ICML 2026 Papers — Page 39
International Conference on Machine Learning · 6554 papers
OmniFit: Bridging Modalities via Layer-Adaptive Token Compression for Omnimodal Large Language Models
Zining Wang (State Key Laboratory of Complex & Critical Software Environment), Xianglong Liu (State Key Laboratory of Complex & Critical Software Environment)
CompressionOptimizationComputational EfficiencyTransformerLarge Language ModelAuto EncoderContrastive LearningImageVideoTextMultimodalityBenchmark
🎯 What it does: Proposes OmniFit, a training-agnostic token compression framework designed to optimize real-time inference for multimodal large language models.
OmniMoE: An Efficient MoE by Orchestrating Atomic Experts at Scale
Jingze Shi (Hong Kong University of Science and Technology), Yuyu Luo (Hong Kong University of Science and Technology)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsContrastive LearningText
🎯 What it does: Propose the OmniMoE framework, integrating shared dense MLP with atomic-level vector experts, achieving fine-grained expert activation and efficient inference in a single-layer Mixture-of-Experts.
OmniSapiens: A Foundation Model for Social Behavior Processing via Heterogeneity-Aware Relative Policy Optimization
Keane Ong (National University of Singapore), Paul Pu Liang (Massachusetts Institute of Technology)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextMultimodalityBenchmark
🎯 What it does: Constructed the OmniSapiens-7B 2.0 unified foundation model, which can perform reasoning on 10 different social behavior tasks (such as emotion recognition, social reasoning, humor recognition, anxiety/depression detection, etc.), and provide interpretable reasoning trajectories.
OmniShow: Unifying Multimodal Conditions for Human-Object Interaction Video Generation
Donghao Zhou (Chinese University of Hong Kong), Pheng-Ann Heng (Chinese University of Hong Kong)
GenerationData SynthesisTransformerPrompt EngineeringVision-Language-Action ModelDiffusion modelFlow-based ModelAuto EncoderImageVideoTextMultimodalityBenchmarkAudio
🎯 What it does: Propose the OMNISHOW unified model, which generates high-quality human-computer interaction videos using four conditions: text, reference images, audio, and pose.
OmniSIFT: Modality-Asymmetric Token Compression for Efficient Omni-modal Large Language Models
Yue Ding (Chinese Academy of Sciences), Liang Wang (Chinese Academy of Sciences)
CompressionComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningAuto EncoderContrastive LearningVideoTextMultimodalityAudio
🎯 What it does: Proposed and implemented OmniSIFT, a modality-asymmetric token compression framework, which first performs spatiotemporal pruning on videos, and then uses the compressed visual information to guide the selection of audio tokens, thereby significantly reducing the token length of Omni-LLM while maintaining or enhancing performance.
OmniVideo-R1: Reinforcing Audio-visual Reasoning with Query Intention and Modality Attention
Zhangquan Chen (Tsinghua University), Ruqi Huang (Tsinghua University)
Representation LearningData-Centric LearningTransformerReinforcement LearningVision Language ModelAuto EncoderContrastive LearningVideoMultimodalityAudio
🎯 What it does: Propose a two-stage reinforcement learning framework called OmniVideo-R1, which enhances multi-modal audio-visual reasoning capabilities through query-oriented localization and modal attention fusion.
OmniVL-Guard: Towards Unified Vision-Language Forgery Detection and Grounding via Balanced RL
Jinjie Shen (Hefei University Of Technology), Zhun Zhong
Anomaly DetectionReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelImageVideoTextMultimodalityChain-of-Thought
🎯 What it does: Propose OmniVL-Guard, a unified multi-modal (image, video, text) forgery detection and localization framework, integrating self-evolving chain-of-thought generation and adaptive reward scaling reinforcement learning, achieving balanced optimization in classification and localization tasks.
OMP: One-step Meanflow Policy with Directional Alignment
Han Fang (Shanghai Jiao Tong University), Yutong Ban (Shanghai Jiao Tong University)
Robotic IntelligenceReinforcement LearningFlow-based ModelOptical FlowPoint CloudStochastic Differential Equation
🎯 What it does: Developed a single-step MeanFlow strategy OMP, which addresses the spectral bias, gradient exhaustion, and high memory complexity issues of traditional MeanFlow in robotic grasping;
On Contraction of Sequential and Offset Rademacher Complexities
Adam Block (Columbia University), Mark Sellke (OpenAI)
OptimizationExplainability and InterpretabilityComputational EfficiencyRepresentation LearningSequentialReview/Survey PaperChain-of-Thought
🎯 What it does: This paper studies the contraction properties of sequence and offset Rademacher complexity, proposing that under specific structural assumptions, this contraction can be extended to sequence and offset Rademacher complexity, and provides the corresponding bounds.
On Densest $k$-Subgraph Mining and Diagonal Loading: Optimization Landscape and Finite-Step Exact Convergence Analysis
Qiheng Lu (University of Virginia), Aritra Konar (KU Leuven)
OptimizationGraph
🎯 What it does: Studied the non-convex continuous relaxation of the Densest k-Subgraph problem, and provided proofs of its compactness, optimization landscape, and convergence based on Frank-Wolfe.
On Effectiveness and Efficiency of Agentic Tool-calling and RL Training
Tong Liu (LMU Munich), Gabriella Kazai (Amazon)
Computational EfficiencyTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: This paper systematically analyzes the evaluation effectiveness of tool calling in large language models and the training efficiency of reinforcement learning (RL), revealing the significant impact of evaluation details on results, and proposes two acceleration methods based on this.
On Efficient Scaling of GNNs via IO-Aware Layers Implementations
Daria Fomina (Yandex), Fedor Velikonivtsev (Yandex)
Computational EfficiencyGraph Neural NetworkTransformerReinforcement LearningAuto EncoderContrastive LearningGraph
🎯 What it does: To address sparse, unstructured GNN computations, this paper designs an IO-aware GPU kernel, achieving acceleration through the fusion of SpMM-based convolution, subtraction-based aggregation, and attention layers (GATv2/Graph Transformer).
