work proposes and implements an interactive multi-turn text-to-image generation system called PASTA, which utilizes reinforcement learning to progressively optimize the generated images through prompts during interactions with users, in order to meet their changing preferences.
Preference Controllable Reinforcement Learning with Advanced Multi-Objective Optimization
Yucheng Yang (Eindhoven University of Technology), Meng Fang (University of Liverpool)
OptimizationReinforcement LearningSequentialBenchmark
🎯 What it does: A Preference Controllable Reinforcement Learning (PCRL) framework is proposed, training preference-conditioned meta-policies to achieve adaptive balance across multiple objectives.
Preference Learning for AI Alignment: a Causal Perspective
Kasia Kobalczyk, Mihaela van der Schaar (University of Cambridge)
Recommendation SystemReinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningText
🎯 What it does: Proposes to view preference learning as a causal inference problem, exploring causal assumptions such as user goal confounding and potential treatment variables in LLM alignment;
Preference learning made easy: Everything should be understood through win rate
Lily H Zhang, Rajesh Ranganath (New York University)
Recommendation SystemOptimizationReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningText
🎯 What it does: A preference learning framework centered on win rate is proposed, proving that win rate is the only evaluation metric consistent with the preference sampling distribution, and categorizing preference learning methods into win rate optimization (WRO) and non-win rate optimization (non-WRO).
Preference Optimization for Combinatorial Optimization Problems
Mingjun Pan (China Mobile), Chun Yuan (Tsinghua University)
OptimizationTransformerReinforcement LearningTabularBenchmark
🎯 What it does: A Preference Optimization (PO) framework is proposed to convert quantitative rewards in COPs into qualitative preferences and train neural solvers through an entropy-regularized RL objective, combined with local search for fine-tuning.
Preference-CFR: Beyond Nash Equilibrium for Better Game Strategies
Qi Ju (Huazhong University of Science and Technology), YunFeng Luo
OptimizationReinforcement Learning from Human FeedbackReinforcement Learning
🎯 What it does: This paper studies a Pref-CFR algorithm that adjusts strategies in incomplete information games by introducing preference and vulnerability parameters, which can generate diverse Nash and ε-Nash equilibrium strategies that align with user preferences while maintaining an acceptable range of revenue loss.
Premise-Augmented Reasoning Chains Improve Error Identification in Math reasoning with LLMs
Sagnik Mukherjee (University of Illinois at Urbana Champaign), Dilek Hakkani Tur
TransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: This paper proposes transforming Linear Reasoning Chains (LRC) into Premise-Augmented Reasoning Chains (PARC), verifying each step based solely on its premises, thereby improving the error identification of large language models in mathematical reasoning.
Preserving AUC Fairness in Learning with Noisy Protected Groups
Mingyang Wu (Purdue University), Shu Hu (Purdue University)
OptimizationImageTabular
🎯 What it does: A robust fairness learning framework is proposed for the AUC optimization process in the presence of label noise in protected groups;
Pretraining Generative Flow Networks with Inexpensive Rewards for Molecular Graph Generation
Mohit Pandey (University of British Columbia), Emmanuel Bengio (Recursion Pharmaceuticals)
GenerationDrug DiscoveryGraph Neural NetworkReinforcement LearningFlow-based ModelGraph
🎯 What it does: A generative flow network (Atomic GFlowNet, A-GFN) was constructed with atoms as the operational unit, and the distribution of drug-like molecules on low-cost properties (such as QED, TPSA, SAS, and the number of rings) was learned through pre-training, followed by fine-tuning on target properties to generate diverse and high-quality molecules.
Prices, Bids, Values: One ML-Powered Combinatorial Auction to Rule Them All
Ermis Soumalias (University of Zurich), Sven Seuken (University of Zurich)
OptimizationComputational EfficiencyTabular
🎯 What it does: This paper proposes a new machine learning-driven hybrid combinatorial auction mechanism (MLHCA) that achieves more efficient iterative combinatorial auctions.
Primal-Dual Neural Algorithmic Reasoning
Yu He (Stanford University), Ellen Vitercik (Stanford University)
OptimizationGraph Neural NetworkGraph
🎯 What it does: A neural algorithm reasoning based on the primal-dual framework (PDNAR) is proposed, which can simulate and surpass various approximation algorithms for NP-hard problems.
