Why GRPO: The SFT vs. RFT vs. RLHF Decision Tree

~15 min read

GRPO exists to solve reasoning-heavy tasks where correctness is verifiable — the book's own decision tree shows exactly when GRPO-style RFT is the right call versus SFT or RLHF.

Why GRPO: The SFT vs. RFT vs. RLHF Decision Tree is a Pro topic

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

Key points

  • GRPO answers a specific branch of a larger decision tree: no labeled data, but the task IS verifiable -> RFT (GRPO is a leading RFT method)
  • No labeled data, task NOT verifiable -> RLHF instead (needs human preference signals)
  • Labeled data exists: large dataset -> SFT; tiny dataset where reasoning helps -> RFT/GRPO
  • SFT trains on static data and often memorizes answers; RFT is online, exploring outputs and learning from a reward signal
  • GRPO uses deterministic reward functions to check correctness automatically, eliminating the need for labeled data entirely