The Evolution of RL: From PPO to MaxRL in LLM Reasoning Training

强化学习的进化：从PPO到MaxRL，LLM推理训练的算法演进史

Jiqizhixin translated Alexander Weers' article on RL algorithms for LLM reasoning from 2024 to 2026. It covers REINFORCE, PPO, GRPO, RLOO, Dr. GRPO, DAPO, CISPO, MaxRL, DPPO, and ScaleRL, comparing critic removal, clipping, normalization, and pass@k goals. The key signal is mechanism choice, not algorithm names.

Why it matters: A strong technical explainer, not a model or paper release. HKR-H comes from the PPO→MaxRL arc, HKR-K from concrete mechanism comparisons, and HKR-R from live RL-recipe choices; the higher technical bar keeps it in low featured.