On Expressive Power of Floating-Point Transformers
Sejun Park (Korea University), Geonho Hwang (Gwangju Institute of Science and Technology)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelAuto EncoderGenerative Adversarial NetworkContrastive Learning
🎯 What it does: This paper studies the expressive power of Transformers implemented with floating-point numbers, revealing the impact of asymmetry under floating-point operations, constraints on finite sequence lengths, and position encoding on representability.
On Group Relative Policy Optimization Collapse in Agent Search: The Lazy Likelihood-Displacement
Wenlong Deng (University of British Columbia), Xiaoxiao Li (University of British Columbia)
TransformerReinforcement LearningAgentic AIPrompt EngineeringTextBenchmark
🎯 What it does: In tool-integrated reinforcement learning, this paper analyzes the training collapse phenomenon that occurs during GRPO training and proposes the LLDS regularization method to prevent a sharp decline in model performance.
On Information Self-Locking in Reinforcement Learning for Active Reasoning of LLM agents
Deyu Zou (Chinese University of Hong Kong), James Cheng (Chinese University of Hong Kong)
TransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringTextMultimodalityBenchmark
🎯 What it does: Investigate and address the information self-locking problem caused by reward learning in large language model agents during active reasoning, and propose a directional criticism-based advantage reweighting method (AREW) to break this bottleneck.
On Learnability and Disambiguation of Multiclass Partial Concept Classes
Jingyuan Xu (Wuhan University), Weiwei Liu (Fudan University)
ClassificationExplainability and InterpretabilityComputational EfficiencyRepresentation LearningContrastive Learning
🎯 What it does: This study investigates the PAC learnability of partial concept classes under multi-class settings (possibly with an infinite label space), and reveals the necessary and sufficient conditions for their learnability.
On Minimum Depth and Width of Floating-Point Neural Networks for Representing Floating-Point Functions
Sejun Park (Korea University), Geonho Hwang (Gwangju Institute of Science and Technology)
Computational EfficiencyRepresentation LearningRecurrent Neural Network
🎯 What it does: Analyze the minimum number of layers and width required for ReLU networks to represent all floating-point functions under floating-point operators, proving that the minimum number of layers is 3, and the width is bounded between 2d and 2d+4.
On Path to Multimodal Historical Reasoning: HistBench and HistAgent
Jiahao Qiu (Princeton University), Mengdi Wang (Princeton University)
Recommendation SystemAutonomous DrivingOptimizationFederated LearningSafty and PrivacyExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningAdversarial AttackHyperparameter SearchData-Centric LearningRobotic IntelligenceMeta LearningDrug DiscoveryAI Code AssistantReinforcement Learning from Human FeedbackNeural Architecture SearchProtein Structure PredictionTransformerLarge Language ModelAgentic AIPrompt EngineeringImageTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed HistBench historical reasoning benchmark and HistAgent historical specialized agent;
On Regret Bounds of Thompson Sampling for Bayesian Optimization
Shion Takeno (Nagoya University), Shogo Iwazaki (MI-6 Ltd)
Information TheoryOptimization
🎯 What it does: This paper provides a rigorous theoretical analysis of the cumulative and lenient cumulative reward of Gaussian Process Thompson Sampling (GP-TS) in Bayesian optimization, presenting new lower bounds, upper bounds, and improved proofs of information gain and time complexity;
On Revisiting Entropy for Identifying Mislabeled Images
Chunlei Li (MedAI Technology (Wuxi) Co Ltd), Lichao Mou (MedAI Technology (Wuxi) Co Ltd)
ClassificationAnomaly DetectionTransformerContrastive LearningImageBiomedical Data
🎯 What it does: Propose a statistical measure called Symbolic Entropy Integral (SEI), based on predicted entropy during training, to automatically identify mislabeled samples in medical image datasets.
On Robustness and Chain-of-Thought Consistency of RL-Finetuned VLMs
Rosie Zhao (Apple), Arnab Kumar Mondal
Explainability and InterpretabilityComputational EfficiencyAdversarial AttackData-Centric LearningReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Systematically evaluate the robustness and chain-of-thought consistency of vision-language models fine-tuned with reinforcement learning on visual reasoning tasks, and construct perturbations with incorrect titles and incorrect chain-of-thought reasoning.
On Stable Long-Form Generation: Benchmarking and Mitigating Length Volatility
Zhitao He (Hong Kong University of Science and Technology), Yi R. Fung (Hong Kong University of Science and Technology)
GenerationTransformerLarge Language ModelPrompt EngineeringTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Studied the issue of length fluctuations in large language models during long text generation and proposed a new metric to measure such fluctuations.
On Structured State-Space Duality
Jerry Yao-Chieh Hu (Northwestern University), Han Liu (Northwestern University)
Computational EfficiencyRepresentation LearningTransformerContrastive LearningTime SeriesSequential
🎯 What it does: Formalized and generalized Structured State‑Space Duality (SSD), proposing an SSD that extends from scalar-identity matrices to general diagonal state matrices, providing its computational complexity and implementation methods, and deriving a necessary and sufficient condition for the existence of 1-semiseparable (1-SS) masked attention.
On Testing Conditional Mean Independence for Manifold-Valued Data
Meiling Zeng (Shanghai University of Finance and Economics), Shouxia Wang (Shanghai University of Finance and Economics)
Review/Survey PaperPhysics Related
🎯 What it does: This paper proposes a nonparametric test method for conditional mean independence with Euclidean predictors and manifold-valued responses, called Manifold Martingale Difference Divergence (MMDD);
On the "Induction Bias" in Sequence Models
Reza Ebrahimi (Qualcomm AI Research), Roland Memisevic (Qualcomm AI Research)
Computational EfficiencyRepresentation LearningData-Centric LearningRecurrent Neural NetworkTransformerSupervised Fine-TuningSequentialChain-of-Thought
🎯 What it does: This paper conducts large-scale experiments on synthetic state tracking tasks to systematically evaluate the in-distribution data efficiency and weight sharing capability of Transformer and RNN (LSTM, DenseSSM, Mamba). It reveals that Transformers are prone to 'length-specific' solutions when multi-length reasoning is required, leading to a sharp increase in data demand;
On the Ability of Transformers to Verify Plans
Yash Sarrof (Saarland University), Michael Hahn (Saarland University)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextSequentialChain-of-Thought
🎯 What it does: The study investigates the length generalization ability of Transformers in plan verification tasks, and systematically analyzes the learnability of different planning subclasses (delete-free, well-formed, STRIPS, conditional effects) under fixed and varying universes (object sets).
