ICLR 2026 Papers — Page 37
International Conference on Learning Representations · 5356 papers
Prompt Curriculum Learning for Efficient LLM Post-Training
Zhaolin Gao (Meta Superintelligence Labs), Liang Tan (Meta Superintelligence Labs)
Computational EfficiencyLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: In the post-training of large language models, an efficient reinforcement learning framework based on prompt curriculum learning (PCL) is proposed, which significantly accelerates convergence by using an online learned value model to select medium difficulty prompts;
Prompt-MII: Meta-Learning Instruction Induction for LLMs
Emily Xiao (Carnegie Mellon University), Graham Neubig (Carnegie Mellon University)
Meta LearningTransformerReinforcement LearningPrompt EngineeringText
🎯 What it does: Train a meta-learning instruction generator to enable LLMs to generate concise task instructions in a single forward pass, replacing traditional in-context learning or manual prompts.
Prompt-Robust Vision-Language Models via Meta-Finetuning
Haohui Liang (Beijing Normal-Hong Kong Baptist University), Cees G. M. Snoek (University of Amsterdam)
ClassificationMeta LearningTransformerPrompt EngineeringVision Language ModelContrastive LearningImageText
🎯 What it does: Developed Promise, a scheme that enhances the robustness of vision-language models to variations in natural language prompts through a meta-learning framework.
PromptHub: Enhancing Multi-Prompt Visual In-Context Learning with Locality-Aware Fusion, Concentration and Alignment
Tianci Luo (Tsinghua University), Shu-Tao Xia (Tsinghua University)
Object DetectionSegmentationTransformerPrompt EngineeringAuto EncoderContrastive LearningImageRetrieval-Augmented Generation
🎯 What it does: Propose the PromptHub framework in visual context learning, using multi-prompt fusion to enhance performance in image segmentation, detection, and coloring tasks.
ProofBridge: Auto-Formalization of Natural Language Proofs in Lean via Joint Embeddings
Prithwish Jana (Georgia Institute of Technology), Vijay Ganesh (Georgia Institute of Technology)
AI Code AssistantTransformerLarge Language ModelContrastive LearningTextRetrieval-Augmented Generation
🎯 What it does: Proposes PROOFBRIDGE, which can automatically convert natural language theorems and proofs into Lean 4 formal code in one go, integrating retrieval and iterative repair processes.
ProofFlow: A Dependency Graph Approach to Faithful Proof Autoformalization
Rafael Medeiros Cabral (Huawei Celia Team), SHEN XIN (Huawei Celia Team)
AI Code AssistantTransformerLarge Language ModelTextGraphBenchmark
🎯 What it does: Developed a three-stage pipeline, PROOFFLOW, to convert natural language proofs into structured Lean code, and introduced the PROOFSCORE evaluation metric along with the PROOFFLOWBENCH dataset.
ProofOptimizer: Training Language Models to Simplify Proofs without Human Demonstrations
Alex Gu (MIT), Aram H. Markosyan (Axiom Math)
OptimizationAI Code AssistantTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Built and trained a language model named ProofOptimizer to automate the simplification of long and redundant formal proofs in the Lean environment that have already been trained using reinforcement learning (RL).
Propaganda AI: An Analysis of Semantic Divergence in Large Language Models
Nay Myat Min (Singapore Management University), Jun Sun (Singapore Management University)
Explainability and InterpretabilityAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Studied a black-box auditing framework called RAVEN for detecting semantic biases and propaganda consistency in large language models induced by high-level concepts (e.g., public figures, ideologies).
PropensityBench: Evaluating Latent Safety Risks in Large Language Models via an Agentic Approach
Udari Madhushani Sehwag (Scale AI), Furong Huang (University of Maryland)
Safty and PrivacyLarge Language ModelAgentic AITextBenchmark
🎯 What it does: Proposed PropensityBench, an agentic method to evaluate the propensity of large language models when using high-risk tools, covering 5,874 tasks across four high-risk domains: cybersecurity, self-replication, life safety, and chemical safety.
Proper Velocity Neural Networks
Ziheng Chen (University of Trento), Nicu Sebe (University of Trento)
ClassificationImageGraphBiomedical Data
🎯 What it does: Proposes a neural network (PVNN) built using an unconstrained Proper Velocity (PV) space, constructing PV versions of polynomial logistic regression (MLR), fully connected layers (FC), convolutional layers, activation functions, and normalization layers by deriving its complete Riemannian geometry toolbox;
ProPerSim: Developing Proactive and Personalized AI Assistants through User-Assistant Simulation
Jiho Kim (Korea Advanced Institute of Science and Technology), Edward Choi (Seoul National University)
Recommendation SystemData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Investigated an AI assistant that simultaneously exhibits proactivity and personalization, learning from user agents in simulated home environments to provide personalized recommendations.
Property-Driven Protein Inverse Folding with Multi-Objective Preference Alignment
Junqi Liu (Peking University), Jian Tang (Mila - Québec AI Institute)
Protein Structure PredictionGraph Neural NetworkReinforcement LearningBiomedical Data
🎯 What it does: Proposes the ProtAlign framework, which aligns multi-objective preferences on pre-trained protein inverse folding models to enhance developability features such as solubility and thermal stability while maintaining design feasibility.
ProRe: A Proactive Reward System for GUI Agents via Reasoner–Actor Collaboration
Gaole Dai (Nanyang Technological University), Lili Qiu (Microsoft Research)
TransformerLarge Language ModelAgentic AISequentialChain-of-Thought
🎯 What it does: Proposes PRORE, an active reward system for GUI agents that combines a general reasoner and domain-specific evaluator agents for state detection and reward determination.
