Training & post-training
Reinforcement learning
Learning behaviour through rewards associated with actions or outcomes.
Also known as: RL
Definition
Reinforcement learning optimises a policy by increasing the likelihood of actions that receive higher rewards in an environment.
Why it matters
Reward design can improve reasoning and agent behaviour but may also encourage shortcuts, gaming or unintended optimisation.
Related concepts
- Reinforcement learning from feedback
An umbrella for reinforcement learning guided by preference or quality feedback.
- Reward model
A model estimating how desirable another model's output or action is.
- Reward hacking
Optimising a measured reward without achieving the intended objective.