On the Accuracy of Newton Step and Influence Function Data Attributions
Ittai Rubinstein (Massachusetts Institute of Technology), Samuel B. Hopkins (Massachusetts Institute of Technology)
OptimizationExplainability and InterpretabilityText
🎯 What it does: This paper compares and evaluates the approximation accuracy of influence functions (IF) and single-step Newton (NS) in data removal tasks through theoretical analysis.
On the Adversarial Robustness of Large Vision-Language Models under Visual Token Compression
Xinwei Zhang (Hong Kong Polytechnic University), Haibo Hu (Hong Kong Polytechnic University)
CompressionAdversarial AttackTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This paper studies the adversarial robustness of large-scale vision-language models under the condition of visual token compression, revealing that traditional attacks suffer from optimization-inference mismatch, and proposes a compression-aware adversarial attack framework called CAGE.
On the Anisotropy of Score-Based Generative Models
Andreas Floros (Imperial College London), Pier Luigi Dragotti (Imperial College London)
GenerationData SynthesisDiffusion modelScore-based ModelImageStochastic Differential Equation
🎯 What it does: This paper analyzes the impact of the network architecture of score-based generative models on generation bias, and proposes Score Anisotropy Directions (SADs) to characterize the model's learning tendencies in different output directions.
On the Collapse of Generative Paths: A Criterion and Correction for Diffusion Steering
Ziseok Lee (Seoul National University), Kyungsu Kim (Seoul National University)
GenerationDrug DiscoveryMixture of ExpertsDiffusion modelScore-based ModelBiomedical DataStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: The study addresses the edge path collapse problem that occurs when model scheduling is performed using proportional density combinations during inference, and proposes criteria for judgment and correction methods.
On the Computational Complexity of Performative Prediction
Ioannis Anagnostides (Carnegie Mellon University), Jingming Yan (University of California, Irvine)
OptimizationFederated LearningComputational EfficiencyReinforcement Learning from Human FeedbackReinforcement LearningContrastive Learning
🎯 What it does: Studies the computational complexity of computing stable points when deploying models in machine learning scenarios that lead to changes in data distribution.
On the Convergence of Decentralized Stochastic Minimax Optimization Algorithm with Compressed Communication
Yihan Zhang (Temple University), Hongchang Gao (Temple University)
OptimizationFederated LearningImage
🎯 What it does: A communication-efficient decentralized stochastic gradient descent ascent (CE-DSGDAR) algorithm based on error compensation is developed for solving distributed stochastic minimax optimization problems.
On the Convergence Rate of LoRA Gradient Descent
Siqiao Mu (Northwestern University), Diego Klabjan (Northwestern University)
OptimizationSupervised Fine-TuningContrastive LearningImageText
🎯 What it does: Perform a non-asymptotic analysis of the convergence rate of original LoRA gradient descent without requiring parameter boundedness or Lipschitz smoothness assumptions, and propose an adaptive learning rate scheme based on theoretical insights;
On the Coordination of Value-Maximizing Bidders
Yanru Guan (Peking University), Tao Lin (Microsoft Research)
OptimizationReinforcement Learning from Human FeedbackTabularFinance Related
🎯 What it does: Studied the coordination problem among multiple automated bidders on online advertising platforms, proposing a coordination mechanism where only the highest-value bidder competes while the others remain silent.
On the Difficulty of Learning a Meta-network for Training Data Selection
Zilin Du (Nanyang Technological University), Boyang Li
OptimizationData-Centric LearningMeta LearningConvolutional Neural NetworkTransformerPrompt EngineeringDiffusion modelAuto EncoderGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: This paper provides theoretical analysis and practical improvements for Meta-Learning for Training Data Selection (MTS), revealing the optimization challenges caused by low gradient signal-to-noise ratio (GSNR) and missing input features, and proposes to significantly enhance the effectiveness of training data selection by increasing the batch size and designing rich distributions and dynamic features.
On the Effect of Misspecifying the Embedding Dimension in Low-rank Network Models
Roddy Taing (University of Wisconsin-Madison), Keith Levin (University of Wisconsin-Madison)
Representation LearningGraph
🎯 What it does: The study investigates the impact of incorrect selection of embedding dimensions on node embedding (ASE) estimation in random dot product graphs (RDPG) and other low-rank network models.
On the Emergence of Implicit Curriculum in RLVR Learning Dynamics
Yu Huang (University of Pennsylvania), Yuxin Chen (University of Pennsylvania)
TransformerSupervised Fine-TuningReinforcement LearningSequentialChain-of-Thought
🎯 What it does: This paper studies the training dynamics of Transformers in multi-step compositional reasoning tasks under reward based solely on the final result (RLVR), and reveals the emergence mechanism of implicit curriculum.
On the Entropy Dynamics in Reinforcement Fine-Tuning of Large Language Models
Shumin Wang (University of Science and Technology of China), Yanyong Zhang (University of Science and Technology of China)
OptimizationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Proposes a theoretical framework regarding entropy changes during the process of reinforcement learning fine-tuning (RFT), and provides a first-order expression for entropy changes caused by single logit updates;
On the Epistemic Uncertainty of Overparametrized Neural Networks
David Rügamer (LMU Munich)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningNeural Radiance FieldAuto EncoderGenerative Adversarial NetworkContrastive LearningTabularTime SeriesReview/Survey PaperStochastic Differential Equation
🎯 What it does: This paper conducts theoretical analysis and experimental verification on the uncertainty of parameterized neural networks, focusing on the parameter space uncertainty caused by the lack of identifiability.
