ICML 2026 Papers — Page 46
International Conference on Machine Learning · 6554 papers
ProMeCD: Unifying Long-Tailed and Noisy Label Learning via White-Box Control
Yu Zhu (Tianjin University of Commerce)
ClassificationExplainability and InterpretabilityData-Centric LearningSupervised Fine-TuningAuto EncoderContrastive LearningImage
🎯 What it does: Propose the ProMeCD framework, which dynamically adjusts gradients through a self-supervised PI controller to address the problem of long-tailed distributions and noisy labels coexisting.
ProMiSE: Protein Multi-State Evaluation Benchmark in Biological Contexts
Bonjae Ku (Seoul National University), Chaok Seok (Seoul National University)
Protein Structure PredictionBiomedical DataBenchmark
🎯 What it does: Proposed the ProMiSE benchmark for systematically evaluating the performance of protein multi-state structure prediction models under three biological contexts: intrinsic, multi-body induced, and ligand induced.
Prompt Injection as Role Confusion
Charles Ye (Independent), Dylan Hadfield-Menell (Massachusetts Institute of Technology)
Safty and PrivacyExplainability and InterpretabilityAdversarial AttackTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextChain-of-Thought
🎯 What it does: Investigate the internal perception mechanism of LLMs regarding roles (such as <user>, <assistant>, <tool>, etc.), discovering that models determine who is speaking based on text style rather than labels, leading to the success of prompt injection attacks.
Prompt Optimization with Minimal Unlabeled Input via Meta-Reasoning
Yuran Sun (University of Hong Kong), Chuan Wu (University of Hong Kong)
OptimizationComputational EfficiencyData-Centric LearningMeta LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes a self-supervised prompt optimization framework called MR.PEA, which automatically improves the prompts of large language models using meta-reasoning under extremely limited unlabeled inputs (only task descriptions, a single example, and format instructions);
Prompt Reinjection: Alleviating Prompt Forgetting in Multimodal Diffusion Transformers
Yuxuan Yao, Siyu Zhu (Fudan University)
GenerationComputational EfficiencyRepresentation LearningTransformerPrompt EngineeringDiffusion modelImageTextMultimodality
🎯 What it does: This paper investigates the prompt forgetting phenomenon in Multimodal Diffusion Transformer (MMDiT) and proposes a no-training, inference-time feasible Prompt Reinjection technique to help recover the semantic information of deep text features.
Prompt Tuning for CLIP on the Pretrained Manifold
Xi Yang (Guizhou University), Jie Wen (Guizhou University)
ClassificationDomain AdaptationRepresentation LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningImageTextMultimodality
🎯 What it does: Prompt tuning of CLIP leads to feature drift in data-scarce or domain-shift scenarios. This paper proposes the ManiPT framework, which suppresses drift and enhances generalization by performing prompt tuning on the pre-trained feature manifold.
PromptDyG: Test-Time Prompt Adaptation on Dynamic Graphs
Guoguo Ai (Nanjing University of Science and Technology), Guansong Pang (Singapore Management University)
Domain AdaptationAnomaly DetectionRepresentation LearningRecurrent Neural NetworkGraph Neural NetworkPrompt EngineeringContrastive LearningGraphTime SeriesBenchmark
🎯 What it does: Proposed a framework called PromptDyG for test-time adaptation in dynamic graphs, which learns lightweight graph prompts on frozen dynamic graph models and corrects graph structure drift by leveraging unsupervised entropy minimization.
PromptPilot: Game-Theoretic Multi-Agent Prompt Optimization for Segment Anything
Guangze Shi (Taiyuan University Of Technology), Mingqiang Wei (Nanjing University Of Aeronautics And Astronautics)
SegmentationOptimizationTransformerReinforcement LearningAgentic AIPrompt EngineeringImageVideoBiomedical Data
🎯 What it does: Propose PromptPilot, a multi-agent reinforcement learning framework that generates high-quality point prompts for SAM with minimal labeled support, without updating model parameters, enabling automated few-shot segmentation.
PromptRL: Prompt Matters in RL for Flow-Based Image Generation
Fu-Yun Wang (Chinese University of Hong Kong), Taesung Park (Reve)
GenerationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringFlow-based ModelImageText
🎯 What it does: This paper proposes PromptRL, a reinforcement learning framework that jointly trains language models (LMs) with flow-matching image generation models (FMs), leveraging LMs to generate diverse prompts to improve exploration efficiency and suppress prompt overfitting;
ProOPF: Benchmarking and Improving LLMs for Professional-Grade Power Systems Optimization Modeling
Chao Shen (Zhejiang University), Mingyang Sun (Peking University)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextTabularBenchmarkRetrieval-Augmented Generation
🎯 What it does: Constructed a dataset ProOPF-D (12K tiered instances) and a benchmark ProOPF-B (121 expert-annotated OPF test cases) for professional-level power system optimization modeling, and automatically converted natural language descriptions into executable OPF code using LLMs.
ProphetKV: User-Query-Driven Selective Recomputation for Efficient KV Cache Reuse in Retrieval-Augmented Generation
Shihao Wang (Harbin Institute Of Technology), PengFei Wang
RetrievalComputational EfficiencyTransformerTextRetrieval-Augmented Generation
🎯 What it does: Propose a selective recomputation method called ProphetKV based on user query-driven mechanisms, to achieve efficient KV cache reuse in long context Retrieval-Augmented Generation (RAG) scenarios.
Propose, Solve, Verify: Self-Play Through Formal Verification
Alex Wilf (Carnegie Mellon University), Sean Welleck (Carnegie Mellon University)
AI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringText
🎯 What it does: This paper proposes a self-play framework based on formal verification called PSV (Propose, Solve, Verify), which generates code problems in formal specifications and uses a formal verifier to determine whether the generated code meets the specification, thereby providing reliable reward signals for code generation training of large language models; during the iterative process, the model can generate more challenging specifications and improve itself through verified solutions.
