Policy Gradient

Writing1 post

• May 31, 2026

  GRPO From First PrinciplesWhy On-Policy RL for LLMs Is Really Just Dynamic Weighted SFT

  A first-principles walk from supervised fine-tuning to policy gradient and GRPO — why on-policy RL for LLMs is really dynamic, group-normalized, reward-weighted SFT.

  18 min read

  • GRPO
  • RLHF
  • LLM
  • Reinforcement Learning
  • Policy Gradient