强化学习Reinforcement Learning (RL)

靠奖励试错学习的范式，与监督、无监督并列的三大训练范式之一。Learning by trial and error against rewards: one of the three training paradigms alongside supervised and unsupervised learning.

强化学习研究智能体如何与环境交互、靠最大化累计奖励学会决策。

它不同于监督学习（有标签答案）和无监督学习（无标签找结构）。

训练三阶段里的 RLHF，就是这个范式用在语言模型微调上。

Reinforcement learning studies how an agent learns to act by interacting with an environment and maximizing cumulative reward.

It differs from supervised learning (labeled answers) and unsupervised learning (finding structure without labels).

RLHF in the three-stage training recipe is this paradigm applied to fine-tuning language models.

聊天模型的「对齐」阶段用的就是它。懂「按奖励学」，才懂模型为什么被练得爱顺着人。The alignment stage of chat models runs on it. Understanding 'learning against rewards' explains why models are trained to please.

去 Models 图谱：看这个模型有哪些量化版、按你的显存该下哪一档Models atlas: which quantisations exist and which tier fits your VRAM ↗

回到名词地图看第1站的其他 34 个词 →Back to the map: the other 34 terms in station 1 →

• Sutton & Barto《Reinforcement Learning: An Introduction》第 2 版（MIT Press 官方页）Sutton & Barto, Reinforcement Learning: An Introduction, 2nd ed. (MIT Press) ↗