ProReGen: Progressive Residual Generation under Attribute Correlations
Ruby Shrestha (Rochester Institute of Technology), Linwei Wang (Rochester Institute of Technology)
GenerationData SynthesisDiffusion modelAuto EncoderGenerative Adversarial NetworkImage
🎯 What it does: Propose the ProReGen method, which decomposes attribute correlations through Robinson's partialling-out into orthogonal inputs and adopts a two-stage progressive residual generation, significantly improving the generation accuracy of minority samples.
PROS: Towards Compute-Efficient RLVR via Rollout Prefix Reuse
Baizhou Huang (Wangxuan Institute of Computer Technology, Peking University), Xiaojun Wan (Wangxuan Institute of Computer Technology, Peking University)
Computational EfficiencyReinforcement LearningText
🎯 What it does: Propose the PROS framework, which reduces redundant inference steps by reusing high-quality prefixes from historical rollouts in RLVR training to generate Augmented Queries;
ProSafePrune: Projected Safety Pruning for Mitigating Over-Refusal in LLMs
Zijun Chen (Hefei University of Technology), Richang Hong (Hefei University of Technology)
Safty and PrivacyComputational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: To address the problem of over-rejection in large language models (LLMs), the ProSafePrune low-rank parameter pruning method is proposed, which directly removes excessive 'harmful encoding' within the model for pseudo-harmful instructions, thereby reducing false rejection rates while maintaining safe rejection capabilities.
Prosperity before Collapse: How Far Can Off-Policy RL Reach with Stale Data on LLMs?
Haizhong Zheng (Carnegie Mellon University), Beidi Chen (Carnegie Mellon University)
Large Language ModelReinforcement LearningTextBenchmark
🎯 What it does: Proposed an offline reinforcement learning algorithm named M2PO, which can maintain stable performance using extremely outdated data (data up to 256 model updates ago) in large language model training;
ProstaTD: Bridging Surgical Triplet from Classification to Fully Supervised Detection
Yiliang Chen, Jing Qin (Hong Kong Polytechnic University)
Object DetectionKnowledge DistillationConvolutional Neural NetworkTransformerVideoBiomedical DataBenchmark
🎯 What it does: This study proposes the ProstaTD dataset and the TDnet benchmark model, constructing the first fully supervised triplet detection resource in surgery, covering complete videos of multi-institutional robotic prostatectomy, precise bounding boxes, and temporary boundaries;
ProTDyn: A Foundation Protein Language Model for Thermodynamics and Dynamics Generation
Yikai Liu (Purdue University), Guang Lin (Purdue University)
GenerationDrug DiscoveryTransformerLarge Language ModelTime SeriesBiomedical Data
🎯 What it does: What was done: Proposed ProTDyn, a unified multi-task autoregressive Transformer model capable of generating protein thermodynamic equilibrium conformation ensembles and multi-scale dynamic trajectories within the same framework.
Protection against Source Inference Attacks in Federated Learning
Andreas Athanasiou (Technical University of Delft), Catuscia Palamidessi (Inria)
Federated LearningConvolutional Neural NetworkImage
🎯 What it does: A defense scheme against source inference attacks in federated learning is proposed, involving parameter-level shuffling combined with Residue Number System (RNS) and unit encoding, and demonstrating a reconstruction attack that can break traditional shuffling.
Protein Structure Tokenization via Geometric Byte Pair Encoding
Michael Sun (MIT), Marinka Zitnik (Apple)
Representation LearningProtein Structure PredictionTransformerBiomedical Data
🎯 What it does: Designed and implemented a geometry-based byte pair encoding (GEOBPE) that segments continuous protein backbones into discrete hierarchical structural tokens.
ProteinAE: Protein Diffusion Autoencoders for Structure Encoding
Shaoning Li (CUHK), Pheng-Ann Heng (CUHK)
Protein Structure PredictionTransformerDiffusion modelFlow-based ModelAuto EncoderBiomedical Data
🎯 What it does: Proposed the PROTEINAE autoencoder, which directly maps protein backbone coordinates to a continuous compact latent space, and trained a latent diffusion model for structure generation based on this.
ProtoKV: Long-context Knowledges Are Already Well-Organized Before Your Query
Zhiyuan Yu (Nanjing University), Sanglu Lu (Nanjing University)
CompressionTransformerText
🎯 What it does: Studied the token distribution during the prefilling stage of large language models when compressing Key-Value (KV) caches, discovering that a small number of special semantic anchor points (SAT) cluster in the key embedding space, and proposed the ProtoKV method for KV cache compression based on this observation.
ProtoTS: Learning Hierarchical Prototypes for Explainable Time Series Forecasting
Ziheng Peng (Renmin University of China), Liang Sun (Alibaba Group)
Explainability and InterpretabilityTime SeriesFinance Related
🎯 What it does: Proposes the ProtoTS framework, which interprets time series prediction using hierarchical prototypes and extracts interactive information from heterogeneous inputs through multi-channel embeddings and bottleneck fusion;
Provable and Practical In-Context Policy Optimization for Self-Improvement
Tianrun Yu (Brigham Young University), Weitong Zhang (University of North Carolina at Chapel Hill)
OptimizationTransformerLarge Language ModelReinforcement LearningTextBenchmarkChain-of-Thought
🎯 What it does: Proposes the In-Context Policy Optimization (ICPO) framework and its practical algorithm ME-ICPO, which achieve test-time scaling and answer improvement for large language models (LLMs) through self-reflection and self-evaluation rewards during inference.
