ICML 2026 Papers — Page 2
International Conference on Machine Learning · 6554 papers
A model of errors in transformers
Suvrat Raju (International Centre for Theoretical Sciences, Tata Institute of Fundamental Research), Praneeth Netrapalli (Google Deepmind)
Explainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Investigate the error rates of LLMs on tasks requiring deterministic outputs (such as arithmetic, list reversal, etc.), and propose a quantitative model that depends on only two parameters to describe how error rates change with task complexity, validated through extensive experiments.
A Narrowing Geometry in Contaminated Reasoning
Jiakuan Xie (University of Chinese Academy of Sciences), Jun Zhao (Chinese Academy of Sciences)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
🎯 What it does: This paper investigates the phenomenon of 'contaminated reasoning' in large language models (LLMs) caused by data leakage, and proposes a method to quantify and interpret it from an internal mechanism perspective.
A New Framework for Cybersecurity Refusals in AI Agents
Eliot Krzysztof Jones (Gray Swan AI), J Zico Kolter (Gray Swan AI)
Safty and PrivacyTransformerLarge Language ModelAgentic AIPrompt EngineeringTextTabularBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed and verified the 'Cybersecurity Refusal Framework' (CRF), which is used to evaluate when AI agents should refuse to execute requests in network attack tasks, and constructed a corresponding evaluation benchmark.
A Noise Sensitivity Exponent Controls Large Statistical-to-Computational Gaps in Single- and Multi-Index Models
Leonardo Defilippis (École Normale Supérieure), Antoine Maillard (École Normale Supérieure)
OptimizationFederated LearningExplainability and InterpretabilityComputational EfficiencyRepresentation LearningMeta LearningReinforcement Learning from Human FeedbackNeural Architecture SearchPrompt EngineeringMixture of ExpertsDiffusion modelScore-based ModelAuto EncoderGenerative Adversarial NetworkContrastive LearningReview/Survey PaperPhysics RelatedChain-of-ThoughtStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Studied statistically feasible but computationally challenging learning problems in single-index models and separable multi-index models, proposing the Noise Sensitivity Index (NSE) as a key quantitative metric for identifying computational bottlenecks in prediction.
A Penalty Approach For Differentiation Through Black-box Quadratic Programming Solvers
Yuxuan Linghu (Shanghai Jiao Tong University), Qi Deng (Shanghai Jiao Tong University)
OptimizationTabularTime SeriesFinance Related
🎯 What it does: This paper proposes a gradient computation framework called dXPP, which embeds the constraints of convex quadratic programming (QP) into a smooth penalty function and uses a black-box QP solver to perform the forward solution. In the backward pass, it only needs to solve a symmetric positive definite linear system of the same dimension as the original variables, thereby achieving efficient and robust gradient derivation.
A Perturbation Approach to Unconstrained Linear Bandits
Andrew Jacobsen (Università degli Studi di Milano), Nicolò Cesa-Bianchi (Università degli Studi di Milano)
OptimizationReinforcement Learning
🎯 What it does: This paper re-examines the perturbation method proposed by Abernethy et al., and introduces the PABLO framework, which transforms any online linear optimization (OLO) algorithm into an unconstrained linear bandit (uBLO) problem, and provides new expected and high-probability adaptive comparator, dynamic regret results;
A Positive Case for Faithfulness: Explanations Help Predict Model Behavior
Harry Mayne (University of Oxford), Noah Y. Siegel (Google DeepMind)
ClassificationExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextTabularRetrieval-Augmented Generation
🎯 What it does: Proposed and evaluated a new metric called Normalized Simulatability Gain (NSG) to measure the self-explainability credibility of LLMs, and conducted systematic experiments on 18 state-of-the-art LLMs (including Gemini-3, GPT-5.2, Claude-4.5, etc.).
A Probabilistic Framework for LLM-Based Model Discovery
Stefan Wahl (University of Tübingen), Daniel Gedon (University of Tübingen)
Explainability and InterpretabilityComputational EfficiencyDrug DiscoveryAI Code AssistantNeural Architecture SearchTransformerLarge Language ModelPrompt EngineeringTextTabularBiomedical Data
🎯 What it does: This paper reformulates model discovery methods based on large language models (LLMs) as Bayesian inference over executable programs, and proposes the ModelSMC algorithm. It utilizes the Sequential Monte Carlo (SMC) framework to enable LLMs to generate, modify candidate models, and assign weights based on likelihood during the iterative process, thereby achieving sampling from the model posterior.
A Progressive Evidence Localization Framework Based on Wasserstein Gradient Flows for Document Visual Question Answering
Haosen Wang (Tianjin University), Zhiyong Feng (Tianjin University)
RecognitionImage TranslationRestorationSegmentationGenerationOptimizationExplainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerMixture of ExpertsVision Language ModelDiffusion modelScore-based ModelFlow-based ModelImageTextMultimodalityBenchmark
🎯 What it does: Propose an evolutionary evidence localization framework based on Wasserstein gradient flow, which gradually transfers the probability distribution from the entire page to the real evidence regions, achieving fine-grained evidence localization and answer generation in DocVQA.
A Provable Expressiveness Hierarchy in Hybrid Linear-Full Attention
Xiaowei Ye (Ecole Polytechnique), Pinyan Lu (Huawei)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelContrastive Learning
🎯 What it does: This paper provides a theoretical analysis of the expressive power of hybrid linear-full attention Transformers, proving that there exists a decidable hierarchy of expressiveness between them and full attention.
A proximal ADMM for multiblock problems with block anti-upper triangular constraints
Zhanwang Deng (Peking University), Wen Huang (Xiamen University)
OptimizationTabularBiomedical DataAlzheimer's Disease
🎯 What it does: Proposed and analyzed an approximate ADMM algorithm for block anti-triangular constrained multi-block problems.
