Xiaomi MiMo's CodeMidas Scales Agentic Coding RL Environments from Source Code Alone
Xiaomi's MiMo team, with Peking University and other institutions, released the paper CodeMidas (CodeMIDAS: Scaling Agentic Coding RL Environments from Code Itself) on HuggingFace around September 21. It proposes an agentic pipeline that takes only source code as input and automatically converts already-implemented functionality in existing repositories into executable RL training environments, backed by large-scale results.
Confirmed
- CodeMidas was proposed by the Xiaomi MiMo team with PKU and others; the paper is public on Hugging Face and was relayed by @XiaomiMiMo, @eliebakouch, @teortaxesTex, @markk and @akhaliq.
- Pipeline: agents inspect open-source projects, extract coding tasks, build execution-based tests, and discard broken or exploitable tasks, yielding verified RL environments.
- Scale: 5,545 verified RL training environments generated from 3,185 projects across 23 programming languages.
- Per a repost by @maierak, DeepSWE improved from 10% to 21.7% after training with CodeMidas.
Unconfirmed
- The DeepSWE 10%→21.7% figure appears only in a repost; the exact benchmark and evaluation setup await paper details.
Why it matters
- Existing RL-for-coding data methods depend on issues, commits and other development artifacts, limiting task sources to repositories with complete dev histories; CodeMidas needs only source code, greatly expanding the usable task pool.
- Automatic filtering of broken and exploitable tasks helps ensure execution-based reward quality and mitigates reward hacking.
- Large-scale, multilingual verified environments offer new data infrastructure for training general coding agents.
2026-09-21 ~ 2026-09-22 · 6 related posts
Primary sources
- Xiaomi's CodeMidas builds 5,545 coding RL environments from raw source code — XiaomiMiMo ·
- Xiaomi's CodeMidas turns open-source code into 5,545 RL environments, boosting coding agents up to 17% — eliebakouch ·
- Xiaomi's CodeMidas mines open-source code into 5,545 verified RL environments, boosting MiMo-V2.5 benchmarks — mark_k ·
- [source] Xiaomi's CodeMidas builds 5,545 coding RL environments from raw source code — XiaomiMiMo · 2026-09-21
- Xiaomi's CodeMidas turns source code into RL environments: 5,545 tasks, 23 languages — teortaxesTex · 2026-09-21
- Xiaomi MiMo's CodeMIDAS scales agentic coding RL environments from code itself — _akhaliq · 2026-09-22
3 near-duplicate retellings: mark_k · eliebakouch · maier_ak