Provable Separations between Memorization and Generalization in Diffusion Models
Zeqi Ye (Northwestern University), Minshuo Chen (Northwestern University)
GenerationTransformerDiffusion modelImage
🎯 What it does: This paper proposes a theoretical framework for memory and generalization in diffusion models. It first proves from a statistical estimation perspective that the true score function is not the minimum of the denoising score matching loss, leading models to tend to memorize training samples. Then, from a network approximation perspective, it proves that approximating the empirical score function requires the network scale to grow with the number of samples, whereas approximating the true score function can maintain compactness. Based on this analysis, the authors propose a one-time pruning method that prunes and fine-tunes attention heads with low importance in small time steps, thereby reducing memorization while maintaining generation quality.
Provably Accelerated Imaging with Restarted Inertia and Score-based Image Priors
Marien Renaud (University of Bordeaux), Yu Sun (Johns Hopkins University)
RestorationSuper ResolutionOptimizationScore-based ModelImageMagnetic Resonance ImagingOrdinary Differential Equation
🎯 What it does: Proposed the RISP (Restarted Inertia with Score-based Priors) algorithm, which accelerates the iterative convergence of RED while maintaining score-based image priors through restarted inertia.
Provably Explaining Neural Additive Models
Shahaf Bassan (Hebrew University of Jerusalem), Guy Katz (Hebrew University of Jerusalem)
Explainability and InterpretabilityComputational EfficiencyTabular
🎯 What it does: Aiming at Neural Additive Models (NAM), this paper proposes an algorithm capable of generating provable cardinal minimal sufficient explanations.
Proving the Limited Scalability of Centralized Distributed Optimization via a New Lower Bound Construction
Alexander Tyurin (Applied AI Institute)
OptimizationFederated Learning
🎯 What it does: Under the homogenized federated learning scenario, the scalability upper bound of centralized distributed non-convex optimization under computational and communication cost constraints is proven, indicating that linear or square-root level scaling of both dimensionality and variance/error cannot be achieved simultaneously.
Proximal Diffusion Neural Sampler
Wei Guo (Georgia Institute of Technology), Yongxin Chen (Georgia Institute of Technology)
GenerationOptimizationDiffusion modelScore-based ModelMultimodalityGraphBiomedical DataPhysics RelatedStochastic Differential Equation
🎯 What it does: Propose the Proximal Diffusion Neural Sampler (PDNS), which achieves step-by-step learning of diffusion neural samplers by applying proximal point methods in the path measure space, and provides a concrete implementation based on weighted denoising cross-entropy (WDCE).
Proximal Supervised Fine-Tuning
Wenhong Zhu (Shanghai Jiao Tong University), Pengfei Liu (Shanghai Jiao Tong University)
Domain AdaptationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextMultimodality
🎯 What it does: Proposes Proximal Supervised Fine-Tuning (PSFT) — a method that introduces the clipped entropy constraint from PPO into the objective function of supervised fine-tuning, aiming to restrict policy drift during the fine-tuning process.
ProxyAttn: Guided Sparse Attention via Representative Heads
Yixuan Wang (Harbin Institute of Technology), Wanxiang Che (Harbin Institute of Technology)
Computational EfficiencyTransformerTextBenchmark
🎯 What it does: Propose a training-agnostic sparse attention method called ProxyAttn, which uses representative heads to compress head dimensions and estimate attention scores for all heads, achieving more precise block importance assessment.
ProxyThinker: Test-Time Guidance through Small Visual Reasoners
Zilin Xiao (Rice University), Vicente Ordonez (Rice University)
Computational EfficiencyKnowledge DistillationReinforcement LearningImageMultimodalityChain-of-Thought
🎯 What it does: Proposed a reasoning-guidance method called PROXYTHINKER, which adjusts the output distribution of large models by utilizing the contrastive differences from RFT reward reinforcement of small models, thereby enhancing visual reasoning capabilities without additional training.
Prune Redundancy, Preserve Essence: Vision Token Compression in VLMs via Synergistic Importance-Diversity
Zhengyao Fang (Harbin Institute of Technology), Wenjie Pei (Harbin Institute of Technology)
CompressionVision Language ModelMultimodality
🎯 What it does: Proposes a training-free, task-agnostic visual token compression framework called PRUNESID, which compresses visual tokens in Vision-Language Models (VLMs) through a two-stage process.
Prune-then-Quantize or Quantize-then-Prune? Understanding the Impact of Compression Order in Joint Model Compression
Minjun Kim (Seoul National University), U Kang (Seoul National University)
ClassificationCompressionConvolutional Neural NetworkTransformerLarge Language ModelImageText
🎯 What it does: Systematically studied the impact of compression order in joint model compression, proposed and validated the 'Progressive Intensity Hypothesis,' and provided corresponding theoretical analysis and multi-task experimental verification.
Pruning as a Cooperative Game: Surrogate-Assisted Layer Contribution Estimation for Large Language Models
Xuan Ding (Chinese University of Hong Kong), Yao Zhu (Zhejiang University)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Proposes a hierarchical pruning framework based on cooperative game theory, which dynamically identifies and prunes layers with the least contribution to model performance by approximating layer Shapley values using a lightweight proxy network and hierarchical Monte Carlo mask sampling.