A Queueing-Theoretic Framework for Stability Analysis of LLM Inference with KV Cache Memory Constraints
Chengyi Nie (Stony Brook University), Zijie Zhou (Hong Kong University of Science and Technology)
OptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper proposes a queuing theory framework for LLM inference that explicitly considers KV cache memory constraints, and provides closed-form criteria for stability and instability.
A Random Matrix Perspective on the Consistency of Diffusion Models
Binxu Wang (Harvard University), Cengiz Pehlevan (Harvard University)
GenerationData SynthesisTransformerDiffusion modelScore-based ModelImage
🎯 What it does: Propose a framework based on random matrix theory to quantitatively explain the high consistency of diffusion model outputs under different subsets of training data, and verify the theoretical predictions on deep networks (UNet, DiT).
A Random Matrix Theory of Masked Self-Supervised Learning
Arie Wortsman Zurich, Yue M. Lu (Harvard University)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningContrastive LearningTabularTime Series
🎯 What it does: This paper conducts a high-dimensional analysis of the linear regression model in masked self-supervised learning (SSR) through random matrix theory, providing convergence properties, generalization error, and spectral distribution of the aggregated predictor, and revealing its performance under different data structures (Gaussian autoregressive, spiked covariance).
A recipe for scalable attention-based ML potentials: unlocking long-range accuracy with all-to-all node attention
Eric Qu (UC Berkeley), Zachary Ward Ulissi
Drug DiscoveryGraph Neural NetworkTransformerMixture of ExpertsDiffusion modelScore-based ModelAuto EncoderGenerative Adversarial NetworkContrastive LearningGraphTabularPhysics Related
🎯 What it does: Propose an expandable attention mechanism machine learning atomic potential energy model named AllScAIP, capable of handling million-level datasets and achieving accurate simulations of long-range interactions.
A Recursive Decomposition Framework for Causal Structure Learning in the Presence of Latent Variables
Zheng Li (Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences), Hao Zhang (Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningGraphTabularBiomedical DataBenchmark
🎯 What it does: Propose a recursive decomposition framework called DICOLA, which can achieve efficient causal structure learning through constraint-based causal learning methods (such as FCI) even in the presence of latent variables, significantly reducing the number of conditional independence (CI) tests.
A Refined Generalization Analysis for Extreme Multi-class Supervised Contrastive Representation Learning
Nong Minh Hieu (Singapore Management University), Antoine Ledent (Singapore Management University)
ClassificationRepresentation LearningContrastive LearningImage
🎯 What it does: A more refined analysis of the generalization error in supervised contrastive learning is conducted, and the U-Statistic estimator is improved, proposing a new debiasing estimation method.
A Regime-Aware Trajectory Prediction Framework for 1000+ Systems Biology Models
Heng Rao (Northeastern University), Minghan Chen (Wake Forest University)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningData-Centric LearningDrug DiscoveryRecurrent Neural NetworkTransformerDiffusion modelScore-based ModelFlow-based ModelTime SeriesBiomedical DataBenchmarkOrdinary Differential Equation
🎯 What it does: Proposes the RegimeFlow framework, which can generate trajectories with uncertainty quantification in long-term prediction tasks across biological systems.
A Regret Minimization Framework on Preference Learning in Large Language Models
Suhwan Kim (Seoul National University), Jungwoo Lee (Seoul National University)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This paper proposes a preference learning framework based on regret minimization (RePO), reinterpreting RLHF from reward maximization as optimizing policies from the relative suboptimality of actions.
A Risk Decomposition Framework for Pre-hoc Fine-tuning Prediction
Yuxiang Luo (Hong Kong University of Science and Technology), Nan Tang (Hong Kong University of Science and Technology)
OptimizationFederated LearningComputational EfficiencyData-Centric LearningTextTabularTime SeriesSequential
🎯 What it does: Proposes a risk decomposition framework for pre-predicting the performance of fine-tuning, decomposing prediction error into an irreducible intrinsic limit and a reducible optimization variance, and provides a budget-optimal exploration principle.
A Robust Optimization Guided Pruning Framework for Vision and Large Language Models
Gabriel Afriat (MIT), Rahul Mazumder (MIT)
OptimizationComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelAuto EncoderContrastive LearningImageText
🎯 What it does: Proposes a robust optimization guided one-shot pruning framework called RobOP, which can sparsify Vision Transformers and large language models (LLMs) without retraining.
A robust PPG foundation model using multimodal physiological supervision
Eloy Geenjaar (Georgia Institute of Technology), Daniel P. Darcy (Dolby Laboratories)
Anomaly DetectionRepresentation LearningConvolutional Neural NetworkTransformerSupervised Fine-TuningContrastive LearningMultimodalityTime SeriesBiomedical DataElectrocardiogram
🎯 What it does: Propose a foundational model that uses simultaneously recorded ECG and respiration signals to generate physiological supervision indicators for multimodal contrastive learning on PPG signals, and performs inference using only PPG after training.
A Semantically Consistent Dataset for Data-Efficient Query-Based Universal Sound Separation
Kai Li (Tsinghua University), Xiaolin Hu (Tsinghua University)
RestorationGenerationData SynthesisComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringDiffusion modelFlow-based ModelContrastive LearningMultimodalityAudio
🎯 What it does: This paper proposes an automated data cleaning and synthesis pipeline, and based on this, constructs the Hive dataset. By utilizing high-purity single-event samples and a semantic consistency hybrid strategy, it provides a high-quality, reproducible data source for training query-based universal sound separation (USS).
A Short and Unified Convergence Analysis of the SAG, SAGA, and IAG Algorithms
Feng Zhu (North Carolina State University), Aritra Mitra (North Carolina State University)
OptimizationContrastive LearningGaussian SplattingStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Proposes a unified convergence analysis framework for three variance-reduced algorithms: SAG, SAGA, and IAG, providing high-probability linear convergence rates;
A Single Layer to Explain Them All: Understanding Massive Values in Large Language Models
Zeru Shi (Rutgers University), Ruixiang Tang (Rutgers University)
Explainability and InterpretabilityComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringText
🎯 What it does: Analyze the origin of extreme activation in large language models, identify the 'Massive Emergence Layer (ME Layer)', and propose a method called WeMask, which applies weight-oriented dimension masking to intervene in the hidden states before self-attention at this layer, significantly improving the model's performance on tasks such as instruction following, mathematical reasoning, and safety alignment.