ProRL: Effective Reinforcement Learning for Proactive Recommendation via Rectified Policy Gradient Estimation
Hongru Hou (Fudan University), Deqing Yang (Fudan University)
Recommendation SystemTransformerReinforcement LearningRectified FlowSequential
🎯 What it does: ProRL generates guiding paths from users' current preferences to target items in active recommendation systems using reinforcement learning combined with an improved policy gradient, addressing the issue of insufficient path quality.
ProSAR: Prototype-Guided Semantic Augmentation and Refinement for Time Series Contrastive Learning
Caiyi Yang (Shanghai Jiao Tong University), Hongkai Xiong (Shanghai Jiao Tong University)
Anomaly DetectionExplainability and InterpretabilityComputational EfficiencyRepresentation LearningData-Centric LearningTransformerContrastive LearningTabularTime Series
🎯 What it does: This paper proposes the ProSAR framework, which achieves self-supervised contrastive learning for time series through prototype-guided semantic enhancement and refinement.
ProtDBench: A Unified Benchmark of Protein Binder Design and Evaluation
Cong Liu (University of Amsterdam), Wenzhi Xiao (ByteDance Seed)
Drug DiscoveryProtein Structure PredictionTransformerDiffusion modelAuto EncoderGenerative Adversarial NetworkContrastive LearningBiomedical DataBenchmark
🎯 What it does: Proposes ProtDBench, a standardized, throughput-aware framework for protein binder design and evaluation, used to uniformly assess generative methods and analyze the impact of evaluation decisions on results.
Protein Autoregressive Modeling via Multiscale Structure Generation
Yanru Qu (ByteDance Seed), Quanquan Gu (ByteDance Seed)
Protein Structure PredictionTransformerDiffusion modelScore-based ModelFlow-based ModelAuto EncoderBiomedical DataStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Propose PAR — a multi-scale autoregressive protein backbone generation framework that generates 3D backbones progressively from coarse to fine levels;
Protein Circuit Tracing via Cross-layer Transcoders
Darin Tsui (Georgia Institute of Technology), Amirali Aghazadeh (Georgia Institute of Technology)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningDrug DiscoveryTransformerLarge Language ModelPrompt EngineeringAuto EncoderContrastive LearningTextSequentialBiomedical Data
🎯 What it does: Propose the ProtoMech framework, which constructs an interpretable surrogate model through a cross-layer Transcoder (CLT) to discover and utilize the internal computational pathways (circuits) of protein language models (pLMs).
Protein Design with Agent Rosetta: A Case Study for Specialized Scientific Agents
Jacopo Teneggi (Polymathic AI), Siavash Golkar (Polymathic AI)
Drug DiscoveryProtein Structure PredictionTransformerLarge Language ModelAgentic AIPrompt EngineeringBiomedical Data
🎯 What it does: Built an LLM-driven intelligent agent, Agent Rosetta, capable of interacting with the Rosetta simulation software, automatically generating, and iteratively optimizing protein sequences.
Protein Fold Classification at Scale: Benchmarking and Pretraining
Dexiong Chen (Max Planck Institute of Biochemistry), Karsten Borgwardt (Max Planck Institute of Biochemistry)
ClassificationProtein Structure PredictionGraph Neural NetworkTransformerSupervised Fine-TuningAuto EncoderContrastive LearningGraphBiomedical DataBenchmark
🎯 What it does: Proposed TEDBench, a large-scale non-redundant protein fold classification benchmark, and evaluated various structural and sequence models on it.
Protein Language Model Embeddings Improve Generalization of Implicit Transfer Operators
Panagiotis Antoniadis (University of Copenhagen), Ole Winther (University of Copenhagen)
Protein Structure PredictionTransformerLarge Language ModelFlow-based ModelBiomedical Data
🎯 What it does: Propose a transferable implicit transition operator (TITO) based on coarse-grained Cα trajectories, and enhance data efficiency and generalization ability for unseen protein systems by incorporating protein language model (pLM) embeddings, structural embeddings, and annotations generated by LLM.
Proteo-R1: Reasoning Foundation Models for De Novo Protein Design
Fang Wu (Stanford University), Yejin Choi (Stanford University)
Drug DiscoveryProtein Structure PredictionTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsDiffusion modelContrastive LearningMultimodalityBiomedical DataRetrieval-Augmented Generation
🎯 What it does: Propose a dual-expert framework called Proteo-R1, which uses a multimodal large language model as the understanding expert to first perform functional residue reasoning on antibody-antigen complexes; subsequently, the inferred key residue identities and embeddings are injected into an AF3-based diffusion generator to complete the co-design of the antibody CDR.
Proteus: Lookup-Free Trellis-Coded Quantization by Lattice-Breaking Compute Codes for 2-Bit LLMs
Zhengwu Yang (Baidu Inc), Dianhai Yu (Baidu Inc)
Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelDiffusion modelAuto EncoderContrastive LearningText
🎯 What it does: Designed and implemented Proteus, a fully lookup-table-free, bitshift trellis-based low-bit quantization framework for 2-bit weight post-training quantization (PTQ) of large language models.
ProtocolBench: Which LLM MultiAgent Protocol to Choose?