PRIME: Deep Imbalanced Regression with Proxies
Jongin Lim (Samsung Electronics), Jinwoo Shin (Korea Advanced Institute of Science and Technology)
Representation LearningTabular
🎯 What it does: The PRIME framework is proposed, which constructs a balanced and ordered feature space using learnable proxies, and achieves representation learning for deep imbalanced regression through proxy alignment.
Primitive Vision: Improving Diagram Understanding in MLLMs
Shan Zhang (Australian Institute for Machine Learning), Yuan Xue (Ohio State University)
Object DetectionSegmentationData SynthesisTransformerLarge Language ModelVision Language ModelImageMultimodality
🎯 What it does: Design and implement the PRIMITIVE framework, which combines the GeoGLIP visual encoder specifically for geometric primitive recognition with a multimodal large language model (MLLM), enhancing the fine-grained perception and reasoning capabilities of mathematical charts through the generation of visual soft prompts.
Primphormer: Efficient Graph Transformers with Primal Representations
Mingzhen He (Shanghai Jiao Tong University), Xiaolin Huang (Shanghai Jiao Tong University)
Computational EfficiencyRepresentation LearningGraph Neural NetworkTransformerGraph
🎯 What it does: This paper proposes Primphormer, a graph transformer that achieves linear complexity through principal pair representation, enabling self-attention without pairwise computation.
Principal-Agent Bandit Games with Self-Interested and Exploratory Learning Agents
Junyan Liu (University of Washington), Lillian J. Ratliff (University of Washington)
OptimizationReinforcement Learning from Human FeedbackReinforcement LearningAgentic AITabular
🎯 What it does: This paper studies repeated principal-agent gambling games, where the principal indirectly explores an unknown environment by incentivizing the agent. Unlike previous research, this paper considers a self-interested learning agent that iteratively updates reward estimates and may explore with some probability.
Principled Algorithms for Optimizing Generalized Metrics in Binary Classification
Anqi Mao (New York University), Yutao Zhong (Google)
ClassificationOptimizationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a unified and theoretically guaranteed algorithm (METRO) for optimizing generalized performance metrics in binary classification (such as F‑β, Jaccard, weighted accuracy, etc.).
Principled Data Selection for Alignment: The Hidden Risks of Difficult Examples
Chengqian Gao (MBZUAI), zhiqiang xu
Reinforcement Learning from Human FeedbackSupervised Fine-TuningText
🎯 What it does: This paper studies the impact of matching preference data difficulty with model capacity on alignment effectiveness, proposing the principle that 'diverse preference data difficulty, with overly difficult examples harming alignment', and based on this, introduces the Selective DPO method to filter out overly difficult alignment samples.
Prior Knowledge Guided Neural Architecture Generation
Jingrong Xie (Sichuan University), Yanan Sun (Sichuan University)
Neural Architecture SearchGraph Neural NetworkDiffusion modelImageAudio
🎯 What it does: A method for generating neural architectures guided by prior knowledge (PG-NAG) is proposed, which directly generates high-performance neural network architectures without searching and evaluating.
Privacy Amplification by Structured Subsampling for Deep Differentially Private Time Series Forecasting
Jan Schuchardt (Technical University of Munich), Stephan Günnemann
OptimizationSafty and PrivacyTime Series
🎯 What it does: This paper proposes a differential privacy training framework for time series forecasting that incorporates structured subsampling and data augmentation, capable of training deep models while ensuring strict event-level and user-level privacy.
Privacy Amplification Through Synthetic Data: Insights from Linear Regression
Clément Pierquin (Craft AI), Matthieu Boussard (Université de Lille)
Data SynthesisSafty and PrivacySupervised Fine-Tuning
🎯 What it does: This paper studies the phenomenon of privacy enhancement of synthetic data under differential privacy, particularly in the context of using linear regression models. Through theoretical analysis, it explores the conditions for privacy leakage and privacy enhancement of synthetic data.
