Reinforcement Fine-Tuning (RFT): Reward Signals Instead of Static Labels

~13 min read

RFT replaces fixed labeled completions with an online reward signal: the model explores different outputs, a reward function scores correctness, and the model learns to generate higher-reward answers via GRPO.

Reinforcement Fine-Tuning (RFT): Reward Signals Instead of Static Labels is a Pro topic

Sign in, then upgrade to Pro or Power to unlock this topic and the full AI Engineering curriculum.

Key points

  • RFT's process (per the book): no static labels — the model explores different outputs, a Reward Function scores correctness, and it learns via GRPO to favor higher-reward answers
  • 'Online' (RFT) vs 'static' (SFT) is the core contrast: RFT evaluates the model's own live-generated attempts rather than training against a fixed pre-written dataset
  • You only need a function that can RECOGNIZE a correct outcome (e.g. checking a final numeric answer), not one that specifies the single correct path to it
  • This is why RFT pairs naturally with reasoning tasks (math, logic) that have automatically-checkable outcomes, via deterministic reward functions
  • The book's summary: SFT memorizes static answers; RFT, being online, learns from rewards and explores new strategies — at the cost of more compute and needing a reliable reward function