RL lifts Qwen 397B agent Pass@1 ~70% on APEX-Agents: SkyRL x Mercor full training postmortem
青稞AI · wechat · 2026-09-18
Edward Hu (first author of LoRA, ex-OpenAI o1 core member) joined Mercor as Head of AI Modeling and partnered with the UC Berkeley SkyRL team on RL for complex knowledge-work agents. Qwen3.5-397B-A17B improved Pass@1 on APEX-Agents from 16.11% to 27.29% (70% relative), while the smaller Qwen3.6-35B-A3B beat Claude Opus 4.5 after RL. Key engineering lessons: harness-only fixes added 5.95 points (22.74%→28.69%), TITO ensures inference/training token consistency to avoid off-policy drift, and fully async RL plus dynamic micro-batching maximizes throughput on long trajectories.
More from coding & agent
- One stale date silently broke a WhatsApp reply agent: 19 messages unanswered — petrusenko_max · 2026-11-02
- The 10X engineer is now just a SKILLS.md: 6 workflow rules for AI coding agents — mdancho84 · 2026-11-02
- Auditing my MCP credential broker: the worst hole was the config file — Relevant_Mousse_9117 · 2026-11-02
- AWS's open-source Strands Harness claims 45% cheaper than Claude Code, reigniting harness debate — blaizedsouza · 2026-11-02
- Xiaomi MiMo's CodeSpecWeave trains coding agents on interdependent specs for long-horizon coding — XiaomiMiMo · 2026-11-02
- Sketchy v0.1.6: AI diagrams use 29% fewer output tokens than plain text — Nkgsporty · 2026-11-02