On the Existence of Consistent Adversarial Attacks in High-Dimensional Linear Classification
Matteo Vilucchio (Ecole Poly-technique Fédérale de Lausanne), Bruno Loureiro (PSL University)
ClassificationAdversarial AttackImageTabular
🎯 What it does: This paper studies the robustness of high-dimensional linear classifiers under consistent attacks and inconsistent attacks, and provides precise asymptotic expressions for the consistent robust error and the consistent boundary error.
On the Expressive Power of GNNs to Solve Linear SDPs
Chendi Qian (RWTH Aachen University), Christopher Morris (RWTH Aachen University)
OptimizationGraph Neural NetworkSupervised Fine-TuningContrastive LearningGraphTabularBenchmark
🎯 What it does: This paper studies the expressive power of graph neural networks (GNNs) in solving linear semidefinite programming (SDP), proving that standard VC-WL and VC-2WL cannot express the optimal solution of SDP, and proposes stronger higher-order WL architectures such as VC-2FWL, which are then converted into a trainable VC-2FMPNN model; experiments are conducted on synthetic and real SDPLIB data.
On the Expressive Power of Permutation-Equivariant Weight-Space Networks
Adir Dayan (Technion Israel Institute Of Technology), Haggai Maron (Technion Israel Institute Of Technology)
OptimizationRepresentation LearningGraph Neural NetworkTransformerNeural Radiance FieldAuto EncoderContrastive LearningImage
🎯 What it does: Study the expressive power of weight space networks, propose a unified theory, and prove the equivalence of mainstream permutation-equivariant networks;
On the Fragility of Data Attribution When Learning Is Distributed
Xian Gao (Auburn University), Wei-Shinn Ku (Auburn University)
OptimizationFederated LearningExplainability and InterpretabilityAdversarial AttackConvolutional Neural NetworkDiffusion modelAuto EncoderGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: In the federated learning environment, a latent space optimization-based attack is proposed, which manipulates data attribution scores using a small number of forged samples while keeping the global model accuracy unchanged.
On the Generalization Gap in Self-Evolving Language Model Reasoning
Zhenting Qi (Harvard University), Cyrus Rashtchian (Google Research)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: The study investigates how language models can self-evolve through self-generated supervisory signals under strict closed-loop conditions to enhance their reasoning capabilities.
On the Generalization in Topology Optimization via Sensitivity-Conditioned Bernoulli Flow Matching
Mohammad Rashed (Technical University of Munich), Nils Thuerey (Technical University of Munich)
OptimizationTransformerDiffusion modelFlow-based ModelImageTabularPhysics Related
🎯 What it does: This paper proposes a binary manifold generative model based on sensitivity conditions to accelerate topology optimization and improve generalization to out-of-distribution samples.
On the Identifiability of Poisson Branching Structural Causal Model Under Latent Confounding
Jie Qiao (Guangdong University of Technology), Zhifeng Hao (Shantou University)
Recommendation SystemAnomaly DetectionOptimizationFederated LearningExplainability and InterpretabilityData-Centric LearningTabularTime SeriesBenchmark
🎯 What it does: This study proposes a Poisson Branching Structure Causal Model with Latent Confounding Variables (LC-PB-SCM), and utilizes this model for causal structure learning.
On the Infinite Width and Depth Limits of Predictive Coding Networks
Francesco Innocenti (University of Oxford), Rafal Bogacz (University of Oxford)
OptimizationComputational EfficiencyRepresentation LearningConvolutional Neural NetworkTransformerContrastive LearningImageText
🎯 What it does: This paper studies the behavior of predictive coding networks (PCN) in the limits of infinite width and infinite depth, and proves that under parameterizations with stable width and depth and feature learning, PCN converges to backpropagation (BP). Subsequently, experiments verify the validity of this theory in nonlinear networks (such as CNN and Transformer).
On the Interaction of Batch Noise, Adaptivity, and Compression, under $(L_0,L_1)$-Smoothness: An SDE Approach
Enea Monzio Compagnoni (University of Basel), Eduard Gorbunov (MBZUAI)
OptimizationFederated LearningImageStochastic Differential Equation
🎯 What it does: Propose a new framework based on SDE that unifies the analysis of convergence and stability of distributed compressed SGD (DCSGD) and distributed SignSGD (DSignSGD) under (L, L₀¹)-smoothness.
On the Interplay of Pre-Training, Mid-Training, and RL on Reasoning Language Models
Charlie Zhang (Carnegie Mellon University), Xiang Yue (Carnegie Mellon University)
Data SynthesisExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: By constructing controllable synthetic reasoning tasks and a process-level evaluation framework, systematically disentangle and quantify the causal contributions of pre-training, mid-training, and RL to the reasoning capabilities of language models.
On the Intrinsic Limits of Transformer Image Embeddings in Non-Solvable Spatial Reasoning
Siyi Lyu (Nanjing University), Feng Yan (Nanjing University)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningRecurrent Neural NetworkTransformerContrastive LearningImagePoint CloudMeshBenchmark
🎯 What it does: Proposes viewing the spatial representation capability of visual Transformers as a group homomorphism problem, and proves that conventional ViT cannot capture the structural properties of non-solvable spatial transformations (such as SO(3)) in a single forward pass.
On the Learnability of Test-Time Adaptation: A Recovery Complexity Perspective
Zhi Zhou (Nanjing University), Yu-Feng Li (Nanjing University)
Domain AdaptationOptimizationSupervised Fine-TuningContrastive LearningImage
🎯 What it does: This paper proposes a new theoretical framework for analyzing the learnability of test-time adaptation (TTA), defining two metrics: (ε,δ)-recovery complexity and (ε,ρ)-TTA learnability. It explores the fundamental limits of TTA's recovery speed and long-term reliability under non-stationary, unlabeled data streams.