A Sketch-and-Project Analysis of Subsampled Natural Gradient Algorithms
Gil Goldshlager (University of California, Berkeley), Lin Lin (University of California, Berkeley)
Optimization
🎯 What it does: This paper studies the theoretical analysis of the subsampled natural gradient (SNG) algorithm, viewing it as a sketch-and-project method.
A Solvable High-Dimensional Model Where Nonlinear Autoencoders Learn Structure Invisible to PCA While Test Loss Misaligns With Generalization
Vicente Conde Mendes (EPFL), Lenka Zdeborová (EPFL)
Representation LearningData-Centric LearningAuto EncoderContrastive Learning
🎯 What it does: A solvable spike model with two potential factors (one that can be captured by PCA and another that only appears in higher-order moments) was constructed under a high-dimensional assumption, and it was proven that a single-layer nonlinear autoencoder can simultaneously recover both factors, while a linear autoencoder cannot recover the hidden factor.
A Solver-Free Training Method for Predict-then-Optimize
Beichen Wan (University of North Carolina at Chapel Hill), Mo Liu (University of North Carolina at Chapel Hill)
OptimizationFlow-based ModelContrastive LearningTabularFinance Related
🎯 What it does: Propose a solver-free training method based on probabilistic measure transformation, generating a new WISE loss that can directly optimize decision costs under a prediction-first optimization framework.
A Spiking Heterogeneous Harmonic Resonate-and-Fire State Space Model for Time Series
Kartikay Agrawal (IIT Guwahati), Ayon Borthakur (IIT Guwahati)
ClassificationComputational EfficiencyConvolutional Neural NetworkSpiking Neural NetworkTransformerTime Series
🎯 What it does: A second-order pulse state space model based on a hybrid harmonic resonance-discharge (SH2RFSSM) is studied for ultra-long time series classification and regression.
A Statistical Framework for Analyzing Specification Resistance to Learnware-Inversion Risks
Hao-Yi Lei (Nanjing University), Zhi-Hua Zhou (Nanjing University)
Safty and Privacy
🎯 What it does: This paper establishes a statistical framework for analyzing the privacy risks introduced by additional specifications in learnware, particularly providing theoretical guarantees for risks associated with the RKME specification.
A Strictly Proper Scoring Rule and a Calibration Metric for Interval-Censored Data Analysis
Hiroki Yanagisawa (AI Lab, CyberAgent), Shunta Akiyama (AI Lab, CyberAgent)
Explainability and InterpretabilityData-Centric LearningScore-based ModelNeural Radiance FieldAuto EncoderGenerative Adversarial NetworkContrastive LearningTabularBiomedical DataBenchmark
🎯 What it does: Proposed a strict completeness scoring rule and a new calibration metric called IC-Cal for interval-censored data.
A Stronger Benchmark for Online Bilateral Trade: From Fixed Prices to Distributions
Anna Lunghi (Politecnico di Milano), Alberto Marchesi (Politecnico di Milano)
OptimizationReinforcement LearningBenchmarkFinance Related
🎯 What it does: The study focuses on online bilateral trading and designs an algorithm under the global budget balance (GBB) benchmark, proving that it achieves a sublinear regret rate of O~(T^{3/4}) in a random environment.
A Studentized Spherical Harmonics–Based Nonparametric Two-Sample Test for Compositional and Directional Data
Binglin Li (North Carolina A&T State University), Seong-Tae Kim (North Carolina A&T State University)
Data SynthesisAnomaly DetectionRepresentation LearningDiffusion modelScore-based ModelContrastive LearningTabularBiomedical DataReview/Survey Paper
🎯 What it does: A studentized nonparametric two-sample test method based on spherical harmonics is proposed, which can be simultaneously applied to compositional data and directional data, without the need for logarithm ratio transformation or handling of zero values.
A Systematic Study of Behavioral Cloning for Scientific Data Annotation
Ishaan Singh Chandok (Harvard University), Core Francisco Park (Harvard University)
Data SynthesisAI Code AssistantReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningVision Language ModelImageTextSequential
🎯 What it does: Proposed a framework based on behavioral cloning, using synthetic tasks to simulate the scientific data annotation process, studying how models learn GUI interactions and annotation strategies;
A Tale of Two Graphs: Separating Knowledge Exploration from Outline Structure for Open-Ended Deep Research
Zhuofan Shi (Microsoft), Dongmei Zhang (Microsoft)
Autonomous DrivingOptimizationFederated LearningExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningMeta LearningDrug DiscoveryAI Code AssistantReinforcement Learning from Human FeedbackNeural Architecture SearchProtein Structure PredictionGraph Neural NetworkTransformerLarge Language ModelAgentic AIPrompt EngineeringTextGraphBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the DualGraph architecture, using two parallel graphs (Outline Graph and Knowledge Graph) to achieve structured writing and knowledge exploration for open-ended deep research agents, and perform goal search and gap discovery through the Knowledge Graph.
A Tale of Two Problems: Multi-Task Bilevel Learning Meets Equality Constrained Multi-Objective Optimization
Zhiyao Zhang (Ohio State University), Jia Liu (Ohio State University)
OptimizationMeta LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringMixture of ExpertsContrastive LearningTextTabularBenchmark
🎯 What it does: Proposes the multi-task bilevel learning (MTBL) problem under the lower-level general convex (LLGC) assumption, transforms it into an equality-constrained multi-objective optimization (ECMO) framework, provides a Pareto steady-state criterion based on KKT conditions, and designs the WC-Penalty algorithm, proving its finite-time convergence rate of O(S/T^{1/2}) in both deterministic and stochastic environments.