Hongyi Du (University of Illinois Urbana-Champaign), Jiaxuan You (University of Illinois Urbana-Champaign)
Autonomous DrivingOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringTextTabularBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes ProtocolBench, a benchmark for systematically evaluating the performance of LLM multi-agent protocols across four dimensions: task success, latency, byte overhead, and fault tolerance, and designs a Constraint-aware ProtocolRouter capable of dynamically selecting the most suitable protocol based on scenario requirements and runtime signals;
ProtoKV: Streaming Video Understanding under Delayed Query with Summary-State Memory
Le Tu Ngoc Minh (KAIST), Dongsu Han (KAIST)
RecognitionComputational EfficiencyTransformerVideo
🎯 What it does: Propose ProtoKV, a constant-sized KV cache mechanism that combines near-window precise storage with remote prototype memory to achieve delayed queries in streaming video understanding;
Prototype Transformer: Towards Language Model Architectures Interpretable by Design
Yordan Yordanov (Vienna University of Technology), Thomas Lukasiewicz (Vienna University of Technology)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose the Prototype Transformer (ProtoT) architecture, replacing the Transformer self-attention mechanism with a prototype-based linear communication module to improve interpretability and support targeted editing
Prototype-Based Test-Time Adaptation of Vision-Language Models
Zhaohong Huang (Xiamen University), Rongrong Ji (Xiamen University)
ClassificationDomain AdaptationTransformerVision Language ModelContrastive LearningImageTextPoint Cloud
🎯 What it does: Proposes a prototype-based test-time adaptation method called PTA, which does not use backpropagation or caching. It leverages class-specific knowledge prototypes to accumulate information in real-time during the test stream, thereby improving the zero-shot performance of vision-language models such as CLIP.
Prototype-Grounded Concept Models for Verifiable Concept Alignment
Stefano Colamonaco (KU Leuven), Giuseppe Marra (KU Leuven)
Explainability and InterpretabilityRepresentation LearningContrastive LearningImage
🎯 What it does: Prototype-Grounded Concept Models (PGCMs) are proposed by associating concepts with visual prototypes, achieving verifiable concept alignment.
Prototype-guided Bilateral Alignment Multimodal Federated Learning
Tianchi Liao (Hong Kong Baptist University), Chuan Chen (Sun Yat-Sen University)
Federated LearningComputational EfficiencyRepresentation LearningTransformerAuto EncoderContrastive LearningMultimodality
🎯 What it does: Propose a multi-modal federated learning framework called MFedPBA, which enables knowledge sharing under model heterogeneity and modal imbalance through bidirectional prototype alignment.
ProtoVAR: Efficient Dataset Distillation via Prototype-Guided Visual Autoregressive Modeling
Mingyu Wang (Zhejiang University), Wei Jiang (Zhejiang University of Water Resources and Electric Power)
ClassificationData SynthesisKnowledge DistillationTransformerDiffusion modelAuto EncoderContrastive LearningImage
🎯 What it does: Propose the ProtoVAR framework, which achieves efficient prototype-guided dataset distillation by utilizing a multi-scale prototype feature library and VAR's logit refinement, generating synthetic data that is both representative and diverse.
Provable Accuracy Collapse of Embedding-Based Representations under Dimensionality Mismatch
Dionysis Arvanitakis (Northwestern University), Yiyuan Luo (University of California Santa Cruz)
Representation LearningContrastive LearningImageText
🎯 What it does: This paper investigates the essential relationship between embedding dimension and accuracy when embedding triplet constraints in Euclidean space, proving that insufficient dimensions lead to an accuracy collapse;
Provable Benefits of RLVR over SFT for Reasoning Models: Learning to Backtrack Efficiently
Stanley Wei (Princeton University), Juno Kim (University of California, Berkeley)
Computational EfficiencyKnowledge DistillationRepresentation LearningGraph Neural NetworkTransformerSupervised Fine-TuningReinforcement LearningGraphChain-of-ThoughtOrdinary Differential Equation
🎯 What it does: Studied the performance differences between supervised fine-tuning (SFT) and reinforcement learning with verifiable rewards (RLVR) during late training in chain-of-thought tasks, and provided theoretical proofs.
Provable Bounds for the Learnability of Sample-Compressible Families from Noisy Samples
Arefe Boushehrian (EPFL), Amir Najafi (Sharif University of Technology)
OptimizationFederated LearningExplainability and InterpretabilityComputational EfficiencyRepresentation LearningData-Centric LearningContrastive Learning
🎯 What it does: This paper studies the PAC learnability of sample compressible distribution families under additive noise, and provides upper bounds on the sample complexity under ℓ₂ and TV distances.
Provable Sample Efficiency of Curriculum Post-Training for Transformer Reasoning
Dake Bu (City University of Hong Kong), Taiji Suzuki (University of Tokyo)
Computational EfficiencyRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextSequentialChain-of-Thought
🎯 What it does: Proposed a Transformer inference method based on curriculum post-training, and provided provable sample efficiency.
Provable Training Data Identification for Large Language Models
Zhenlong Liu (Southern University of Science and Technology), Hongxin Wei (Southern University of Science and Technology)
Safty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelContrastive LearningTextMultimodality
🎯 What it does: This paper proposes a training data identification method for large-scale language models called PTDI, addressing the issue of strictly controlling error rates in high-risk scenarios.
Provably Convergent Actor-Critic for MARL through Risk-aversion
Yizhou Zhang (California Institute of Technology), Eric Mazumdar (California Institute of Technology)
Reinforcement Learning
🎯 What it does: This paper proposes a multi-agent reinforcement learning (MARL) convergence algorithm that achieves the globally optimal stationary policy in infinite-horizon discounted generalized Markov games through risk-averse quantile response equilibrium (RQE).
Provably Data-driven Lagrangian Relaxation for Mixed Integer Linear Programming
TUNG QUOC LE, Viet Anh Nguyen (Chinese University of Hong Kong)
OptimizationTabular
🎯 What it does: In mixed integer linear programming, a data-driven Lagrangian relaxation learning framework is proposed;
Provably Data-driven Multiple Hyper-parameter Tuning with Structured Loss Function
TUNG QUOC LE, Viet Anh Nguyen (Chinese University Of Hong Kong)
OptimizationHyperparameter SearchData-Centric LearningContrastive Learning
🎯 What it does: Proposed a theoretical framework for multi-dimensional hyperparameter tuning, and provided upper and lower bounds on sample complexity.