Pruning Long Chain-of-Thought of Large Reasoning Models via Small-Scale Preference Optimization
Bin Hong (University of Science and Technology of China), Zhenya Huang (University of Science and Technology of China)
OptimizationComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: Reduce the output length of large-scale inference models while maintaining inference performance through small-scale preference optimization.
PSDNorm: Temporal Normalization for Deep Learning in Sleep Staging
Theo Gnassounou, Alexandre Gramfort (École Polytechnique)
ClassificationConvolutional Neural NetworkTransformerTime SeriesBiomedical Data
🎯 What it does: Proposes the PSDNorm layer, achieving more robust normalization in sleep staging tasks by aligning the power spectral density of feature maps through Monge mapping in deep learning models.
Pseudo-Non-Linear Data Augmentation: A Constrained Energy Minimization Viewpoint
Pingbang Hu (University of Illinois Urbana-Champaign), Mahito Sugiyama (National Institute of Informatics)
ClassificationData-Centric LearningImageTabularAudio
🎯 What it does: Propose a pseudo-nonlinear data augmentation framework based on energy models and information geometry, which constructs a statistical manifold on a custom poset, uses forward projection to obtain low-dimensional representations, and recovers augmented data via backward projection (based on nearest-neighbor local submanifolds), achieving learning-agnostic, interpretable, and controllable augmentation.
PSP: Prompt-Guided Self-Training Sampling Policy for Active Prompt Learning
Sen Tao (Shenyang Institute of Automation, Chinese Academy of Sciences), Zheng-Jun Zha
ClassificationTransformerReinforcement LearningPrompt EngineeringVision Language ModelImage
🎯 What it does: For active prompt learning using pre-trained vision-language models, we propose a Prompt-Guided Self-Training Sampling Policy (PSP), which achieves sample selection and pseudo-label generation based on prompt information through two major modules, VSSP and UST, in a self-training framework.
PT$^2$-LLM: Post-Training Ternarization for Large Language Models
Xianglong Yan (Shanghai Jiao Tong University), Yulun Zhang (Shanghai Jiao Tong University)
CompressionComputational EfficiencyLarge Language ModelText
🎯 What it does: Propose PT-LLM, a ternary framework that compresses large language models to approximately 1.58-bit without retraining, significantly reducing memory and computational costs while maintaining high accuracy.
PTNET: A PROPOSAL-CENTRIC TRANSFORMER NET- WORK FOR 3D OBJECT DETECTION
Jianping Zhong (Harbin Institute of Technology, Weihai), Qingming Huang (University of Chinese Academy of Sciences)
Object DetectionAutonomous DrivingTransformerPoint Cloud
🎯 What it does: This paper proposes a proposal-based Transformer network called PTNet for 3D object detection.
PTQ4ARVG: Post-Training Quantization for AutoRegressive Visual Generation Models
Xuewen Liu (Institute of Automation, Chinese Academy of Sciences), Qingyi Gu (Institute of Automation, Chinese Academy of Sciences)
GenerationComputational EfficiencyTransformerImage
🎯 What it does: Studied post-training quantization methods for autoregressive visual generation models (ARVG), and proposed the PTQ4ARVG framework.
PU-BENCH: A UNIFIED BENCHMARK FOR RIGOROUS AND REPRODUCIBLE PU LEARNING
Qiuyi Chen (Xi'an Jiaotong-Liverpool University), Wei Wang (Xi'an Jiaotong-Liverpool University)
ClassificationImageTextMultimodalityTabularBenchmark
🎯 What it does: Proposes PU-Bench, a unified and reproducible benchmark for learning with incomplete positive and negative labels, integrating 18 advanced algorithms, 8 multimodal datasets, and providing standardized data generation, training, and evaluation processes.
Pulp Motion: Framing-aware multimodal camera and human motion generation
Robin Courant (École Polytechnique), Vicky Kalogeiton (École Polytechnique)
GenerationData SynthesisVision-Language-Action ModelDiffusion modelAuto EncoderVideoTextMultimodality
🎯 What it does: Propose a method for jointly generating human motion and camera trajectories, achieving multimodal consistency through an auxiliary framework.
Purifying Generative LLMs from Backdoors without Prior Knowledge or Clean Reference
Jianwei Li (North Carolina State University), Jung-Eun Kim (North Carolina State University)
Safty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposes a framework for removing backdoors in instruction-tuned large language models (LLMs) without prior trigger knowledge and without relying on a clean reference model.
Purrception: Variational Flow Matching for Vector-Quantized Image Generation
Răzvan-Andrei Matișan (UvA-Bosch Delta Lab University of Amsterdam), Floor Eijkelboom (UvA-Bosch Delta Lab University of Amsterdam)
GenerationRepresentation LearningTransformerDiffusion modelFlow-based ModelAuto EncoderImage
🎯 What it does: Proposes a variable flow matching method named Purrception for generating high-resolution images in the vector quantized (VQ) latent space.
Pursuing Minimal Sufficiency in Spatial Reasoning
Yejie Guo (Shanghai Jiao Tong University), Ming-Hsuan Yang
Representation LearningTransformerLarge Language ModelAgentic AIVision Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposed and implemented a zero-training dual-agent framework called MSSR, which first extracts 3D scene information via the Perception Agent, then constructs a minimal sufficient set (MSS) through iterative refinement and request mechanisms by the Reasoning Agent to answer spatial reasoning questions.