A Task-centric Theory for Iterative Self-Improvement with Easy-to-Hard Curricula
Chenruo Liu (New York University), Qi Lei (New York University)
OptimizationData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Studied the iterative self-improvement mechanism of large language models on mathematical reasoning tasks, and provided task-centric finite sample theory.
A Theoretical Framework for Modular Learning of Robust Generative Models
Corinna Cortes (Google Research), Yutao Zhong (Google Research)
GenerationData SynthesisOptimizationComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelReinforcement LearningMixture of ExpertsDiffusion modelScore-based ModelAuto EncoderGenerative Adversarial NetworkContrastive LearningTextSequential
🎯 What it does: A modular generative model robustness framework based on game theory is proposed, with a feasible gating function designed and the existence of an optimal robust gate proved.
A Theoretical Framework for Statistical Evaluability of Generative Models
Shashaank Aiyer (University of Maryland), Han Shao (University of Maryland)
GenerationData SynthesisExplainability and InterpretabilityScore-based ModelGenerative Adversarial NetworkContrastive Learning
🎯 What it does: Proposes a theoretical framework for evaluating generative models and establishes the assessability results of commonly used evaluation metrics.
A Theory of Contrastive Learning with Natural Images
Antonio Torralba (MIT), Yair Weiss (Hebrew University of Jerusalem)
Representation LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: By analyzing the contrastive learning (CL) loss, this paper derives that the optimal representation under natural images and simple data augmentation is the squared magnitude measurement of frequency, and proves that the optimal network consists of one convolutional layer (sine filter) + squared/ReLU + global average pooling + linear projection, achieving partial whitening;
A Theory of Data Acquisition and Pricing at Scale
Andrew Ilyas (Carnegie Mellon University), Grigoris Velegkas (Google Research)
OptimizationFederated LearningComputational EfficiencyData-Centric LearningContrastive LearningTabular
🎯 What it does: In large-scale machine learning, considering the data procurement and pricing problem under computational constraints, we propose the 'Pricing and Attribution Operator' model, and provide the gradient operator and solving strategy for linear regression.
A Theory of How Pretraining Shapes Inductive Bias in Fine-Tuning
Nicolas Anguita, Clémentine Carla Juliette Dominé
OptimizationFederated LearningComputational EfficiencyKnowledge DistillationRepresentation LearningMeta LearningConvolutional Neural NetworkTransformerSupervised Fine-TuningAuto EncoderContrastive LearningImageText
🎯 What it does: This paper constructs a theoretical model to perform a global analysis of the pre-training and fine-tuning (PT+FT) process, derives the generalization error formula under different initialization and task parameters in diagonal linear networks, and verifies the theory through large-scale experiments.
A theory of learning data statistics in diffusion models, from easy to hard
Lorenzo Bardone (EPFL), Sebastian Goldt (SISSA)
GenerationData SynthesisDiffusion modelScore-based ModelImage
🎯 What it does: Studied the distributional simplification bias of diffusion models when learning data statistical distributions, and constructed a analytically tractable mixed cumulative model and a single-neuron denoiser to theoretically analyze the learning dynamics;
A Tight Theory of Error Feedback Algorithms in Distributed Optimization
Daniel Berg Thomsen (Inria), Aymeric Dieuleveut (Institut Polytechnique De Paris)
OptimizationFederated Learning
🎯 What it does: This paper provides a compact convergence analysis of two main error feedback algorithms (the classical error feedback method and error feedback 21), identifies the optimal step size selection, and constructs optimal Lyapunov functions for each method.
A Time-Reparameterized Cumulative Intensity Extrapolation Sampler for Discrete Flow Matching
Feiyang Fu (Zhejiang University), Hehe Fan (Zhejiang University)
GenerationData SynthesisTransformerDiffusion modelScore-based ModelFlow-based ModelImageTextTime SeriesStochastic Differential Equation
🎯 What it does: This paper proposes a TR-CIE (Time Re-parameterized Cumulative Intensity Extrapolation) sampler based on time re-parameterization for improving sampling quality in discrete flow matching (DFM) with a limited number of model evaluations.
A Two-Layer Framework for Joint Online Configuration Selection and Admission Control
Owen Shen (Massachusetts Institute of Technology), Patrick Jaillet (Massachusetts Institute of Technology)
OptimizationReinforcement LearningTabularTime Series
🎯 What it does: Studied the online configuration selection and admission control problem, and proposed a two-layer framework to handle decision-making within each time period.
A Two-Tier Perspective on Inference-Time Parallelism in Multi-Agent LLM Systems
Zihan Xu (Beijing University of Posts and Telecommunications), Hai Jiang (Beijing University of Posts and Telecommunications)
Computational EfficiencyAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelMixture of ExpertsTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes the TIPEX framework, unifying two-layer parallelism strategies in multi-agent LLM systems—Replica Parallelism and Structural Parallelism—enabling controllable parallel execution while maintaining the original agent pipeline unchanged.