On the Limits of LLM Adaptability: Impact of Model-Internalized Priors on Annotation Task Performance
Etienne Casanova (California Institute of Technology), R. Michael Alvarez (California Institute of Technology)
ClassificationExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsContrastive LearningText
🎯 What it does: This study investigates the interaction between internalized priors (solidified cognition about task definitions) within large language models and prompt instructions, evaluating the model's adaptability and error-correction ability across different definitions of toxicity.
On the Limits of Test-Time Compute: Sequential Reward Filtering for Better Inference
Yue Yu (Indiana University Bloomington), Dongruo Zhou (Indiana University Bloomington)
Computational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningMixture of ExpertsTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Analyze the theoretical limits of adding computation during the inference phase (Test-Time Compute) for large language models, and propose a reward-filtered sequential reasoning method (RFSeqBoN) based on a reward model.
On the Optimization Trajectory of DeepWalk Embeddings
Christopher Harker (University of Utah), Aditya Bhaskara (University of Utah)
OptimizationRepresentation LearningContrastive LearningGraph
🎯 What it does: Investigate the gradient descent optimization trajectory of the DeepWalk algorithm when using negative sampling (SGNS), proving that under small norm initialization and spectral gap assumptions, the update trajectory converges to a low-rank space and can recover community structures in the stochastic block model.
On the Origin of Neural Scaling Laws: from Random Graphs to Natural Language
Maissam Barkeshli (Meta Superintelligence Labs), Andrey Gromov (Meta Superintelligence Labs)
Computational EfficiencyRepresentation LearningData-Centric LearningTransformerLarge Language ModelContrastive LearningTextGraph
🎯 What it does: Investigated and verified that even when data does not exhibit power-law characteristics, transformers still display neural scaling laws in random walk tasks on random graphs, and tracked the evolution of scaling exponents by progressively reducing data complexity (from bigrams to natural language).
On the Plasticity and Stability for Post-Training Large Language Models
Wenwen Qiang (Institute of Software Chinese Academy of Sciences), Hui Xiong (Hong Kong University of Science and Technology)
OptimizationRepresentation LearningData-Centric LearningTransformerLarge Language ModelReinforcement LearningMixture of ExpertsTextBenchmark
🎯 What it does: Studied the plasticity-stability conflict of the GRPO algorithm during the post-training process of large language models (LLMs), and proposed a probabilistic conflict resolution (PCR) framework to achieve smoother and more effective post-training.
On the Power of (Approximate) Reward Models for Inference-Time Scaling: Sequential Monte Carlo and Beyond
Youheng Zhu (Northwestern University), Yiping Lu (Northwestern University)
OptimizationComputational EfficiencyReinforcement Learning from Human FeedbackTextBenchmark
🎯 What it does: This paper provides a theoretical analysis of the Sequential Monte Carlo (SMC) method in inference-time scaling, clarifying the core role of the approximate reward model in this process, and presents upper and lower bounds on computational complexity related to the Bellman error.
On the Power of Source Screening for Learning Shared Feature Extractors
Muxing Wang, Lili Su (Northeastern University)
Federated LearningRepresentation LearningData-Centric LearningContrastive LearningImageTabular
🎯 What it does: Studies how to select data sources to improve the performance of shared feature extraction, demonstrating that using a subset can achieve optimality.
On the Power of Statistics in Class-Incremental Learning with Pretrained Models
Zhiwen Cao (Sichuan University), Jiancheng Lv (Sichuan University)
ClassificationRepresentation LearningTransformerPrompt EngineeringContrastive LearningGaussian SplattingImageVideoBenchmark
🎯 What it does: The study investigates the role of class-level feature statistics in class-incremental learning under large-scale pre-trained models, and proposes two minimalist reference methods (SCLIP and SViT) to verify this.
On the Provable Suboptimality of Momentum SGD in Nonstationary Stochastic Optimization
Sharan Sahu (Cornell University), Martin T. Wells (Cornell University)
OptimizationStochastic Differential Equation
🎯 What it does: Analyzed the tracking error of SGD and momentum methods (Heavy-Ball and Nesterov) in non-stationary environments, and provided finite-time bounds decomposed into three components: convergence, noise, and drift;
On the Relationship Between Activation Outliers and Feature Death in Sparse Autoencoders
Elana Simon (Stanford University), James Zou (Stanford University)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerAuto EncoderContrastive LearningImageTextTabularBiomedical DataReview/Survey Paper
🎯 What it does: This study investigates the relationship between activation outliers and feature death in sparse autoencoders (SAE), proposes a quantification metric γ, and explains the death mechanism from a geometric perspective. It finds that mean-centering can eliminate most deaths, and also explores residual deaths caused by low-rank structures, providing corresponding preprocessing recommendations.
On the Robustness of Langevin Dynamics to Score Function Error
Daniel Yiming Cao (Cornell University), Yuchen Wu (Cornell University)
OptimizationExplainability and InterpretabilityRepresentation LearningRobotic IntelligenceScore-based ModelTabularTime SeriesStochastic Differential Equation
🎯 What it does: This paper studies the robustness of Langevin dynamics sampling when using an estimated score function (L2 or Lp error), proving that sampling may still deviate from the target distribution even with minimal error.
On the Role of Batch Size in Stochastic Conditional Gradient Methods
Rustem Islamov (University of Basel), Volkan Cevher (EPFL)
OptimizationText
🎯 What it does: Studied the role of batch size in random conditional gradient methods under the µ-Kurdyka-Łojasiewicz condition, particularly momentum-based random conditional gradient algorithms (such as Scion).
On the Role of Computation in Reinforcement Learning
Raj Ghugare (Princeton University), Benjamin Eysenbach (Princeton University)
Recurrent Neural NetworkTransformerReinforcement LearningMixture of ExpertsContrastive LearningSequentialBenchmark
🎯 What it does: This paper studies the impact of computational resources on policy performance and generalization in reinforcement learning, proposes a policy model with variable computational budget, and experimentally verifies that more computation can enhance performance and long-term generalization.