Provably Efficient Policy-Reward Co-Pretraining for Adversarial Imitation Learning
Tian Xu (Nanjing University), Yang Yu (Nanjing University)
Reinforcement LearningGenerative Adversarial NetworkContrastive LearningTabular
🎯 What it does: Proposed a joint policy-reward co-pretraining adversarial imitation learning framework called CoPT-AIL, and provided theoretical proof.
Provably Label-Efficient Conformal Prediction
Andrew Ilyas (Carnegie Mellon University), Jiahao Zhang (Carnegie Mellon University)
ClassificationExplainability and InterpretabilityComputational EfficiencyData-Centric LearningContrastive LearningImageText
🎯 What it does: This paper studies how to automatically determine the calibration sample size for split conformal prediction without compromising distribution-free coverage guarantees, in environments where labels are expensive, thereby achieving an optimal balance between total label cost and prediction set size.
Provably Learning Attention with Queries
Satwik Bhattamishra (University of Oxford), Varun Kanade (University of Oxford)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerReinforcement LearningAuto EncoderContrastive Learning
🎯 What it does: The paper studies recovering the weight parameters of a single-head soft maximum attention Transformer model using value queries or membership queries under the black-box query setting, and provides polynomial-time recovery algorithms with query complexity for various scenarios.
Provably Protecting Fine-Tuned LLMs from Training Data Extraction while Preserving Utility
Tom Segal (Ben-Gurion University of Negev), Asaf Shabtai (Ben-Gurion University of Negev)
Safty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: To address training data extraction attacks in fine-tuning large language models, the paper proposes the SCP-Δr algorithm based on the SpaRPS sparsity attribute and base model smoothing, achieving theoretical and empirical protection while maintaining practicality.
Proximal Splitting Methods for Hybrid Differentiable Models
Abdel-Rahim Mezidi (Université Jean Monnet Saint-Etienne), Amaury Habrard (Université Jean Monnet Saint-Etienne)
OptimizationTransformerDiffusion modelTabularTime SeriesPhysics Related
🎯 What it does: Designed a hybrid differentiable model framework based on proximal splitting, combining numerical solvers with neural networks into a trainable 'unrolled' architecture, and verified its performance in PDE simulations.
Proximal-Based Generative Modeling for Bayesian Inverse Problems
Boyang Zhang (University of Chinese Academy of Sciences), Ya-Feng Liu (Beijing University of Posts and Telecommunications)
RestorationSuper ResolutionOptimizationDiffusion modelScore-based ModelImageStochastic Differential Equation
🎯 What it does: This paper proposes a proximal operator-based generative model framework called PGM, which uses the Moreau score instead of the traditional time-dependent posterior score to achieve efficient sampling for Bayesian inverse problems.
Proximal-IMH: Proximal Posterior Proposals for Independent Metropolis–Hastings with Approximate Operators
Youguang Chen (University of Texas at Austin), George Biros (University of Texas at Austin)
OptimizationComputational EfficiencyData-Centric LearningScore-based ModelContrastive LearningGaussian SplattingImagePoint CloudGraphTabularTime SeriesPhysics RelatedStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Propose a Proximal-IMH method, which corrects the bias caused by the approximate operator and improves the acceptance rate by proximal optimization of samples from the approximate posterior in independent Metropolis-Hastings sampling.
Proxy Compression for Language Modeling
Lin Zheng (University of Hong Kong), Lingpeng Kong (University of Hong Kong)
CompressionComputational EfficiencyRepresentation LearningAI Code AssistantTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextSequential
🎯 What it does: Propose a Proxy Compression training scheme, which mixes the compressed sequences generated by an external compressor with the original UTF-8 byte sequences during training, allowing a single language model to learn both representations simultaneously; during inference, only the original byte sequences are used, completely eliminating the dependence on the compressor.
PRPO: Paragraph-level Policy Optimization for Vision-Language Deepfake Detection
Tuan Nguyen (Qatar Computing Research Institute, Hamad Bin Khalifa University), Issa Khalil (Qatar Computing Research Institute, Hamad Bin Khalifa University)
Anomaly DetectionExplainability and InterpretabilityTransformerLarge Language ModelReinforcement LearningVision Language ModelImageTextMultimodality
🎯 What it does: Proposed a multimodal large language model based on paragraph-level reinforcement learning, which can provide interpretable and visually consistent reasoning and decision-making when detecting deepfake images.
Pruning at Initialisation through the lens of Graphon Limit: Convergence, Expressivity, and Generalisation
Hoang Pham (University of Warwick), Long Tran-Thanh (University of Warwick)
Computational EfficiencyKnowledge DistillationRepresentation LearningConvolutional Neural NetworkSpiking Neural NetworkContrastive LearningImage
🎯 What it does: Conduct a Tsunami analysis on the sparse subnetwork of the "initialization pruning" algorithm, proving that its mask converges to a deterministic bipartite Tsunami graph as the width approaches infinity, and thereby providing theoretical expressions for the representational capacity and generalization of sparse networks.
PS-PPO : Prefix-Sampling PPO for Critic-Free RLHF
Doo Hwan Hwang (KAIST), Kee-Eung Kim (KAIST)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: Proposes PS-PPO, a value-free RLHF method that samples prefix cutoff points under each prompt, truncates gradients, and performs importance weighting.
PSBench: Editing Image via GUI Agents in Photoshop
Yinuo Zhang (Harbin Institute of Technology), Dianbo Sui (Harbin Institute of Technology)
TransformerLarge Language ModelAgentic AIPrompt EngineeringVision-Language-Action ModelImageTextMultimodalityBenchmark
🎯 What it does: Proposed the PSBench benchmark for evaluating GUI agents in image editing tasks within Adobe Photoshop, containing 600 manually annotated tasks covering core functions such as layer operations, canvas manipulation, and filters, and providing fine-grained evaluation metrics and the Non-Destructive Editing Consistency (NDEC) metric.