A Unified Approach to Interpreting Knowledge Distillation for Large Language Models via Interactions
Qingzhuo Wang (Tongji University), Zhihua Wei (Tongji University)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
🎯 What it does: This paper proposes using an interaction framework to uniformly explain the knowledge distillation (KD) process of large language models (LLMs), and finds that the core mechanism of KD is compressed interaction, especially suppressing complex interactions while retaining significant simple interactions from the teacher model; based on this, the proposed complex interaction penalty (CIP) loss can explicitly enhance the sparsity of the student model on complex interactions, thereby improving performance;
A Unified Density Operator View of Flow Control and Merging
Riccardo De Santi (ETH Zurich), Andreas Krause (ETH Zurich)
Drug DiscoveryReinforcement LearningDiffusion modelFlow-based ModelGraphBiomedical DataOrdinary Differential Equation
🎯 What it does: This paper proposes a unified probabilistic space framework that integrates reward-driven fine-tuning of pre-trained flow models with multi-model fusion (flow merging), and introduces a Reward-Guided Flow Merging (RFM) algorithm based on mirror descent to implement this framework;
A Unified Framework for Deep Hypergraph Clustering Beyond Homophily
Bowen Zhao (Xidian University), Qianqian Wang (Xidian University)
Representation LearningGraph Neural NetworkContrastive LearningGraph
🎯 What it does: Propose a unified framework Uni-DHC to achieve unsupervised clustering on hypergraphs
A Unified Framework for Diffusion Model Unlearning with f-Divergence
Nicola Novello (University of Klagenfurt), Andrea M Tonello
GenerationData SynthesisSafty and PrivacyDiffusion modelScore-based ModelContrastive LearningImage
🎯 What it does: Propose a unified f-divergence framework for concept forgetting in diffusion models, generalizing the original KL (MSE) loss to arbitrary f-divergence and providing closed-form and variational implementations.
A Unified Sparse Attention via Multi-Granularity Compression
Siran Liu (Peking University), Yongchao He (ScitiX AI)
Computational EfficiencyTransformerLarge Language ModelTextMultimodality
🎯 What it does: Propose a unified sparse attention mechanism called UniSparse, which generates composite tokens through multi-granularity compression and dynamically selects efficient block-level attention.
A unified theory of feature learning in RNNs and DNNs
Jan Philipp Bauer (University College London), Agostina Palmigiano (University College London)
Representation LearningRecurrent Neural NetworkSequential
🎯 What it does: A unified mean-field theory is proposed to describe the representation kernels of recurrent neural networks (RNNs) and deep feedforward networks (DNNs) during the feature learning (µP) phase, revealing how weight sharing leads to temporal consistency and better generalization on sequential tasks.
A Unifying Relational Perspective on Expressive Lottery Tickets
Lorenz Kummer (University of Vienna), Nils Morten Kriege (University of Vienna)
Computational EfficiencyRepresentation LearningDrug DiscoveryGraph Neural NetworkContrastive LearningGraphTime SeriesBiomedical Data
🎯 What it does: Studied the impact of sparse parameters on the expressive power of multi-relational and temporal graph neural networks, and proposed the Relational Strong Expressive Lottery Ticket Hypothesis (RSELTH), proving the existence of sparse subnetworks that retain the 1-RWL expressive power.
A Unifying View of Variational Generative Wasserstein Flows
Paul Caucheteux (ENSAE), Anna Korba (CMAP)
GenerationData SynthesisOptimizationDiffusion modelScore-based ModelFlow-based ModelRectified FlowGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: This paper proposes a unified generative Wasserstein flow (GWF) framework that unifies various JKO-based generative models (such as VWGF, S-JKO, RWP-GAN, etc.) into a single formulation, and extends this framework to the cases of IPM and squared MMD, providing corresponding new algorithms.
A Very Big Video Reasoning Suite
Maijunxian Wang (University Of California Berkeley), Hokin Deng (Carnegie Mellon University)
Data SynthesisExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelDiffusion modelImageVideoTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed the VBVR (Very Big Video Reasoning) toolkit, which includes the large-scale video reasoning dataset VBVR-Dataset (approximately 2 million images, 1 million video clips, 150 reasoning tasks) and a verifiable, interpretable evaluation framework called VBVR-Bench.
A-MemGuard: A Proactive Defense Framework For LLM-Based Agent Memory
Qianshan Wei, XiaoFeng Wang
Safty and PrivacyExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed and implemented A-MemGuard — an active defense framework for LLM agent memory, combining consensus verification and dual memory structures to detect and correct memory injection attacks in real time.
A$^2$SG: Adaptive and Asymmetric Surrogate Gradients for Training Deep Spiking Neural Networks
Yechan Kang (Korea Institute of Science and Technology), Seongsik Park (Korea Institute of Science and Technology)
ClassificationComputational EfficiencyHyperparameter SearchData-Centric LearningConvolutional Neural NetworkSpiking Neural NetworkTransformerContrastive LearningImageVideoStochastic Differential Equation
🎯 What it does: Introducing adaptive and asymmetric surrogate gradients during the training of deep spiking neural networks to enhance training stability and generalization ability
A²RBench: An Automatic Paradigm for Formally Verifiable Abstract Reasoning Benchmark Generation
Qingchuan Ma (Xiamen University), Rongrong Ji (Xiamen University)
TransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Through an automated pipeline, LLMs are used to generate, expand, evaluate, and analyze abstract reasoning tasks, ensuring task uniqueness and formal verifiability via cyclic consistency verification using forward/backward functions.
A3: an Analytical Low-Rank Approximation Framework for Attention
Jeffrey T. H. Wong (Imperial College London), Yiren Zhao (Imperial College London)
CompressionComputational EfficiencyRepresentation LearningTransformerLarge Language ModelAuto EncoderText
🎯 What it does: Propose the A3 low-rank approximation framework, which decomposes the Transformer into three functional components: QK, OV, and MLP, and provides closed-form solutions to reduce the hidden dimension and compress the model.
AAD-1: Asymmetric Adversarial Distillation for One-Step Autoregressive Video Generation
Haobo Li (AutoLab, SAI, Shanghai Jiao Tong University), Zhipeng Zhang (AutoLab, SAI, Shanghai Jiao Tong University)
GenerationData SynthesisKnowledge DistillationTransformerDiffusion modelScore-based ModelAuto EncoderGenerative Adversarial NetworkContrastive LearningVideoOrdinary Differential Equation
🎯 What it does: Propose the AAD-1 framework to achieve single-step autoregressive video generation.
ABC-Bench: An Agentic Bio-Capabilities Benchmark for Biosecurity
Andrew Bo Liu (SecureBio), Seth Donoughe (SecureBio)
Drug DiscoveryAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBiomedical DataBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed and evaluated ABC-Bench, targeting the capability of LLM agents in dual-use biological tasks (fragment design, screening evasion, liquid handling robot programming) and actual experimental validation.
