从最初的RLHF（人类反馈强化学习），转向RLVR（基于可验证反馈强化学习），再转向前沿的“自然语言奖励”。

Enhancing Math Reasoning in Small-sized LLMs via Preview Difficulty-Aware Interv...

活性物质是指由自我驱动实体组成的系统，这些实体消耗能量以产生运动，表现出复杂的非平衡动力学，挑战了传统模型。随着机器学习的快速发展，强化学习（RL）已成为应对活...

分层强化学习（HRL）通过分解复杂的决策过程，在长时域和稀疏奖励任务中表现出有效性，但由于层级间不稳定、子目标调度效率低下、响应延迟以及可解释性差等问题，其在现...

Quantum Reinforcement Learning by Adaptive Non-local Observables

Deeply felt affect- the emergence of valence in deep active inference

A Survey of Reinforcement Learning for Large Reasoning Models

元强化学习教程 A Tutorial on Meta-Reinforcement Learning

A Survey of Reinforcement Learning for Optimization in Automation

Lifelong Reinforcement Learning with Similarity-Driven Weighting by Large Models

Are Reasoning Models More Prone to Hallucination?

https://www.nature.com/articles/s42256-025-00983-2

在后期训练中，我们超越了标准的有监督微调。我们实施了一个顺序的强化学习流程——从推理强化学习开始，接着是智能体强化学习，最后是通用强化学习。

Epistemically-guided forward-backward exploration

基于人类反馈的强化学习（RLHF）已成为部署最新机器学习系统的重要技术和叙事工具。在本书中，我们希望为具备一定量化背景的读者提供对核心方法的简明介绍。本书首先回...

Provably Safe Reinforcement Learning: Conceptual Analysis, Survey, and Benchmark...

Towards Large-Scale In-Context Reinforcement Learning by Meta-Training in Random...

DIME:Diffusion-Based Maximum Entropy Reinforcement Learning DIME：基于扩散的最大熵强化学习

Reinforcement Learning for Reasoning in Large Language Models with One Training ...

SFT Memorizes, RL Generalizes: A Comparative Study of Foundation Model Post-trai...