Pseudo-Mallows for Efficient Probabilistic Preference Learning
Sylvia Qinghua Liu (University of Oslo), Ida Scheel (University of Oslo)
Recommendation SystemComputational EfficiencyTabularSequential
🎯 What it does: Propose the Pseudo-Mallows distribution to approximate the Bayesian Mallows model posterior, enabling fast variational inference in preference learning and recommendation.
PSG-Nav: Probabilistic Scene Graph Navigation via Multiverse Decision Making
Rufeng Chen (Hong Kong University of Science and Technology (Guangzhou)), Sihong Xie (Hong Kong University of Science and Technology (Guangzhou))
Robotic IntelligenceGraph Neural NetworkTransformerLarge Language ModelReinforcement LearningTextPoint CloudGraphRetrieval-Augmented Generation
🎯 What it does: Propose Probabilistic Scene Graph Navigation (PSG-Nav), which achieves robust handling of semantic uncertainty in open-vocabulary navigation by constructing a 3D probabilistic scene graph and combining multi-universe decision-making with experience calibration.
PSMix: Robust Point Cloud Recognition through Spectral Domain Mixing
Xin Wei (Xidian University), Xinbo Gao (Xidian University)
RecognitionGraph Neural NetworkDiffusion modelGenerative Adversarial NetworkContrastive LearningPoint Cloud
🎯 What it does: The paper proposes a point cloud data augmentation framework called PSMix based on spherical harmonic transformation, which can perform rotation-aware hierarchical mixing in the frequency domain to generate structurally consistent and diverse augmented samples.
PsumQuant: In-line Post-training Partial Sum Quantizer for Energy Efficient NPU Inference
Sangwoo Hwang (Korea University), Jaeha Kung (Korea University)
Computational EfficiencyConvolutional Neural NetworkTransformerImageText
🎯 What it does: This paper proposes a PSumQuant method tailored for NPU, achieving input-aware low-precision quantization of partial sums, significantly reducing the bit width and energy consumption of intermediate accumulations;
pTNAS: Progressive Neural Architecture Search for Tabular Data
Naili Xing (National University of Singapore), Beng Chin Ooi (Zhejiang University)
Neural Architecture SearchReinforcement LearningTabularBenchmark
🎯 What it does: Proposed a progressive neural architecture search framework called pTNAS for tabular data, adopting a filter-refinement strategy and introducing a budget-aware coordinator;
Pull Requests as a Training Signal for Repo-Level Code Editing
Qinglin Zhu (King's College London), Yeyun Gong (Microsoft Research Asia)
Data-Centric LearningAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Propose the CleanPR framework, which constructs a large-scale reproducible Search/Replace edit dataset using cleaned and verified GitHub PRs, and enhances repository-level code editing capabilities through in-training and Agentless-aligned SFT.
PULSE: Generative Phase Evolution for Non-Stationary Time Series Forecasting
Yangyou Liu (Sichuan University), Yuankai Wu (Sichuan University)
OptimizationComputational EfficiencyTransformerDiffusion modelFlow-based ModelAuto EncoderGenerative Adversarial NetworkContrastive LearningTabularTime SeriesBenchmark
🎯 What it does: Propose the PULSE framework, which utilizes physics-driven phase-anchored decomposition and generative phase routing to address the phase forgetting problem under the non-stationarity of time series.
Push, Pop, Parallelize: Stack-Augmented Linear Attention via the Delta Rule
Anh T Nguyen (University of Illinois Chicago), Bing Liu (University of Illinois Chicago)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelMixture of ExpertsContrastive LearningTextBenchmark
🎯 What it does: Introduces DeltaStack, a linear attention architecture that embeds a differentiable stack into DeltaNet, enabling parallel operations on the stack.
Pushing Forward Pareto Frontiers of Proactive Agents with Behavioral Agentic Optimization
Yihang Yao (Carnegie Mellon University), Ding Zhao (Carnegie Mellon University)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIPrompt EngineeringWorld ModelTextBenchmark
🎯 What it does: Proposes the Behavioral Agentic Optimization (BAO) framework, which combines multi-round behavioral reinforcement and behavioral regularization to optimize the task performance of proactive LLM agents and reduce user interaction costs.
Pushing the Boundaries of Natural Reasoning: Interleaved Bonus from Formal-Logic Verification
Chuxue Cao (Hong Kong University of Science and Technology), Sirui Han (Hong Kong University of Science and Technology)
Data SynthesisExplainability and InterpretabilityComputational EfficiencyAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextMultimodalityChain-of-Thought
🎯 What it does: Propose a framework that integrates formal logic verification in real-time into the reasoning process of large language models, and improves the model's reasoning consistency and correctness through a two-phase training approach (supervised fine-tuning + reinforcement learning).
Pushing the Limits of Block Rotations in Post-Training Quantization
Sai Sanjeet (Advanced Micro Devices Inc), Nicholas J. Fraser (Advanced Micro Devices Inc)
Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: Propose the PeRQ framework, which first performs permutation on activation to balance the block-wise ℓ1 quality, and then uses block Hadamard rotation and quantization to achieve accuracy improvement in PTQ.
PuzzleMoE: Efficient Compression of Large Mixture-of-Experts Models via Sparse Expert Merging and Bit-packed inference
Yushu Zhao (Tsinghua University), Minjia Zhang (University of Illinois Urbana Champaign)
CompressionComputational EfficiencyTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: Propose a method called PuzzleMoE, which efficiently compresses large-scale MoE language models through fine-grained expert merging and bit-level compression, while improving inference speed while maintaining high accuracy.
PVDepth: Panoramic Video Depth Estimation via Geometry-Aware Spatiotemporal Adaptation
Chuanxin Song (Peking University), Peixi Peng (Peking University)
Depth EstimationAutonomous DrivingTransformerDiffusion modelContrastive LearningImageVideo
🎯 What it does: Propose a depth estimation framework for omnidirectional videos, PVDepth, to address the ERP geometric distortion and spatiotemporal inconsistency issues.
