The paper shows SAO outperforms GRPO on SWE-Bench Verified.
14.1K
4 posts, first seen 5h ago
New GLM paper on the PPO algo they use for GLM 5.2 Clearly the PPO hype was overblown (DIS here is Icepop without recompute which we first did in Intellect-3) This also comes with 3x longer trainer steps, 2x memory on trainer (with equal inference compute), and extra compute to train the value model before RL starts Wouldn't be surprised if taking that into account makes GRPO >> Also clearly SWE vs AIME type RL didnt have a difference much, so the statements going around on long horizon and tool calling being key for PPO also seem overblown
Users in the replies criticized PPO's inefficiencies in the GLM paper, calling out its need for 3x steps, 2x memory, and a separate value model versus GRPO and SAO as brutal.
Based on 1 visible X reactions from 2 accounts; directional sample.
Ask a question below.
Published answers will appear here.