policy

Definition

In reinforcement learning, the function mapping a state to a distribution over actions — the thing being learned. Policy-gradient methods update it directly; value-based methods derive it from an estimated return. When an LLM is tuned with RL, the model itself is the policy and tokens are its actions. Unrelated to the everyday sense of a rule or guideline.

Explainers using this term

• Paper Walkthrough: One Training Example Keeps On-Policy Distillation Improving for Hundreds of StepsRethinking On-Policy Distillation of Large Language Models II: One Training Example
• Paper Walkthrough: SecOPD — Grading One Token at a Time to Cut Adaptive Prompt Injection by an Order of MagnitudeSecOPD: Mitigating Adaptive Prompt Injections by On-Policy Distillation
• Paper Explained: Does On-Policy Distillation Really Distill? From Noisy Teacher to Self-ImprovementDoes On-Policy Distillation Really Distill? From Noisy Teacher to Self-Improvement