PyHealth 2.0: A Comprehensive Open-Source Toolkit for Accessible and Reproducible Clinical Deep Learning
John Wu (University of Illinois Urbana-Champaign), Jimeng Sun (University of Illinois Urbana-Champaign)
Explainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerPrompt EngineeringAuto EncoderContrastive LearningMultimodalityTabularBiomedical DataElectronic Health RecordsBenchmark
🎯 What it does: Created PyHealth 2.0, a deep learning toolkit that integrates multimodal data, tasks, models, and evaluation tools to achieve reproducibility and accessibility in clinical AI.
PyVision-RL: Forging Open Agentic Vision Models via RL
Shitian Zhao (Shanghai AI Lab), Chen Wei (Rice University)
Autonomous DrivingOptimizationFederated LearningExplainability and InterpretabilityComputational EfficiencyRepresentation LearningMeta LearningDrug DiscoveryAI Code AssistantReinforcement Learning from Human FeedbackNeural Architecture SearchTransformerSupervised Fine-TuningReinforcement LearningAgentic AIPrompt EngineeringVision Language ModelVision-Language-Action ModelImageVideoTextMultimodalityRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed the PyVision-RL framework, a tool for training open-source multimodal models to achieve tool use and multi-round reasoning for images and videos through reinforcement learning;
Q-CLIP: Unleashing the Power of Vision-Language Models for Video Quality Assessment through Unified Cross-Modal Adaptation
Yachun Mi (Harbin Institute of Technology), Shaohui Liu (Harbin Institute of Technology)
Computational EfficiencyRepresentation LearningTransformerSupervised Fine-TuningPrompt EngineeringVision Language ModelContrastive LearningVideoTextMultimodality
🎯 What it does: This paper proposes Q-CLIP, a unified cross-modal adaptation framework fully based on contrastive vision-language models, for achieving efficient and accurate video quality assessment.
Q-Delta: Beyond Key–Value Associative State Evolution
Sumin Park (Korea Advanced Institute of Science and Technology), Noseong Park (Korea Advanced Institute of Science and Technology)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Propose a new linear attention update rule called Q-Delta, which integrates query information into state evolution, thereby achieving more flexible memory writing and retrieval.
Q-DiT4SR: Exploration of Detail-Preserving Diffusion Transformer Quantization for Real-World Image Super-Resolution
Xun Zhang (Shanghai Jiao Tong University), Yulun Zhang (Shanghai Jiao Tong University)
Super ResolutionTransformerDiffusion modelImage
🎯 What it does: A post-training quantization framework called Q-DiT4SR for Diffusion Transformer is designed to achieve efficient Real-World image super-resolution.
Q-Flow: Stable and Expressive Reinforcement Learning with Flow-based Policy
JaeHyeok Doo (KAIST AI), Minjoon Seo (KAIST AI)
Reinforcement LearningFlow-based ModelTabularTime SeriesSequentialOrdinary Differential Equation
🎯 What it does: Propose the Q-Flow framework, which explicitly propagates terminal rewards of flow models to intermediate latent states, achieving stable policy optimization without BPTT by utilizing gradients from intermediate values.
Q-SAM: Unlocking Sharpness-Aware Minimization for Generalization in Offline Reinforcement Learning
Da Wang (Shanxi University), Jiye Liang (Shanxi University)
Reinforcement LearningTabularTime Series
🎯 What it does: Propose Q-SAM, an additive sharpness minimization framework for offline reinforcement learning, which dynamically balances sharpness and sample weights using the upper bound of Q-values;
Q-Sched: Pushing the Boundaries of Few-Step Diffusion Models with Quantization-Aware Scheduling
Natalia Frumkin (University of Texas at Austin), Diana Marculescu (University of Texas at Austin)
GenerationTransformerDiffusion modelContrastive LearningImageText
🎯 What it does: For diffusion models with extremely low number of steps, a method called Quantization-Aware Scheduling (Q-Sched) is proposed, which significantly improves the image quality of the compressed model by introducing learnable pre-adjustment coefficients in the sampler, while keeping the model weight quantization unchanged.
Q-Tab: Quantized Tabular Data Generator
Julian Wustl (University of Augsburg), Claudius Schnörr
Data SynthesisTransformerDiffusion modelGenerative Adversarial NetworkContrastive LearningTabular
🎯 What it does: This paper proposes a table data generator called Q-Tab, which utilizes residual Lookup-Free Quantization (LFQ) to jointly quantize numerical features into an expandable discrete code space, and uses a BERT-style MLM to learn the joint distribution of these discrete codes, thereby achieving high-quality synthetic table data.
QEDBench: Quantifying the Alignment Gap in Automated Evaluation of University-Level Mathematical Proofs
Santiago Gonzalez (Yale University), Quanquan C. Liu (Yale University)
Large Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Constructed the QEDBENCH benchmark for evaluating and aligning automatic assessment of university-level mathematical proofs
QHyer: Q-conditioned Hybrid Attention-mamba Transformer for Offline Goal-conditioned RL
Xing Lei (Xi'an Jiaotong University), Donglin Wang (Westlake University)
TransformerReinforcement LearningFlow-based ModelTabularSequentialBenchmark
🎯 What it does: Propose a sequence modeling framework called QHyer for offline goal-conditioned reinforcement learning, which combines NF for estimating Q-values and the Hybrid Attention-Mamba structure to address the RTG dependency on trajectories and challenges in time modeling.
QiMeng-ChipV-RTL: Exploiting Information Locality for IP-level Verilog Generation
Hanqi Lyu (University of Science and Technology of China), Yunji Chen (SKL of Processors, Institute of Computing Technology, CAS)
AI Code AssistantTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed a multi-agent framework called ChipV-RTL, which utilizes the principle of information locality to decompose the generation of long documents and long code into subtasks of short documents and short code, achieving IP-level Verilog code generation automatically.