ABCD: All Biases Come Disguised
Mateusz Nowak (Dartmouth College), Peter Chin (Dartmouth College)
Explainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes a bias-free multiple-choice evaluation protocol that reduces label, position, and few-shot distribution biases in LLMs during MCQ evaluation by unifying unordered labels and full-text answer generation.
Abductive Reasoning with Probabilistic Commonsense
Joseph Cotnareanu (McGill University), Mark Coates (McGill University)
Explainability and InterpretabilityKnowledge DistillationReinforcement Learning from Human FeedbackTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: Propose a probabilistic attributable common-sense reasoning framework called PACS, which estimates the truth value of propositions by sampling common-sense beliefs of different individuals using LLMs and logic solvers.
ABSINT-AI: Agentic Heap Abstractions for Abstract Interpretation
Michael Wang (Massachusetts Institute of Technology), Armando Solar-Lezama (Massachusetts Institute of Technology)
Explainability and InterpretabilityComputational EfficiencyAI Code AssistantTransformerLarge Language ModelAgentic AIPrompt EngineeringText
🎯 What it does: Dynamically selects per-allocation-site heap abstraction strategies for JavaScript code using a language model-driven agentic approach, improving analysis precision while maintaining the soundness of abstract interpretation.
Absorbing Quantization Error by Deformable Noise Scheduler for Diffusion Models
Mingrui Yang (AI Chip Center for Emerging Smart Systems), Xiaojuan Qi (The University of Hong Kong)
GenerationData SynthesisTransformerDiffusion modelScore-based ModelFlow-based ModelRectified FlowContrastive LearningImageText
🎯 What it does: This paper studies how errors caused by post-training quantization disrupt the distribution of diffusion models, and proposes a quantization error absorption framework that requires no additional steps and can be used in random, deterministic, and flow matching models.
Abstraction Induces the Brain Alignment of Language and Speech Models
Emily Cheng (Universitat Pompeu Fabra), Richard Antonello (Columbia University)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelContrastive LearningTextBiomedical DataAudio
🎯 What it does: Studied the correspondence between intermediate layers of language models (LLM) and speech models with human brain activity, demonstrating that high intrinsic dimension (Id) and semantic abstraction are key factors driving brain model similarity, rather than merely next-word prediction capability;
AC-ODM: Actor–Critic Online Data Mixing for Sample-Efficient LLM Pretraining
Jing Ma (Renmin University of China), Mingjie Liao (LiblibAI)
Computational EfficiencyData-Centric LearningTransformerLarge Language ModelReinforcement LearningMixture of ExpertsText
🎯 What it does: Propose the Actor-Critic Online Data Mixing (AC-ODM) framework, which dynamically adjusts the data domain sampling weights during the pre-training of LLMs.
Accelerated and Stable Convergence with Anchored Generalized Optimistic Method
Motahareh Sohrabi (Universite de Montreal), Gauthier Gidel (Universite de Montreal)
OptimizationTabularTime SeriesStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Proposed a general optimistic algorithm with anchoring called GOMA for solving monotone Lipschitz variational inequalities, achieving optimal last-iterate convergence rate.
Accelerated Dual Method for Distributed Optimization: An Inexact-Gradient View of Local Updates
Junchi YANG, Feng Qiu (Northwestern University)
OptimizationFederated LearningTabular
🎯 What it does: Proposes the use of Local ADA (Local Accelerated Dual Ascent) method in distributed optimization, allowing an arbitrary number of local stochastic gradient descent steps and synchronization only during communication rounds;
Accelerated Multiple Wasserstein Gradient Flows for Multi-objective Distributional Optimization
Dai Hai Nguyen (Hokkaido University), Hiroshi Mamitsuka (Kyoto University)
OptimizationDiffusion modelScore-based ModelContrastive LearningImageTabularStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Study multi-objective distribution optimization, propose an accelerated variant of MWGraD in the Wasserstein space, A-MWGraD, achieving continuous-time analysis and discrete implementation of multi-objective gradient flows.
Accelerating Langevin Monte Carlo via Efficient Stochastic Runge-Kutta Methods beyond Log-Concavity
Bin Yang (Central South University), Xiaojie Wang (Central South University)
OptimizationComputational EfficiencyScore-based ModelTabularTime SeriesStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes a high-order, Hessian-free Langevin Monte Carlo sampling algorithm based on stochastic Runge–Kutta, which requires only two gradient computations per step, aiming to accelerate sampling from high-dimensional target distributions.
Accelerating Q-learning through Efficient Value-Sharing across Actions
Prabhat Nagarajan (University of Alberta), Marlos C. Machado (University of Alberta)
Reinforcement LearningTabular
🎯 What it does: Propose a parameter-free mean expansion layer (ME layer) that reduces the norm of action values and accelerates Q-learning by sharing a baseline value among actions;
Accelerating Regression Tasks with Quantum Algorithms
Chenghua Liu (Chinese Academy of Sciences), Zhengfeng Ji (Tsinghua University)
OptimizationComputational EfficiencyTabularPhysics Related
🎯 What it does: Proposes a general quantum framework for sparsifying a wide range of generalized linear model (GLM) regression tasks, achieving square quantum acceleration on the sample size m.
Accordion-Thinking: Self-Regulated Step Summaries for Efficient and Readable LLM Reasoning
Zhicheng Yang (Hong Kong University of Science and Technology (Guangzhou)), Jing Tang (Hong Kong University of Science and Technology)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
🎯 What it does: Propose the Accordion-Thinking framework, enabling LLMs to dynamically generate and use concise step summaries during the reasoning process, thereby significantly reducing context length and computational cost while maintaining reasoning accuracy.