Pusa V1.0: Unlocking Temporal Control in Pretrained Video Diffusion Models via Vectorized Timestep Adaptation
Yaofang Liu (City University of Hong Kong), Jean-michel Morel
GenerationTransformerSupervised Fine-TuningDiffusion modelVideoTextBenchmarkOrdinary Differential Equation
🎯 What it does: Proposed Pusa V1.0, which achieves fine-grained temporal control on pre-trained text-to-video (T2V) diffusion models through vectorized time step adaptation (VTA), supporting multi-tasks such as zero-shot image-to-video generation, start-end frame control, and video extension.
Pushing on Multilingual Reasoning Models with Language-Mixed Chain-of-Thought
Guijin Son (OneLineAI), Youngjae Yu (Seoul National University)
Data-Centric LearningTransformerSupervised Fine-TuningTextMultimodalityChain-of-Thought
🎯 What it does: Propose the Language-Mixed Chain-of-Thought (CoT) reasoning framework and construct the largest post-training Korean dataset YI-SANG. Using this dataset, a series of multilingual reasoning models (KO-REAson series) with sizes ranging from 4B to 35B were trained, achieving cross-lingual and multimodal 'free-lunch' improvements.
Pushing Test-Time Scaling Limits of Deep Search with Asymmetric Verification
Weihao Zeng (Hong Kong University of Science and Technology), Junxian He (Hong Kong University of Science and Technology)
RetrievalOptimizationComputational EfficiencyLarge Language ModelAgentic AIText
🎯 What it does: This paper systematically studies the test-time computational scalability methods for deep search agents, leveraging the 'asymmetric verification' property to achieve optimal computational allocation on two dimensions: search and verification;
PuzzleWorld: A Benchmark for Multimodal, Open-Ended Reasoning in Puzzlehunts
Hengzhi Li (MIT), Paul Pu Liang (MIT)
TransformerLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposes the PUZZLEWORLD benchmark, containing 667 real-world puzzlehunt problems, providing fine-grained multi-modal inputs, cognitive skill labels, and step-by-step reasoning trajectories to evaluate open-ended multi-modal reasoning capabilities.
Pyramid Patchification Flow for Visual Generation
Hui Li (Fudan University), Siyu Zhu (Fudan University)
GenerationTransformerDiffusion modelFlow-based ModelImageTextMultimodality
🎯 What it does: Designed and implemented a pyramid patchification process called PPFlow, which significantly reduces the number of tokens while maintaining latent full resolution by using large patches at high noise levels and small patches at low noise levels, achieving acceleration of Diffusion Transformers.
PYRREGULAR: A Unified Framework for Irregular Time Series, with Classification Benchmarks
Francesco Spinnato (University of Pisa), Cristiano Landi (University of Pisa)
ClassificationTime SeriesBenchmark
🎯 What it does: Proposed the pyrregular framework, which unifies the storage and processing of irregular time series using a single sparse array format, and constructed a standard repository of 34 naturally irregular time series classification datasets, followed by a systematic benchmark evaluation of 12 mainstream classifiers.
pySpatial: Generating 3D Visual Programs for Zero-Shot Spatial Reasoning
Zhanpeng Luo (Carnegie Mellon University), Yaqi Xie (Carnegie Mellon University)
Autonomous DrivingExplainability and InterpretabilityTransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Propose the pySpatial framework, which utilizes MLLM to generate executable Python programs, invoking tools such as 3D reconstruction, camera motion description, and perspective synthesis to achieve zero-shot 3D space reasoning.
Q-Learning with Adjoint Matching
Qiyang Li (University Of California Berkeley), Sergey Levine (University Of California Berkeley)
Reinforcement LearningFlow-based ModelBenchmark
🎯 What it does: Developed a Q-learning based TD algorithm called QAM, which efficiently optimizes flow/diffusion strategies in continuous action spaces using adjoint matching, and addresses the problem of gradient instability.
Q-Learning with Fine-Grained Gap-Dependent Regret
Haochen Zhang (Pennsylvania State University), Lingzhou Xue (Pennsylvania State University)
Reinforcement LearningTabular
🎯 What it does: This paper addresses periodic tabular MDPs in model-free reinforcement learning, providing the first fine-grained upper bounds based on sub-optimality gaps and proposing a novel analytical framework. It also conducts an in-depth analysis of the existing AMB algorithm, identifying flaws in its design and theoretical proof, and proposes two improved versions: UCB-based ULCB-Hoeffding and non-UCB Refined AMB, which are proven to outperform the original AMB in both fine-grained upper bounds and empirical performance.
Q-learning with Posterior Sampling
Priyank Agrawal (Columbia University), Azmat Azati (BNP Paribas)
Reinforcement LearningTabular
🎯 What it does: Proposed a Q-learning algorithm based on posterior sampling (PSQL) to address the exploration-exploitation dilemma in reinforcement learning.
Q-RAG: Long Context Multi‑Step Retrieval via Value‑Based Embedder Training
Artyom Sorokin (AXXX), Evgeny Burnaev (AXXX)
RetrievalReinforcement LearningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed and trained Q-RAG, an efficient model that performs multi-step retrieval in the embedding space using reinforcement learning.
Q&C: When Quantization Meets Cache in Efficient Generation
Xin Ding (University of Science and Technology of China), Zhibo Chen (University of Science and Technology of China)
GenerationDiffusion modelImageVideoText
🎯 What it does: This paper systematically studies the efficiency improvement of jointly applying quantization and caching mechanisms to diffusion models, and proposes corresponding solutions to the problems they bring.
