Reinforcement learning (RL)

Reinforcement learning trains a model by rewarding good outcomes, such as a correct answer or passing tests, instead of copying example text.

  • The basics
  • Updated

In ordinary training a model learns to imitate text. In reinforcement learning it tries something, a score says how well it went, and the model is nudged toward the attempts that scored higher. The score can come from a checker, for example whether code passes its tests or a maths answer is right, or from a model that stands in for human preferences.

Many recent reasoning models and coding agents are post-trained this way; Xiaomi's MiMo V2.6 and DeepReinforce's Ornith both describe reinforcement-learning runs. A reward is only as good as its checker, so RL tends to improve a model on the kinds of tasks it was rewarded for. Results are usually reported by the developers themselves.

Related terms

  • Reasoning model A reasoning model works through a problem step by step in a 'thinking' phase, visible or hidden, before it gives its final answer.
  • Fine-tuning Fine-tuning continues a pretrained model's training on your own examples so that it behaves better for a specific task, style or domain.
  • AI agent An AI agent is a system in which a model plans and carries out multi-step tasks by calling tools, such as running code, searching the web or editing files.

All terms