机器学习：2025–2026 年度综述

An annual survey of the cs.LG arXiv category, covering research published from September 2025 – August 2026. Built on a close-reading set of 168 papers and organised into 46 chapters, it maps the year by the field's research sub-directions.

Book framework (share of the year's output):

• Reinforcement Learning and Time Series Forecasting (30.4%) — This subfield focuses on applying deep reinforcement learning and graph-based methods to dynamic temporal data, multi-agent coordination, and time series foundation models.
• Physics-Informed and Interpretable Predictive Modeling (15.8%) — This area addresses the integration of physical constraints and interpretability into machine learning models for complex prediction tasks, risk estimation, and scientific discovery.
• Federated Learning and Online Decision Making (13.2%) — This subfield explores distributed learning paradigms, including federated learning, knowledge distillation, and online policy optimization under heterogeneous data conditions.
• Generative AI and Causal Discovery in Autonomous Systems (9.7%) — This direction investigates diffusion models, causal inference, and autonomous agent control for applications in drug discovery, navigation, and complex system simulation.
• Large Language Models and Agentic Reasoning (9.2%) — This hot topic centers on the deployment, evaluation, and enhancement of large language models, focusing on reasoning capabilities, agent frameworks, and benchmarking.
• Vision-Language Models and Synthetic Data Generation (5.6%) — This subfield deals with the training and application of vision-language models, in-context learning, and the generation or evaluation of synthetic data for scalable AI.
• Uncertainty Quantification and Domain Adaptation (4.8%) — This area focuses on estimating uncertainty, adapting models to domain shifts, and utilizing ensemble or retrieval-based methods for robust generalization.
• Transformer Architectures and Symbolic Regression (3.1%) — This subfield examines the theoretical and practical aspects of transformer mechanisms, attention efficiency, and the application of symbolic regression in physical sciences.
• Representation Learning and Hierarchical Clustering (2.7%) — This direction explores latent space representations, self-supervised contrastive learning, and hierarchical structures for molecular, geometric, and network data.
• Parameter-Efficient Fine-Tuning and Feature Selection (1.4%) — This subfield addresses techniques for efficient model adaptation, such as low-rank adaptation, weight pruning, and feature selection in high-dimensional spaces.
• Efficient Inference and Bayesian Compression (1.1%) — This area focuses on optimizing model inference speed and size through compression, pruning, and Bayesian methods for resource-constrained environments.
• Adversarial Robustness and Security (0.6%) — This subfield investigates defenses against adversarial attacks, data poisoning, and evasion techniques to ensure the security and robustness of machine learning systems.
• Evaluation Frameworks and Predictive Control (0.6%) — This direction develops systematic evaluation metrics, control policies, and reliability assessments for AI systems, particularly in traffic and conversational contexts.
• Quantum Machine Learning and Active Learning (0.3%) — This niche area explores the intersection of quantum computing and machine learning, including quantum circuit simulation and active learning with noisy labels.

Each sub-direction is covered across three chapters — how it framed its questions, the strength of its key findings, and its open debates and frontiers — so the reader gets a traceable picture of what this part of the field actually produced over the year.