On the Salience of Low-Probability Tokens for AI-Generated Text Detection: A Multiscale Uncertainty Perspective
Yikai Guo (Beijing Institute of Computer Technology and Application), Haoran Luo (Nanyang Technological University)
Anomaly DetectionTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
🎯 What it does: Proposed a multi-scale uncertainty detection framework called Uncertainty/Uncertainty++, based on low probability words and R'enyi entropy, for detecting zero-shot AI-generated text.
On the Sample Efficiency of Inverse Dynamics Models for Semi-Supervised Imitation Learning
Sacha Morin (Universite de Montreal), Sebastien Lachapelle (Samsung AI Lab)
Robotic IntelligenceConvolutional Neural NetworkRecurrent Neural NetworkTransformerReinforcement LearningVision-Language-Action ModelDiffusion modelScore-based ModelFlow-based ModelRectified FlowAuto EncoderGenerative Adversarial NetworkContrastive LearningWorld ModelOptical FlowImageVideoSequential
🎯 What it does: The paper studies the sample efficiency advantage of inverse dynamics models (IDM) in semi-supervised imitation learning, unifies three categories of methods: VM-IDM, IDM annotation, and LAPO, and proposes an improved LAPO+ algorithm;
On the Separability of Information in Diffusion Models
Akhil Premkumar (University of California San Diego)
GenerationExplainability and InterpretabilityDiffusion modelScore-based ModelAuto EncoderContrastive LearningImageTabular
🎯 What it does: This paper conducts theoretical and experimental analysis on the allocation of information in diffusion models, proposing to decompose the model's neural entropy into total correlation (TC), which describes the correlation within pixels, and mutual information I(X;Y), which describes semantic correlation. It proves that most information in image diffusion is used to reconstruct fine-grained perceptual details, and that mutual information is much smaller than neural entropy, thus explaining the effectiveness of classifier-free guidance (CFG).
On the Sharp Input-Output Analysis of Nonlinear Systems under Adversarial Attacks
Jihun Kim (University of California Berkeley), Javad Lavaei (University of California Berkeley)
OptimizationFederated LearningExplainability and InterpretabilityComputational EfficiencyRepresentation LearningAdversarial AttackAuto EncoderContrastive LearningTabularTime SeriesSequentialStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper studies how to learn input-output mappings from control inputs and partially observed outputs in nonlinear dynamical systems where sparse, probabilistic adversarial attacks are present.
On the Spectral Unreachability of Brain Graph Learning
Jiaming Zhuo (Hebei University of Technology), Liang Yang (Hebei University of Technology)
ClassificationAnomaly DetectionRepresentation LearningGraph Neural NetworkSupervised Fine-TuningAuto EncoderContrastive LearningGraphBiomedical DataAlzheimer's DiseaseBenchmark
🎯 What it does: Proposed and verified a deep learning framework called HISP-NET, which can avoid boundary blurring caused by spectral domain low-pass filtering in brain network classification, and used it to achieve fine-grained functional module segmentation and disease diagnosis.
On the Theoretical Limitations of Embedding-based Link Prediction
Samy Badreddine (Sony AI), Luciano Serafini (Fondazione Bruno Kessler)
Knowledge DistillationRepresentation LearningGraph Neural NetworkTransformerMixture of ExpertsContrastive LearningGraph
🎯 What it does: This paper studies the rank bottleneck caused by the linear output layer in embedded link prediction models, and provides necessary dimension upper bounds for different prediction tasks (ranking, sign, and distribution reconstruction);
On the Theory of Continual Learning with Gradient Descent for Neural Networks
Hossein Taheri (University of California San Diego), Arya Mazumdar (University of California San Diego)
OptimizationFederated LearningRepresentation LearningMeta LearningTransformerImageTabular
🎯 What it does: Studied the dynamics of gradient descent in continuous learning for two-dimensional hidden layer quadratic activation neural networks, and derived a closed-form forgetting upper bound on XOR-cluster Gaussian noise data;
On the Variability of Concept Activation Vectors
Julia Wenkmann (University of Würzburg), Damien Garreau (University of Würzburg)
Explainability and InterpretabilityComputational EfficiencyConvolutional Neural NetworkRecurrent Neural NetworkGraph Neural NetworkTransformerSupervised Fine-TuningAuto EncoderGenerative Adversarial NetworkContrastive LearningImageTextMultimodalityTabular
🎯 What it does: Analyze the variability of concept activation vectors (CAVs) in the TCAV method, and provide theoretical and experimental guidance
On Training Large Language Models for Long-Horizon Tasks: An Empirical Study of Horizon Length
Sunghwan Kim (Yonsei University), Jinyoung Yeo (Yonsei University)
Autonomous DrivingOptimizationData-Centric LearningRobotic IntelligenceTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningWorld ModelTextSequentialChain-of-Thought
🎯 What it does: Investigate the impact of the required interaction steps (horizon length) for a task on the training of large language models (LLMs), and alleviate instability during long-sequence training through 'horizon reduction' methods such as macro actions and subgoal decomposition, thereby improving the model's performance on tasks with long horizons.
On Uniform Error Bounds for Kernel Regression under Non-Gaussian Noise
Johannes Teutsch (Technical University of Munich), Armin Lederer (National University of Singapore)
OptimizationExplainability and InterpretabilityComputational EfficiencyTabular
🎯 What it does: This paper proposes non-asymptotic uniform error bounds for kernel regression under non-Gaussian noise, and provides general theoretical upper bounds for various noise distributions (sub-Gaussian, bounded, sub-exponential, variance/moment bounded) as well as correlated noise;
Once-for-All: Scalable Simultaneous Forecasting via Equilibrium State Estimation
Beinan Xu (RMIT University), Feng Liu (University of Melbourne)
OptimizationComputational EfficiencyRecurrent Neural NetworkTransformerDiffusion modelScore-based ModelFlow-based ModelRectified FlowAuto EncoderContrastive LearningTabularTime SeriesFinance RelatedStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Proposed the Equilibrium State Estimation (ESE) framework, which predicts the future states of all systems simultaneously by leveraging the equilibrium state of multiple systems, avoiding redundant calculations and information loss caused by predicting one by one.