QeRL: Beyond Efficiency - Quantization-enhanced Reinforcement Learning for LLMs
Wei Huang (NVIDIA), Yukang Chen (NVIDIA)
Computational EfficiencyLarge Language ModelReinforcement LearningText
🎯 What it does: Propose a framework named QeRL, combining 4-bit NVFP4 quantization with LoRA fine-tuning to achieve efficient training of large language models in reinforcement learning tasks.
QKV Projections Require a Fraction of Their Memory
Malik Khalaf (Technion Israel Institute of Technology), Assaf Schuster (Technion Israel Institute of Technology)
CompressionComputational EfficiencyTransformerImageText
🎯 What it does: In large-scale language model training, the PAMM method is proposed to compress the Q/K/V projection input activations in attention layers, significantly reducing memory usage.
QLCoder: A Query Synthesizer For Static Analysis of Security Vulnerabilities
Claire Wang (University of Pennsylvania), Mayur Naik (University of Pennsylvania)
AI Code AssistantTransformerLarge Language ModelAgentic AITextRetrieval-Augmented Generation
🎯 What it does: Propose a LLM-based agent framework called QLCoder, which can automatically generate and verify CodeQL static analysis queries from CVE descriptions.
QLIP: A Dynamic Quadtree Vision Prior Enhances MLLM Performance Without Retraining
Kyle R. Chickering (University of California Davis), Muhao Chen (University of California Davis)
Vision Language ModelContrastive LearningImageMultimodality
🎯 What it does: Propose QLIP, a lightweight, pluggable quadtree-based improvement scheme for the CLIP vision encoder, enabling MLLM to perform reasoning on images of arbitrary resolutions without requiring retraining.
QPrompt-R1: Real-Time Reasoning for Domain-Generalized Semantic Segmentation via Group-Relative Query Alignment
Fengyuan Lu (Nanjing University), Qi Fan (Nanjing University)
SegmentationDomain AdaptationComputational EfficiencyTransformerReinforcement LearningPrompt EngineeringImage
🎯 What it does: Propose the QPrompt-R1 architecture and the Group-Relative Query Alignment (GRQA) training strategy to achieve real-time domain-generalized semantic segmentation;
Qronos: Correcting the Past by Shaping the Future... in Post-Training Quantization
Shihao Zhang (University of California, San Diego), Rayan Saab (University of California, San Diego)
OptimizationComputational EfficiencyTransformerText
🎯 What it does: Designed Qronos, a post-training quantization algorithm that simultaneously corrects quantization errors in weights and activations, and rectifies cumulative errors from previous layers.
QuadGPT: Native Quadrilateral Mesh Generation with Autoregressive Models
Jian Liu (Hong Kong University of Science and Technology), Chunchao Guo (Tencent Hunyuan)
GenerationData SynthesisTransformerReinforcement LearningPoint CloudMesh
🎯 What it does: Proposes QuadGPT, an end-to-end autoregressive model capable of directly generating native quadrilateral (with a small number of triangles) meshes, and further enhances topological quality through reinforcement learning after training.
Quadratic Direct Forecast for Training Multi-Step Time-Series Forecast Models
Eric Wang, Zhouchen Lin (Peking University)
OptimizationMeta LearningTransformerTime Series
🎯 What it does: Proposed a quadratic form weighted learning objective along with the MetaDF/QDF algorithms, using an adaptive weight matrix to enhance the training of multi-step time series prediction models
Quagmires in SFT-RL Post-Training: When High SFT Scores Mislead and What to Use Instead
Feiyang Kang (FAIR at Meta), Newsha Ardalani (FAIR at Meta)
TransformerSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Investigate the predictability of RL after SFT, evaluate whether SFT scores can reliably predict RL outcomes, and propose new evaluation metrics;
QuaMo: Quaternion Motions for Vision-based 3D Human Kinematics Capture
Cuong Le (Linköping University), Bastian Wandt (Independent researcher)
Pose EstimationVideoSequentialOrdinary Differential Equation
🎯 What it does: Proposes QuaMo, an online 3D human motion capture method based on quaternion dynamics.
Quant-dLLM: Post-Training Extreme Low-Bit Quantization for Diffusion Large Language Models
Tianao Zhang (Shanghai Jiao Tong University), Yulun Zhang (Shanghai Jiao Tong University)
GenerationComputational EfficiencyTransformerLarge Language ModelDiffusion modelText
🎯 What it does: Provide a 2-bit weight quantization scheme for diffusion large language models
Quantifying Cross-Attention Interaction in Transformers for Interpreting TCR-pMHC Binding
Jiarui Li (Tulane University), Ramgopal R. Mettu (Tulane University)
Explainability and InterpretabilityDrug DiscoveryTransformerBiomedical DataBenchmark
🎯 What it does: Propose a post-hoc explanation method called QCAI to quantify cross-attention in Transformer decoders, aiding in the interpretation of TCR-pMHC binding prediction models.
Quantile Advantage Estimation: Stabilizing RLVR for LLM Reasoning
Junkang Wu (University of Science and Technology of China), Xiangnan He (University of Science and Technology of China)
Large Language ModelReinforcement LearningTextBenchmark
🎯 What it does: Propose a quantile-based advantage estimation (QAE) that replaces the traditional mean baseline with a quantile baseline to achieve stable reinforcement learning training for large language model inference tasks.
Quantitative Bounds for Length Generalization in Transformers
Zachary Izzo (NEC Labs America), Jason D. Lee (Princeton University)
TransformerSequential
🎯 What it does: This paper provides through theoretical analysis the upper limit on training sequence length required for Transformers in length generalization (LG) tasks, and quantifies the upper bounds for single-layer and double-layer, finite and infinite precision attention. Further experiments verify the relationship between these upper bounds and training length.
