Xiaomi's CodeMidas turns source code into RL environments, doubling DeepSWE to 21.7%
maier_ak · x · 2026-09-22
Xiaomi's MiMo team released CodeMidas, an agentic pipeline that converts existing open-source codebases into executable RL training environments for coding agents.
- Key idea: prior methods rely on development artifacts like issues and commits; CodeMidas uses source code alone. Agents explore implemented functionality to draft behavioral specs, build tests grounded in the original code's execution, then validate and filter tasks via execution checks and repeated solution rollouts.
- Scale: 5,545 training tasks from 3,185 open-source repos across 23 programming languages and 15 technical domains.
- Results: training MiMo-V2.5 with GRPO improves all five benchmarks, including DeepSWE +11.7% (reported as doubling from 10.0% to 21.7%), ProgramBench +17%, and Terminal-Bench v2.1 +8.5%.
- Why it matters: a scalable way to generate RL environments for coding agents without needing issue/commit metadata. Paper: arXiv 2609.22068.
More from coding & agent
- Jev-style models called a major unlock for fast, cheap browser agents — multiply_matrix · 2026-09-22
- Generating responsive UIs in under 2 seconds with shadcn plus Mobbin MCP — msharmas · 2026-09-22
- Nat Friedman admits Meta's Muse was inspired by OpenClaw, bought hundreds of Mac minis for the team — EdwardSun0909 · 2026-09-22
- SkillLift cuts agent skill-evolution token cost 40-70% by ranking, not rollouts — dair_ai · 2026-09-22
- Deel launches Akai: 8,000 agents doing work of ~600 employees, added $140M ARR in 90 days — AIwithGhotai · 2026-09-22
- exe.dev wins over developers: SSH into root VMs, plus the underrated Shelley coding agent — davidcrawshaw · 2026-09-22