One Batch Is Enough: A Unified Dataset Condensation Framework for General Time Series Analysis
Wei Shao (Zhejiang University), Yunjun Gao (Zhejiang University)
ClassificationData SynthesisCompressionAnomaly DetectionKnowledge DistillationRepresentation LearningConvolutional Neural NetworkRecurrent Neural NetworkTransformerContrastive LearningTabularTime Series
🎯 What it does: Proposes UniTSC, a unified dataset compression framework for general time series analysis, which can retain most of the original data's performance under extremely small sample sizes;
One Bias After Another: Mechanistic Reward Shaping and Persistent Biases in Language Reward Models
Daniel Fein (Stanford University), Nick Haber (Stanford University)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmark
🎯 What it does: This paper quantifies various biases, including length bias, confidence bias, position bias, agreement bias (sycophancy), and model-style sensitivity, through systematic experiments on five high-quality reward models (RMs). It proposes a mechanism for reward shaping based on linear activation probes (DiffMean) and zero-space projection, achieving bias mitigation for low-complexity biases with minimal labeled data.
One Bug, Hundreds Behind: LLMs for Large-Scale Bug Discovery
Qiushi Wu (IBM Research), Douglas Lee Schales (IBM Research)
Anomaly DetectionComputational EfficiencyAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the BUGSTONE framework, which generates security coding rules using a single patch and combines LLVM static analysis with LLM for large-scale RPB (Recurring Pattern Bugs) detection.
One Coin Has Two Sides: Single Poistive Multi Label Learning from Salient Annotations
Xiaoyu Wang (Southeast University), Yuheng Jia (Southeast University)
RecognitionObject DetectionRepresentation LearningConvolutional Neural NetworkTransformerPrompt EngineeringAuto EncoderContrastive LearningImage
🎯 What it does: Propose the Salient Single-Positive Multi-Label Learning (SalSPML) problem and design the PiSA method, which utilizes salient labels to generate class prototypes to guide embedding learning and perform prototype-driven negative sample rejection, thus addressing the challenge of biased single-positive multi-label learning towards salient labels.
One Intervention per Component is Enough: Towards Identifiability in Linear Stochastic Dynamics from Steady State
Saber Salehkaleybar (Leiden University)
OptimizationFederated LearningExplainability and InterpretabilityComputational EfficiencyData-Centric LearningContrastive LearningTabularTime SeriesBiomedical DataBenchmarkStochastic Differential Equation
🎯 What it does: Studied how to utilize steady-state observations and intervention data to recover the drift, bias, and diffusion parameters of multivariate Ornstein-Uhlenbeck processes, and proved that under specific SCC conditions, only one intervention per strongly connected component is needed to achieve parameter generic identifiability (subject only to global scale uncertainty constraints)
One LR Doesn’t Fit All: Heavy-Tail Guided Layerwise Learning Rates for LLMs
Di He (Shenzhen Institutes of Advanced Technology Chinese Academy of Sciences), Shiwei Liu (Max Planck Institute for Intelligent Systems)
OptimizationComputational EfficiencyRepresentation LearningTransformerLarge Language ModelAuto EncoderContrastive LearningText
🎯 What it does: Proposes a hierarchical learning rate allocation strategy LLR based on the Heavy-Tail Self-Regulation (HT-SR) theory, which automatically calculates the heavy-tail index of the weight spectrum for each layer of the Transformer language model and dynamically adjusts the learning rate.
One Model to Translate Them All: Universal Any-to-Any Translation for Heterogeneous Collaborative Perception
Yang Li (Beijing University of Posts and Telecommunications), Jinglin Li (Beijing University of Posts and Telecommunications)
Domain AdaptationAutonomous DrivingComputational EfficiencyRepresentation LearningTransformerMixture of ExpertsContrastive LearningImageMultimodalityPoint Cloud
🎯 What it does: Proposes UniTrans, a unified model that can perform zero-shot translation between any two heterogeneous perceptual intermediate features without requiring retraining.
One Probe Won’t Catch Them All: Towards Targeted Deception Detection
Vikram Natarajan (LASR Labs), Joseph Isaac Bloom
Anomaly DetectionExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: This paper systematically analyzes 2,000 linear probes to investigate the feasibility and limitations of detecting deceptive behavior in large language models.
One Step Forward and K Steps Back: Better Reasoning with Denoising Recursion Models
Chris Cameron (Huawei Technologies), Yingxue Zhang (Huawei Technologies)
ClassificationComputational EfficiencyRepresentation LearningData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelDiffusion modelTextBenchmarkChain-of-Thought
🎯 What it does: Propose and train a new recurrent transformer method called Denoising Recursive Model (DRM), whose training process first generates noisy targets through discrete mask diffusion, and then gradually denoises them in k recursive steps, thereby achieving learning of long-term reasoning paths.
One Tool Is Enough: Reinforcement Learning of LLM Agents for Repository-Level Code Navigation
Zhaoxi Zhang (Peking University), Yunfang Wu (Peking University)
AI Code AssistantTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringText
🎯 What it does: Proposes RepoNavigator, an LLM agent that achieves repository-level code localization using only a single 'jump' tool, through reinforcement learning;
One-shot Conditional Sampling: MMD meets Nearest Neighbors
Anirban Chatterjee (University of Chicago), Rohan Hore (Carnegie Mellon University)
RestorationGenerationSuper ResolutionScore-based ModelImage
🎯 What it does: Proposed a single-step conditional sampling framework called CGMMD based on Maximum Mean Discrepancy (MMD), training the conditional generator by minimizing the empirical ECMMD loss.
One-shot Entropy Minimization for Language Model Reasoning
Zitian Gao (IQuest Research), Bryan Dai (IQuest Research)
Computational EfficiencyRepresentation LearningData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBenchmark
🎯 What it does: This paper proposes an unsupervised post-training method called One-shot Entropy Minimization (EM), which requires only one unlabeled sample and about 10 gradient steps. By minimizing the model's own entropy, the performance of various reasoning and generation tasks is significantly improved.