Quantization-Aware Diffusion Models For Maximum Likelihood Training
Shohei Taniguchi (University of Tokyo), Yutaka Matsuo (University of Tokyo)
GenerationDiffusion modelImageStochastic Differential Equation
🎯 What it does: This study proposes a quantization-aware diffusion model (QDPM), which directly embeds the quantization process into the inverse diffusion process, ensuring that generated samples converge to discrete quantization points.
Quantized Gradient Projection for Memory-Efficient Continual Learning
Dongjun Kim (University of Texas at Austin), Haris Vikalo (University of Texas at Austin)
Computational EfficiencyImageBenchmark
🎯 What it does: Propose a continual learning framework named Quantized Gradient Projection Memory (QGPM), which achieves low-memory and privacy-friendly model updates by quantizing and compressing the gradient subspace.
Quantized Visual Geometry Grounded Transformer
Weilun Feng (Chinese Academy of Sciences), Yongjun Xu (Chinese Academy of Sciences)
Pose EstimationDepth EstimationComputational EfficiencyTransformerImagePoint Cloud
🎯 What it does: A complete scheme for post-training quantization (PTQ) tailored to large visual geometric Transformers (VGGT) is proposed.
QuantSparse: Comprehensively Compressing Video Diffusion Transformer with Model Quantization and Attention Sparsification
Weilun Feng (Chinese Academy of Sciences), Yongjun Xu (Chinese Academy of Sciences)
GenerationData SynthesisComputational EfficiencyTransformerDiffusion modelVideo
🎯 What it does: Proposes the QuantSparse framework, which jointly compresses video diffusion Transformers by combining model quantization with sparse attention, achieving significant storage compression and inference acceleration while maintaining or even enhancing generation quality.
Quantum machine learning advantages beyond hardness of evaluation
Riccardo Molteni (Universiteit Leiden), Vedran Dunjko (Universiteit Leiden)
RecognitionComputational Efficiency
🎯 What it does: This paper proves that there exists an exponential quantum advantage in the learning and identification tasks of quantum functions, i.e., classical algorithms require exponential time to identify unknown quantum label functions.
Quartet of Diffusions: Structure-Aware Point Cloud Generation through Part and Symmetry Guidance
Chenliang Zhou (University of Cambridge), Cengiz Oztireli (University of Cambridge)
GenerationData SynthesisTransformerDiffusion modelAuto EncoderPoint Cloud
🎯 What it does: Proposed the Quartet of Diffusions framework, which utilizes four diffusion models to achieve part decomposition and symmetry constraints during point cloud generation.
Quasi-Equivariant Metanetworks
Viet-Hoang Tran (National University of Singapore), Tan Minh Nguyen
Representation LearningMeta LearningConvolutional Neural NetworkTransformerImageTextBenchmark
🎯 What it does: Proposes the quasi-equivariant metanetworks framework, relaxing strict equivariance constraints and enhancing functional identification and expression capabilities in the weight space.
Quasi-Monte Carlo Methods Enable Extremely Low-Dimensional Deep Generative Models
Miles Martinez (Duke University), Alex H Williams
GenerationComputational EfficiencyImageAudio
🎯 What it does: Proposed a class of deep generative models called QLVM, which directly approximate the marginal likelihood using randomized quasi-Monte Carlo integration, thereby obtaining interpretable embeddings in two- or three-dimensional low-dimensional spaces.
Query-Aware Flow Diffusion for Graph-Based RAG with Retrieval Guarantees
Zhuoping Zhou (Samsung SDS Research America), Seungjai Min (Samsung SDS Research America)
GenerationRetrievalGraph Neural NetworkFlow-based ModelTextGraphBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes QAFD-RAG, a training-agnostic graph retrieval-augmented generation framework that dynamically acquires subgraphs and generates answers using query-aware flow diffusion.
Query-Guided Spatial–Temporal–Frequency Interaction for Music Audio–Visual Question Answering
Kun Li (University of Twente), Sami Sebastian Brandt (IT University of Copenhagen)
TransformerPrompt EngineeringVision Language ModelVideoTextMultimodalityBenchmarkAudio
🎯 What it does: For the music audio-visual question answering task, the QSTar method is proposed, which utilizes query-guided multimodal alignment, spatial-temporal-frequency domain interaction, and prompt-based query context reasoning to perform full-process alignment of audio, video, and text features and predict answers.
Query-Level Uncertainty in Large Language Models
Lihu Chen (Imperial College London), Gaël Varoquaux (Imperial College London)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper proposes a training-free query layer uncertainty assessment method called Internal Confidence, used to determine whether an LLM can answer a given query before generating any token.
Query-Specific Causal Graph Pruning Under Tiered Knowledge
Yizuo Chen (University of California, Los Angeles), Jane E. Barker (Amazon)
Explainability and InterpretabilityComputational EfficiencyGraph
🎯 What it does: Proposes a method that utilizes hierarchical knowledge to prune edges in a causal graph, thereby simplifying the graph structure while maintaining the identifiability of causal effects;
QueryStream: Advancing Streaming Video Understanding with Query-Aware Pruning and Proactive Response
Kairui Zhang (ShanghaiTech University), Changsheng Xu (ShanghaiTech University)
Computational EfficiencyTransformerVision Language ModelVideoBenchmark
🎯 What it does: Proposes the QueryStream framework, integrating Query-Aware Differential Pruning (QDP) and Relevance-Triggered Active Response (RTAR), for real-time streaming video understanding, significantly reducing the number of visual tokens while improving response accuracy.
