Huawei and CUHK Release Lego-RL for Coding Agent RL Training
青稞AI · wechat · 2026-08-29
Huawei and The Chinese University of Hong Kong (CUHK) released Lego-RL, a reinforcement learning training framework designed for Coding Agents. Its key innovation enables direct RL training on native harnesses (like OpenHands, ClaudeCode) without code modification, addressing the performance degradation seen in traditional methods that alter control flows.
Key Value & Data:
- On SWE-bench Verified, using Qwen2.5-35B-A3B, Lego-RL boosted scores for OpenHands/ClaudeCode/OpenCode from 64.0/62.4/57.2 to 70.4/68.2/66.6. The gain exceeds that of upgrading directly to the Qwen3.6 Base model.
- Revealed that different harnesses cause a 7 point variance, proving RL training must occur on the exact harness used in deployment.
Three Technical Pillars:
- Faithful Optimization: Uses an in-process Proxy and R3 (Rollout Routing Replay) to fix token misalignment caused by runtime history rewriting and MoE routing drift, boosting training/inference correlation to 0.9993.
- Reliable Execution: Features sandboxed execution with anti-cheat measures, task difficulty filtering, and proper handling of infrastructure failures to ensure Reward signal integrity.
- Observable Training: Provides AgentPlugin and LiveUI for real-time diagnostics to distinguish between policy degradation and environment issues.
Architecture:
Built on verl (training) and Harbor (sandbox execution), connected via AgentLoopWorker and an in-process Proxy. An adapter pattern allows instant support for any harness compatible with OpenAI/Anthropic protocols.
More from coding & agent
- Multi-agent self-improvement loop: scorer agents grade past runs, then auto-PR AGENTS.md fixes — blaizedsouza · 2026-09-23
- Austrian builder event demos liquid rockets, SLAM robots, and production-ready multi-agent systems — DominiqueCAPaul · 2026-09-23
- Xiaomi Open-Sources MiMo-V2.6: 1.02T MoE Multimodal Agent Family Under MIT — burny_tech · 2026-09-23
- Theo: running 40+ agent threads is fine — watching multiple at once is 'the ultimate LARP' — iamsahaj_xyz · 2026-09-23
- Yacine: Astra is 'ridiculously good' at perf optimization — may demote it to a subagent — yacineMTB · 2026-09-23
- Yacine: coaxed Astra into decent code by painstaking task-splitting — DS-Flash would've been faster — yacineMTB · 2026-09-23