One-Shot Weighted Ensemble Estimation for Federated Quantile Regression: Optimal Statistical Guarantees under Heterogeneous Structured Data
Guang Yang (Nanjing Audit University), Bei Jiang (University of Alberta)
OptimizationFederated LearningComputational EfficiencyTabularBenchmark
🎯 What it does: A data-driven one-time weighted ensemble estimator is proposed for federated quantile regression (FQR), aiming to improve communication efficiency and statistical optimality.
One-Step Gradient Delay is Not a Barrier for Large-Scale Asynchronous Pipeline Parallel LLM Pretraining
Philip Zmushko (Institute of Science and Technology Austria), Samuel Horváth (Mohamed bin Zayed University of Artificial Intelligence)
OptimizationFederated LearningComputational EfficiencyTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: Studying asynchronous pipeline parallel (Async PP) training for large language models, exploring the impact of one-step gradient delay on convergence;
One-Step Graph-Structured Neural Flows for Irregular Multivariate Time Series Classification
Mengzhou Gao (Hangzhou Dianzi University), Pengfei Jiao (Hangzhou Dianzi University)
ClassificationAnomaly DetectionGraph Neural NetworkFlow-based ModelAuto EncoderContrastive LearningTime SeriesBiomedical DataElectronic Health Records
🎯 What it does: Propose a first-order graph structure neural flow model, GSNF, for classifying irregular multivariate time series.
One-step Latent-free Image Generation with Pixel Mean Flows
Yiyang Lu, Kaiming He (Massachusetts Institute Of Technology)
GenerationTransformerDiffusion modelScore-based ModelFlow-based ModelAuto EncoderImage
🎯 What it does: Propose a first-order implicit latent space pixel-level MeanFlow model that can directly map noise to complete pixel images.
One-step Optimal Transport via Regularized Distribution Matching Distillation
Denis Rakitin (HSE University), Dmitry Vetrov (Constructor University)
Image TranslationKnowledge DistillationDiffusion modelScore-based ModelImageText
🎯 What it does: Propose a one-stage unpaired translation method called RDMD, which utilizes diffusion models for distribution matching and incorporates regularization to control transportation costs, directly approximating the optimal transport mapping.
One-Step Residual Shifting Diffusion for Image Super-Resolution via Distillation
Daniil Selikhanovych (Kandinsky Lab), Alexander Korotin (Applied AI Institute)
Super ResolutionKnowledge DistillationConvolutional Neural NetworkTransformerDiffusion modelScore-based ModelGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: Designed a first-order residual shift diffusion (RSD) model, distilling ResShift to obtain a single-step inference, ultra-high perceptual quality super-resolution model.
One-Way Policy Optimization for Self-Evolving LLMs
Shuo Yang (Peking University), Li Yuan (Peking University)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Designed and implemented One-Way Policy Optimization (OWPO), a self-evolving LLM training method under the RLVR framework that achieves decoupling of direction and magnitude.
OnePO: Direct One-stage Policy Optimization for SFT-free Domain Adaptation
Junying Chen (Chinese University of Hong Kong), Benyou Wang (Chinese University of Hong Kong)
Domain AdaptationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBiomedical Data
🎯 What it does: Propose a SFT-free one-stage policy optimization method called OnePO, which quickly injects domain knowledge into reinforcement learning using teacher outputs and then automatically exits, thus avoiding the bottleneck of traditional two-stage SFT+RL.
OneSearch: A Preliminary Exploration of the Unified End-to-End Generative Framework for E-commerce Search
Ben Chen (Kuaishou Technology), Kun Gai (Kuaishou Technology)
RetrievalRecommendation SystemTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringMixture of ExpertsTextTabularSequentialRetrieval-Augmented Generation
🎯 What it does: Designed and deployed OneSearch, an end-to-end generative retrieval and ranking framework to replace traditional multi-stage e-commerce search systems.
Online Bayesian Experimental Design for Partially Observed Dynamical Systems
Sara Perez-Vieites (University of Helsinki), Dominik Baumann (Aalto University)
Autonomous DrivingOptimizationRobotic IntelligenceDrug DiscoveryReinforcement Learning from Human FeedbackTabularTime SeriesSequentialBenchmarkStochastic Differential Equation
🎯 What it does: Propose an online Bayesian experimental design method for partially observable dynamic systems, which can select the most informative experiment design after each update step.
Online Change Point Detection for Multivariate Inhomogeneous Poisson Processes Time Series
Xiaokai Luo (University of Notre Dame), OSCAR HERNAN MADRID PADILLA
Anomaly DetectionOptimizationComputational EfficiencyTime SeriesPhysics Related
🎯 What it does: Developed a low-rank matrixization and restricted SVD detector for online multivariate inhomogeneous Poisson point process time series, used for change point detection under β-mixing dependencies.
Online Compatible Reward Identification from Preference Feedback
Simone Drago (Politecnico di Milano), Alberto Maria Metelli (Politecnico di Milano)
Information TheoryOptimizationExplainability and InterpretabilityReinforcement Learning from Human FeedbackReinforcement Learning
🎯 What it does: Proposes a compatibility reward identification framework based on human preference feedback, aiming to recover a consistent and interpretable reward function across the entire state-action space from preference information derived from trajectory pairs.
Online Conformal Prediction via Universal Portfolio Algorithms
Tuo Liu (King Abdullah University of Science and Technology), Francesco Orabona (King Abdullah University of Science and Technology)
OptimizationFederated LearningReinforcement Learning from Human FeedbackReinforcement LearningScore-based ModelContrastive LearningTabularTime SeriesFinance Related
🎯 What it does: Propose a parameter-agnostic online conformal prediction framework (UP-OCP) based on a universal portfolio algorithm, achieving the optimal trade-off between long-term coverage and prediction interval size.