QUEST: A robust attention formulation using query-modulated spherical attention
Hariprasath Govindarajan (Linköping University), Fredrik Lindsten (Linköping University)
ClassificationSegmentationTransformerImageTextTime Series
🎯 What it does: Investigates the training instability issue in Transformer attention mechanisms, proposing and validating a novel Query-modulated Spherical Attention (QUEST) that can serve as a drop-in replacement for standard attention.
QuestA: Expanding Reasoning Capacity in LLMs via Question Augmentation
Jiazheng Li (Tsinghua University), Jingzhao Zhang (Tsinghua University)
OptimizationComputational EfficiencyData-Centric LearningReinforcement LearningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Propose a technique called QUESTA that helps models reason better by inserting partial solutions (hints) for difficult problems during the reinforcement learning training process.
Queue Length Regret Bounds for Contextual Queueing Bandits
Seoungbin Bae (KAIST), Dabeen Lee (Seoul National University)
OptimizationReinforcement LearningTabular
🎯 What it does: Proposed a novel context-aware framework called Contextual Queueing Bandits, aiming to learn unknown service rates while scheduling. The framework allows individual jobs to carry heterogeneous context features, and the agent selects jobs based on these features to match them with servers, maximizing the departure rate.
QuoKA: Query-Oriented KV Selection for Efficient LLM Prefill
Dalton Jones (Qualcomm AI Research), Christopher Lott (Qualcomm AI Research)
Computational EfficiencyTransformerTextBenchmark
🎯 What it does: Propose a training-agnostic, hardware-agnostic sparse attention algorithm called QUOKA to accelerate the pre-filling stage of Transformers.
Quotient-Space Diffusion Models
Yixian Xu (Peking University), Chang Liu
Drug DiscoveryProtein Structure PredictionDiffusion modelBiomedical DataStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper studies symmetry-aware diffusion models in scientific tasks such as molecular structure and protein backbone generation, proposing a diffusion framework based on quotient spaces;
QuRL: Low-Precision Reinforcement Learning for Efficient Reasoning
Yuhang Li (NVIDIA Research), Brucek Khailany (NVIDIA Research)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Proposes Quantized Reinforcement Learning (QuRL) method, which accelerates LLM reinforcement learning training during the rollout phase by using low-precision (INT8/FP8) quantized actor models, while retaining full-precision parameters for gradient updates.
QuRL: Rubrics As Judge For Open-Ended Question Answering
Xiyu Wei (Peking University), Sujian Li (Peking University)
Explainability and InterpretabilityData-Centric LearningReinforcement Learning from Human FeedbackReinforcement LearningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes the QuRL framework, which automatically generates case-level rubrics for each question using text retrieved from the web, and uses these rubrics as verifiable rewards to guide reinforcement learning in open-ended QA models;
QVGen: Pushing the Limit of Quantized Video Generative Models
Yushi Huang (Hong Kong University of Science and Technology), Jun Zhang (Hong Kong University of Science and Technology)
GenerationComputational EfficiencyKnowledge DistillationDiffusion modelVideo
🎯 What it does: Propose a low-bit quantization-aware training framework (QVGen) for video diffusion models, achieving high-quality video generation under 3-bit/4-bit quantization through auxiliary modules and rank decay;
QVLA: Not All Channels Are Equal in Vision-Language-Action Model's Quantization
Yuhao Xu (Shanghai Jiao Tong University), Zhipeng Zhang (Shanghai Jiao Tong University)
OptimizationComputational EfficiencyRobotic IntelligenceVision-Language-Action Model
🎯 What it does: Proposes the QVLA quantization framework for vision-language-action models, conducting quantization analysis on the action space for robot control and achieving unified channel-level bit-width allocation and pruning;
QWHA: Quantization-Aware Walsh-Hadamard Adaptation for Parameter-Efficient Fine-Tuning on Large Language Models
Hyesung Jeon (Seoul National University), Jae-Joon Kim (Seoul National University)
Computational EfficiencyKnowledge DistillationRepresentation LearningLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposes QWHA, a sparse adapter combined with Walsh-Hadamard Transform (WHT), for efficient fine-tuning on large language models with low-bit quantization;
R-Horizon: How Far Can Your Large Reasoning Model Really Go in Breadth and Depth?
Yi Lu (Fudan University Meituan LongCat Team), Xunliang Cai (Fudan University Meituan LongCat Team)
Reinforcement LearningTextSequentialBenchmarkChain-of-Thought
🎯 What it does: Construct multi-step, interdependent reasoning tasks through query composition (R-HORIZON), and design evaluation benchmarks and reinforcement learning training data based on this to explore the long-term reasoning capabilities of large reasoning models (LRM).
R-WoM: Retrieval-augmented World Model For Computer-use Agents
Kai Mei (Rutgers University), Jiarong Jiang (AWS Agentic AI)
TransformerLarge Language ModelWorld ModelTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose a retrieval-enhanced world model, R-WoM, which uses external tutorials to guide LLM's future state simulation and reward evaluation in computer usage tasks.
R-Zero: Self-Evolving Reasoning LLM from Zero Data
Chengsong Huang (Tencent AI Seattle Lab), Dong Yu (Washington University in St. Louis)
OptimizationTransformerLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: Proposed the R-Zero framework, leveraging a challenger and solver from a homologous LLM to self-generate, filter, and learn incrementally difficult training sets through relative policy optimization without any manually annotated data, achieving self-evolving reasoning capability improvements.