Mercor Releases Open RL Guide for 397B Qwen Model

Robert Nishihara posted about an agentic RL training guide for the Qwen 3.5 397B model created with SkyRL. The materials, linked from a tweet by edwardjhu, are presented as coming from Mercor Research and include training scripts for applying DPPO. The post states the approach yields very large improvements when trained on roughly 2000 expert-labeled tasks. No further confirmation or independent details appear in the packet.