Accuracy and Normalized Accuracy under Length Bias: Analysis, Guidelines, and a Bayesian Alternative
Koen Oostermeijer (Aleph Alpha Research)
Explainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelContrastive LearningTextBenchmark
🎯 What it does: Investigated and quantified the length bias in multi-choice evaluation based on log-likelihood, finding that standard unnormalized evaluation favors short answers, while length normalization often overcompensates, favoring long answers; proposed a Bayesian Accuracy (Bayesian Accuracy) scoring rule, which significantly reduces length bias by introducing a length prior and subtracting the global length trend from the log-likelihood.
Accuracy-First Rényi Differential Privacy and Post-Processing Immunity
Ossi Räisä (CISPA Helmholtz Center for Information Security), Antti Honkela (University of Helsinki)
ClassificationData SynthesisSafty and PrivacyAuto EncoderContrastive LearningGaussian SplattingImageTabularStochastic Differential Equation
🎯 What it does: In the paper, the authors propose a new accuracy-priority-based random differential privacy (RDP) framework, proving that the framework possesses post-processing invariance, and apply it to synthetic data generation based on the Brownian mechanism and fine-tuning of image classifiers, demonstrating a significant reduction in privacy budget while meeting the accuracy threshold.
Accurate Evaluation of Quickest Changepoint Detectors via Non-parametric Survival Analysis
Taiki Miyagawa (NEC Corporation), Akinori F. Ebihara (NEC Corporation)
Anomaly DetectionComputational EfficiencyData-Centric LearningTabularTime SeriesBenchmark
🎯 What it does: Proposed nonparametric estimation methods KM-ARL and KM-ADD based on survival analysis, to accurately evaluate the average run length (ARL) and average detection delay (ADD) of rapid change point detectors on real data with limited and irregular sequence lengths
Accurate Large-sample Uncertainty Quantification using Stochastic Gradient Markov Chain Monte Carlo
Yu Wang (Boston University), Jonathan H. Huggins (Boston University)
OptimizationData-Centric LearningTabularTime SeriesSequentialStochastic Differential Equation
🎯 What it does: A discrete-time SG(M)D (SGD, SGLD) uncertainty quantification method is studied and proposed for scenarios with large samples and potential model errors.
Accurate, Private, Secure, Federated U-statistics with Higher Degree
Quentin Sinh (University of Lille), Jan Ramon (University of Lille)
Federated LearningSafty and PrivacyMixture of ExpertsGraphTabularFinance Related
🎯 What it does: In the federated learning environment, a protocol based on multi-party secure computation (MPC) is proposed, which can securely compute arbitrary-order (k ≥ 2) U-statistics under central differential privacy.
Achieving Logarithmic Regret in KL-Regularized Zero-Sum Markov Games
Anupam Nayak (Carnegie Mellon University), Yuejie Chi (Yale University)
OptimizationReinforcement Learning
🎯 What it does: This paper proposes a two-player zero-sum matrix game and Markov game learning algorithm based on KL regularization (OMG and SOMG), achieving logarithmic regret.
Achieving Structurally Robust Gromov Wasserstein Distance via Adaptive Dual-Mask
Kangke Cheng (University of Science and Technology of China), Hu Ding (University of Science and Technology of China)
OptimizationComputational EfficiencyRepresentation LearningData-Centric LearningGraph Neural NetworkContrastive LearningGraphTabularBenchmark
🎯 What it does: This paper proposes the Structural Robust Gromov-Wasserstein distance (SRGW), which adaptively filters edge-level noise during the matching process through a dual masking mechanism, while maintaining strict edge constraints, addressing the global error propagation problem of traditional GW under structural noise.
ACO-MoE-LoRA: Evolving-while-Training for Adapting Segment Anything Model 2 to Specialized Domains
Kaiyi Luo (Soochow University), Jiaqing Fan (Soochow University)
SegmentationDomain AdaptationTransformerSupervised Fine-TuningMixture of ExpertsImageBiomedical Data
🎯 What it does: Propose the ACO-MoE-LoRA framework, combining ant colony optimization with potential space Mixture-of-Experts to implement the 'Evolving-while-Training' strategy, enabling adaptive migration of SAM2 to specialized domains.
ACON: Optimizing Context Compression for Long-horizon LLM Agents
Minki Kang (KAIST), Saravan Rajmohan (Microsoft)
CompressionOptimizationComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: To address the context expansion problem in long-term LLM agents, the ACON framework is proposed, which utilizes natural language optimization compression criteria to achieve context compression.
Acoustic Interference: A New Paradigm Weaponizing Acoustic Latent Semantic for Universal Jailbreak against Large Audio Language Models
Yanyun Wang (Hong Kong University of Science and Technology (Guangzhou)), Li Liu (Hong Kong University of Science and Technology (Guangzhou))
Adversarial AttackTransformerLarge Language ModelPrompt EngineeringDiffusion modelAuto EncoderContrastive LearningAudio
🎯 What it does: This paper proposes the Acoustic Interference (AI) attack, which achieves a non-optimized general jailbreak of large audio language models (LALMs) through the use of Acoustic Latent Semantics (ALS) in harmless audio.
ACTG-ARL: Differentially Private Conditional Text Generation with RL-Boosted Control
Yuzheng Hu (University of Illinois Urbana Champaign), Peter Kairouz (Google Research)
GenerationSafty and PrivacyReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBiomedical DataElectronic Health Records
🎯 What it does: Propose a hierarchical framework ACTG-ARL for generating controllable text under differential privacy.
Action Manifold Smoothing: A Lipschitz Pathway Perspective on High-Dimensional Reinforcement Learning
Zhihao Lin (University of Glasgow)
Reinforcement Learning
🎯 What it does: Propose Action Manifold Smoothing (AMS) by using orthogonal neighborhood averaging on the TD target to smooth the Q function, thus addressing the discrete-continuous mismatch problem in high-dimensional continuous control.