QiMeng-LibBench: Benchmarking LLM Agents for Library-Scale Cross-Architecture Migration
Weijia Li (Intelligent Software Research Center, Institute of Software, CAS), Ling Li (Intelligent Software Research Center, Institute of Software, CAS)
Computational EfficiencyAI Code AssistantTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes the QM-LibBench benchmark, which investigates the effectiveness of LLM agents in migrating across ISA high-performance libraries.
QiMeng-PerceptOS: Semantic-Aware Kernel Optimization for OS-Intensive Workloads via Hardware-Software Alignment
Huilai Chen (Chinese Academy of Sciences), Yunji Chen (Chinese Academy of Sciences)
OptimizationTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Proposed QiMeng-PerceptOS, a perception-driven closed-loop large language model (LLM) framework based on perception, used to automatically optimize Linux kernel configurations for specific applications, significantly improving the performance of various workloads.
QPKO: Differentiable QP-Embedded Deep Koopman Framework for Modeling Nonlinear Systems
Runze Tian (Xi'an Jiaotong University), Peng Kou (Xi'an Jiaotong University)
OptimizationComputational EfficiencyRobotic IntelligenceRecurrent Neural NetworkTransformerReinforcement LearningAuto EncoderContrastive LearningTime SeriesSequentialStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Propose a differentiable QP embedded deep Koopman framework (QPKO), which maps observable functions to a global linear model through a differentiable QP module, enabling end-to-end training.
QPoint: End-to-End Lightweight Point Cloud Processing via Robust Quaternion Feature Learning
Zhiming Zhou (Anhui University), Ajmal Saeed Mian (University Of Western Australia)
ClassificationSegmentationTransformerContrastive LearningPoint Cloud
🎯 What it does: Proposed a lightweight point cloud processing framework called QPoint, which achieves local feature rotation robustness by learning quaternion rotations, and realizes global rotation invariance through quaternion attention, completing tasks such as point cloud classification, segmentation, and few-shot learning.
QTALE: Quantization-Robust Token-Adaptive Layer Execution for LLMs
KangHyun Noh, Yulhwa Kim (Sungkyunkwan University)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: Propose the QTALE framework, combining token-adaptive layer execution with low-bit quantization to achieve efficient LLM inference.
Quadratically Regularized Optimal Transport: Localization Bounds and Affine Case Analysis
Long Nguyen Chi, Binh Nguyen
OptimizationScore-based ModelContrastive LearningTabularTime Series
🎯 What it does: This paper studies quadratic regularized optimal transport (QOT), and presents the geometric aggregation properties of its solution when approximating Monge transport;
Quant VideoGen: Auto-Regressive Long Video Generation via 2-Bit KV-Cache Quantization
Haocheng Xi (University of California, Berkeley), Kurt Keutzer (University of California, Berkeley)
GenerationComputational EfficiencyTransformerLarge Language ModelVision-Language-Action ModelDiffusion modelContrastive LearningVideo
🎯 What it does: Proposes Quant VideoGen (QVG), an untrained KV-Cache quantization framework that significantly reduces the memory consumption of autoregressive video diffusion models while maintaining video quality;
Quantifying and Optimizing Simplicity via Polynomial Representations
Tianren Zhang (Tsinghua University), Feng Chen (Tsinghua University)
OptimizationExplainability and InterpretabilityRepresentation LearningAuto EncoderContrastive LearningImageTextTime Series
🎯 What it does: Proposed a function space succinctness metric and differentiable regularizer based on polynomial approximation, and verified that it can better predict and improve model generalization.
Quantifying Biases in LLM-as-a-Judge Evaluations
Magda Dubois (UK AI Security Institute), Cozmin Ududec (UK AI Security Institute)
Explainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose a statistical framework based on Bayesian generalized linear models to quantify and correct various systematic biases in LLM evaluators during assessment tasks (such as self-bias, length bias, writing style bias, etc.)
Quantifying Cross-Domain Knowledge Distillation in the Presence of Domain Shift
Xiangchao Li (University of Science and Technology of China), Xin Tong (University of Hong Kong)
Domain AdaptationKnowledge DistillationTabular
🎯 What it does: Theoretically analyze cross-domain knowledge distillation (KD), providing a high-dimensional asymptotic excess risk analysis when there is model shift and covariate shift between the source and target domains; and verify the theoretical predictions through experiments.
Quantifying Error Propagation and Model Collapse in Diffusion Models
Nail B. Khelifa, Ramji Venkataramanan (University of Cambridge)
GenerationData SynthesisDiffusion modelScore-based ModelImageTabularStochastic Differential Equation
🎯 What it does: This paper studies how errors propagate across generations and lead to model degradation when training diffusion models with a mixture of partial real samples and synthetic samples during recursive training.
Quantifying Frontier LLM Capabilities for Container Sandbox Escape
Rahul Marchand (University of Oxford), Harry Coppock (UK AI Security Institute)
Safty and PrivacyTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes the SANDBOXESCAPEBENCH benchmark for security evaluation of LLMs' escape capabilities within container sandboxes.
Quantifying LLM Attention-Head Stability: Implications for Circuit Universality
Karan Bali (Mila -Quebec Artificial Intelligence Institute), Danilo Bzdok (Mila -Quebec Artificial Intelligence Institute)
OptimizationExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
🎯 What it does: This paper quantifies and analyzes the stability of attention heads in Transformer models across different layers and model sizes by comparing attention heads from multiple randomly initialized identical Transformer models.