Privacy Attacks on Image AutoRegressive Models
Antoni Kowalczuk (CISPA Helmholtz Center for Information Security), Adam Dziedzic (Warsaw University of Technology)
GenerationSafty and PrivacyDiffusion modelImage
🎯 What it does: This paper conducts a comprehensive privacy risk assessment of Image Autoregressive Models (IAR) and proposes and experiments with new Membership Inference Attacks (MIA), Dataset Inference (DI), and Training Data Extraction Attacks;
Privacy-Preserving Federated Convex Optimization: Balancing Partial-Participation and Efficiency via Noise Cancellation
Roie Reshef (Technion), Kfir Yehuda Levy (Technion)
OptimizationFederated LearningSafty and PrivacyTabular
🎯 What it does: In the scenario of federated learning with partial device participation, a new noise cancellation mechanism is proposed, which can ensure differential privacy while maintaining the same convergence rate and linear computational complexity as fully participating schemes.
Privacy-Shielded Image Compression: Defending Against Exploitation from Vision-Language Pretrained Models
Xuelin Shen (Shenzhen University), Wenhan Yang (Pengcheng Laboratory)
CompressionOptimizationSafty and PrivacyAdversarial AttackVision Language ModelAuto EncoderImageText
🎯 What it does: The framework PSIC implements privacy protection during the image compression phase, generating a bitstream that can be decoded on demand into an encrypted version and a full version.
Private Federated Learning using Preference-Optimized Synthetic Data
Charlie Hou (Carnegie Mellon University), Giulia Fanti (Carnegie Mellon University)
OptimizationFederated LearningSafty and PrivacyTransformerLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: A private federated learning method based on policy optimization, POPri, is proposed, which utilizes LLM to fine-tune synthetic data generation based on client feedback as rewards.
Private Lossless Multiple Release
Joel Daniel Andersson (Basic Algorithms Research Copenhagen), Rasmus Pagh (Basic Algorithms Research Copenhagen)
Safty and Privacy
🎯 What it does: A general framework is proposed to achieve lossless multiple releases of differential privacy mechanisms under multi-level privacy budgets, ensuring that any order and any number of releases maintain the same distribution and privacy loss as a single release with the highest privacy budget.
Private Model Personalization Revisited
Conor Snedeker (Ohio State University), Raef Bassily (Ohio State University)
Federated LearningSafty and PrivacyTabular
🎯 What it does: This paper proposes a federated personalized learning algorithm (Private FedRep) under user-level differential privacy (DP) protection, achieving personalized model training for heterogeneous user data through the sharing of low-dimensional representations, and provides dimension-independent risk bounds for private initialization and classification problems.
Proactive Agents for Multi-Turn Text-to-Image Generation Under Uncertainty
Meera Hahn (Google DeepMind), Zi Wang (Google DeepMind)
GenerationTransformerLarge Language ModelAgentic AIImageTextMultimodality
🎯 What it does: This paper proposes an active multi-turn text-to-image generation agent that can actively ask questions when uncertain and demonstrate its understanding of user intentions through an editable belief map.
Probabilistic Factorial Experimental Design for Combinatorial Interventions
Divya Shyamal (Massachusetts Institute of Technology), Caroline Uhler (Massachusetts Institute of Technology)
TabularSequential
🎯 What it does: A probabilistic factor experimental design method is proposed for the experimental design of combinatorial interventions, addressing the issue of how to effectively conduct experiments in the presence of potential interaction effects among multiple treatments.
Probabilistic Group Mask Guided Discrete Optimization for Incremental Learning
Fengqiang Wan (Nanjing University of Science and Technology), Yang Yang (Nanjing University of Science and Technology)
OptimizationImage
🎯 What it does: This paper proposes a discrete optimization method based on probabilistic group masking (PGM) for efficient parameter allocation in incremental learning.
Probabilistic Interactive 3D Segmentation with Hierarchical Neural Processes
Jie Liu (University of Amsterdam), Efstratios Gavves (University of Amsterdam)
SegmentationDomain AdaptationPoint Cloud
🎯 What it does: This paper proposes a probabilistic interactive 3D segmentation framework NPISeg3D based on hierarchical neural processes.