Action-Sufficient Goal Representations
Jinu Hyeon (Seoul National University), Taesup Moon (Seoul National University)
Reinforcement LearningAuto EncoderContrastive LearningTabularTime SeriesSequential
🎯 What it does: Investigate the sufficiency of subgoal representations in offline goal-conditioned reinforcement learning, introducing the concept of 'action sufficiency' and proving its superiority.
Activation Oracles: Training and Evaluating LLMs as General-Purpose Activation Explainers
Adam Karvonen (MATS), Samuel Marks (Anthropic)
Explainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Train and evaluate Activation Oracles (AO), models capable of answering any question about LLM activation vectors in natural language; and evaluate them on various discrete-outside tasks such as 'secret retention' and error alignment.
Activation with Intrinsic-Extrinsic Consensus
Tian Qiu (Zhejiang University), Mingli Song (Zhejiang University)
ClassificationExplainability and InterpretabilityComputational EfficiencyConvolutional Neural NetworkTransformerAuto EncoderContrastive LearningImageTextTabularTime SeriesSequentialBenchmark
🎯 What it does: Proposed a new activation mechanism called AIEC, which identifies and suppresses irrelevant feature channels by leveraging the consensus between channel activation statistics (intrinsic) and linear decision weights (extrinsic).
Activation-Free Backbones for Image Recognition: Polynomial Alternatives within MetaFormer-Style Vision Models
Jeffrey Wang (University of Wisconsin-Madison), Grigorios Chrysos (University of Wisconsin-Madison)
ClassificationSegmentationComputational EfficiencyTransformerContrastive LearningImage
🎯 What it does: Designed and implemented a polynomial visual backbone called PolyNeXt, which replaces traditional ReLU/GELU and softmax exponential operations with Hadamard product to construct PolyMLP, PolyConv, and PolyAttn modules, and embed them into the MetaFormer framework;
Active Attacks: Red-teaming LLMs via Adaptive Environments
Taeyoung Yun (KAIST), Minsu Kim (KAIST)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringText
🎯 What it does: Proposed an active attack framework (Active Attacks) that automatically generates diverse and effective red team attack prompts by periodically performing safe fine-tuning on the victim LLM and resetting the attacker LLM.
Active Budget Allocation for Efficient Scaling Law Estimation via Surrogate-Guided Pruning
Viktoria Schram (University of Melbourne), Trevor Cohn (University of Melbourne)
Computational EfficiencyRepresentation LearningHyperparameter SearchMixture of ExpertsTextTabular
🎯 What it does: This paper proposes an active budget allocation strategy based on Successive Halving and surrogate models, aimed at efficiently obtaining learning curves under limited computational budgets, thereby enabling rapid estimation of scaling laws.
Active Continual Learning with Metaplastic Binary Bayesian Neural Networks
Kellian Cottart (Universite Paris-Saclay, CNRS, Centre de Nanosciences et de Nanotechnologies), Damien Querlioz (Universite Paris-Saclay, CNRS, Centre de Nanosciences et de Nanotechnologies)
ClassificationRecognitionAnomaly DetectionComputational EfficiencyKnowledge DistillationMeta LearningTransformerContrastive LearningImageTabularBenchmark
🎯 What it does: Proposed a margin-friendly continual learning method called BiMU, which utilizes a variational objective with bounded memory to achieve online, buffer-free continual learning and active querying for binary Bayesian neural networks (BNNs).
Active Curriculum Refinement for Reinforcement Learning
Zhenya Liu (University of Chicago), Yuxin Chen (University of Chicago)
Reinforcement LearningTabularTime SeriesSequential
🎯 What it does: This paper proposes the PATH framework, which improves the training efficiency of reinforcement learning by utilizing a predefined curriculum DAG through random path exploration and regret-based active path sampling.
Active Exploring like a Pigeon: Reinforcing Spatial Reasoning via Agentic Vision-Language Models
Wei Deng (State Key Laboratory of Networking and Switching Technology), Mengshi Qi (State Key Laboratory of Networking and Switching Technology)
Autonomous DrivingRobotic IntelligenceTransformerSupervised Fine-TuningReinforcement LearningAgentic AIPrompt EngineeringVision Language ModelSimultaneous Localization and MappingImageTextMultimodality
🎯 What it does: Proposed an active exploration framework based on a vision-language model, utilizing a dynamic cognitive map and spatial assertion code (SAC) to achieve spatial reasoning;
Active Learning with Foundation Model Priors: Efficient Learning under Class Imbalance
Jiancheng Zhang (University of California Riverside), Yinglun Zhu (University of California Riverside)
ClassificationData-Centric LearningConvolutional Neural NetworkRecurrent Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningImageText
🎯 What it does: Proposed an active learning framework based on foundation model priors, which selects clean samples and performs uncertainty sampling on noisy samples through joint decision-making with a small model in scenarios where class imbalance and label noise coexist.
Active Learning with Low-Rank Structure for Data Selection
Vincent Cohen-Addad (Google Research), Samson Zhou (Texas A&M University)
OptimizationData-Centric LearningSupervised Fine-TuningContrastive LearningTextTabularFinance Related
🎯 What it does: This paper proposes an active learning data selection framework based on low-rank structures, constructing a core set using low-rank approximation and residual sampling.
Active Policy Optimization for Individualized Dosing via Gradient Variance Minimization
Yi Wan (University of Science and Technology of China), Huanhuan Chen (University of Science and Technology of China)
OptimizationDrug DiscoveryReinforcement LearningTabularBiomedical Data
🎯 What it does: This paper proposes a personalized dose strategy learning framework tailored for budget-constrained environments, advocating for directly minimizing the posterior gradient variance to drive sample selection, rather than traditional global prediction uncertainty;
Active Regression for Single-Index Models with Unknown Link Functions
Chansophea Wathanak In (Nanyang Technological University), Xuan Wu (Shanghai Jiaotong University)
OptimizationData-Centric LearningTabular
🎯 What it does: This paper studies the active regression problem under a single index model (with an unknown 1-Lipschitz link function) in the setting where only the label coordinates are queryable.