Quantifying the Generalization Gap in Seizure Detection: A Large-Scale Empirical Benchmark via the SzCORE Challenge
Jonathan Dan (EPFL), David Atienza (EPFL)
Anomaly DetectionConvolutional Neural NetworkRecurrent Neural NetworkTransformerAuto EncoderGenerative Adversarial NetworkContrastive LearningTime SeriesBiomedical DataBenchmark
🎯 What it does: This paper organizes the SzCORE challenge, collecting 28 advanced algorithms to conduct a rigorous blind evaluation on 4,360 hours of EMU continuous EEG, assessing their generalization ability on real clinical data.
Quantifying the noise sensitivity of the Wasserstein metric for images
Erik Lager (Tel Aviv University), Amit Moscovich (Tel Aviv University)
Data SynthesisImageBiomedical Data
🎯 What it does: This paper studies the sensitivity of the Wasserstein metric (especially the signed Wasserstein distance) at the pixel level to noise, and provides theoretical error bounds for the Gaussian noise model.
Quantifying the Salience of Geo-Cultural Values for Pluralistic Safety Alignment
Arkadiy Saakyan (Columbia University), Lora Aroyo (Google DeepMind)
Safty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Conduct a meta-analysis of the geographic and cultural attributes of AI safety assessment datasets, quantify the impact of cultural regions on safety evaluations using a multi-level model, and evaluate the performance of LLMs in mimicking multi-cultural assessments and identifying culturally sensitive items.
Quantifying the Uncertainty of Foundation Models with Singular Value Ensembles
Mehmet Ozgur Turkoglu (Agroscope), Helge Aasen (Agroscope)
ClassificationRecognitionImage TranslationExplainability and InterpretabilityComputational EfficiencyRepresentation LearningConvolutional Neural NetworkTransformerLarge Language ModelMixture of ExpertsContrastive LearningImageText
🎯 What it does: Propose Singular Value Ensemble (SVE), a parameter-free significant depth ensemble that achieves uncertainty quantification by freezing the pre-trained weights of singular vectors and only learning the singular values of each member.
Quantile-Free Uncertainty Quantification in Graph Neural Networks
Soyoung Park (Chungnam National University), Sungsu Lim (Chungnam National University)
Graph Neural NetworkGraph
🎯 What it does: Proposed a graph neural network framework called QpiGNN, which does not require quantile inputs, to achieve confidence interval estimation for node regression tasks;
Quantitative Estimation of Target Task Performance from Unsupervised Pretext Task in Semi/Self-Supervised Learning
Lin-Han Jia (Nanjing University), Yu-Feng Li (Nanjing University)
ClassificationRepresentation LearningConvolutional Neural NetworkTransformerAuto EncoderContrastive LearningImage
🎯 What it does: Propose a low-cost method to quantitatively estimate the impact of unsupervised pre-training tasks on the performance of downstream tasks.
Quantum Algorithms for Triangle Cut Sparsification
Shan Jiang (University of Science and Technology of China), Pan Peng (University of Science and Technology of China)
OptimizationComputational EfficiencyPhysics Related
🎯 What it does: This paper proposes a quantum algorithm for constructing triangle cut sparsifiers and provides a quantum accelerated method for triangle enumeration.
Quantum latent distributions in deep generative models
Omar Bacarreza (ORCA Computing), William R. Clements (ORCA Computing)
GenerationData SynthesisDrug DiscoveryDiffusion modelScore-based ModelFlow-based ModelGenerative Adversarial NetworkOptical FlowGraphTabularBiomedical DataPhysics RelatedStochastic Differential Equation
🎯 What it does: Studied the role of quantum latent distributions in deep generative models, provided theoretical foundations, and verified their advantages through GAN experiments.
Quantum Robust Inner Minimization for Reinforcement Learning with Quadratic Speed-Up in Query Complexity
Hyun Kyu Lee (Ulsan National Institute of Science and Technology), Sung Whan Yoon (Ulsan National Institute of Science and Technology)
Reinforcement LearningTabularTime Series
🎯 What it does: Developed and evaluated a quantum-accelerated robust reinforcement learning method, QRIM, which achieves fast search over the minimization (worst-case) within the uncertainty set through quantum access to the environment, thereby reducing the query complexity from O(|U|) to O(√|U|) while keeping the external RL pipeline unchanged.
QuantumBoost: A lazy, yet fast, quantum algorithm for learning with weak hypotheses
Amira Abbas (Google Quantum AI), Ronald de Wolf (QuSoft)
OptimizationExplainability and InterpretabilityComputational EfficiencyRepresentation LearningPhysics Related
🎯 What it does: Proposed a quantumized weak learner boosting algorithm named QuantumBoost.
QuantWear: Quantum-scale Wear Particle Detection for Jet Engine Diagnosis
Zheng Wang (Hunan University), Yuanqing Zheng (Hong Kong Polytechnic University)
Object DetectionDepth EstimationAnomaly DetectionConvolutional Neural NetworkTransformerDiffusion modelContrastive LearningGaussian SplattingOptical FlowImageVideoPhysics Related
🎯 What it does: Developed and validated the QuantWear system, which utilizes quantum optics for sodium-potassium spectral filtering, YOLOv8 detection, and shape reconstruction techniques, enabling accurate counting and three-dimensional reconstruction of engine wear particles in high-temperature jets.
QuArch: A Benchmark for Evaluating LLM Reasoning in Computer Architecture
Shvetank Prakash (Harvard University), Vijay Janapa Reddi (Google DeepMind)
TransformerLarge Language ModelPrompt EngineeringTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Constructed the QUARCH benchmark for systematically evaluating the knowledge and reasoning capabilities of large language models in the field of computer architecture.
Quartet II: Accurate LLM Pre-Training in NVFP4 by Improved Unbiased Gradient Estimation
Andrei Panferov (Institute of Science and Technology Austria), Dan Alistarh (Institute of Science and Technology Austria)
OptimizationComputational EfficiencyData-Centric LearningTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Proposed a full NVFP4 quantization training framework named Quartet II for end-to-end pre-training of large-scale language models.