Probably Approximately Global Robustness Certification
Peter Blohm (TU Wien), SAGAR MALHOTRA
ClassificationOptimizationAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: A sampling method based on ε-net is proposed, providing a probability approximation global robustness (PAG) guarantee for classifiers at high confidence points;
Probing Visual Language Priors in VLMs
Tiange Luo (University of Michigan), Honglak Lee (University of Michigan)
GenerationData SynthesisOptimizationTransformerVision Language ModelDiffusion modelImageTextMultimodalityBenchmark
🎯 What it does: This paper proposes the ViLP benchmark and the Image-DPO method, aimed at examining the over-reliance of visual language models on visual language priors and enhancing their visual reasoning capabilities.
Procurement Auctions via Approximately Optimal Submodular Optimization
Yuan Deng (Google Research), Song Zuo (Google Research)
OptimizationGraph
🎯 What it does: Proposed a procurement auction mechanism that transforms submodular optimization algorithms to satisfy IC, IR, and NAS, and improved the analysis of the distorted greedy algorithm.
ProDiff: Prototype-Guided Diffusion for Minimal Information Trajectory Imputation
Tianci Bu (National University of Defense Technology), Xin Lu
GenerationData SynthesisDiffusion modelTime SeriesSequential
🎯 What it does: A method called ProDiff is proposed, which can complete trajectory missing point interpolation using only endpoint information. It utilizes prototype learning and diffusion models to achieve trajectory reconstruction.
Product of Experts with LLMs: Boosting Performance on ARC Is a Matter of Perspective
Daniel Franzen (Johannes Gutenberg University Mainz), David Hartmann (Lambda)
TransformerLarge Language ModelSupervised Fine-TuningMixture of ExpertsText
🎯 What it does: For the ARC-AGI abstract reasoning task, the authors propose a two-stage reasoning framework that utilizes LLM to generate candidate solutions and re-scores them through various data augmentations and Product of Experts.
Programming Every Example: Lifting Pre-training Data Quality Like Experts at Scale
Fan Zhou (Shanghai Jiao Tong University), Pengfei Liu (Shanghai Jiao Tong University)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposes the PROX framework, which uses small language models to generate and execute refined programs (such as deletion, normalization, etc.) to automatically improve the quality of pre-training data.
Progressive Tempering Sampler with Diffusion
Severi Rissanen (Aalto University), José Miguel Hernández-Lobato (University of Cambridge)
GenerationData SynthesisOptimizationDiffusion modelMultimodalityStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes a sampler that combines Parallel Tempering with diffusion models, called the Progressive Tempering Sampler with Diffusion (PTSD), which generates samples from a high-temperature model to a low-temperature model through temperature guidance.
Progressively Label Enhancement for Large Language Model Alignment
Biao Liu (Southeast University), Xin Geng (Southeast University)
GenerationRecommendation SystemTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: A PLE (Progressively Label Enhancement) framework is proposed and implemented, which fully utilizes all generated data during the LLM alignment training process through dynamic thresholds and label enhancement methods, improving the consistency of the model with human preferences.
Projection Optimization: A General Framework for Multi-Objective and Multi-Group RLHF
Nuoya Xiong (Carnegie Mellon University), Aarti Singh (Carnegie Mellon University)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: A projection optimization framework is proposed to address the nonlinear aggregation problem in multi-objective and multi-agent RLHF, providing both offline and online algorithms that achieve efficient solutions with nearly no training.
Projection Pursuit Density Ratio Estimation
Meilin Wang (Renmin University of China), Zheng Zhang (Renmin University of China)
Tabular
🎯 What it does: A density ratio estimation method based on projection pursuit (ppDRE) is proposed to address the curse of dimensionality faced by traditional linear sieve methods in high-dimensional data.
Promoting Ensemble Diversity with Interactive Bayesian Distributional Robustness for Fine-tuning Foundation Models
Ngoc-Quan Pham (Qualcomm AI Research), Trung Le (Monash University)
OptimizationTransformerSupervised Fine-TuningImageText
🎯 What it does: An interactive Bayesian distribution robust framework (IBDR) is proposed, which enhances the ensemble diversity and robustness of base model fine-tuning through interactions among particles.