The new CodeMidas system from Xiaomi allows coding agents to autonomously create their own training environments by mining open-source software. This pipeline generated 5,545 verified reinforcement learning (RL) environments from 3,185 projects across 23 programming languages. The software inspects open-source code to extract tasks and build execution-based tests, discarding broken or exploitable tasks to eliminate the need for humans to write RL tasks by hand.

Training the MiMo-V2.5 model on these tasks improved whole-program construction by 17%, issue repair by 11.7%, and Terminal-Bench performance by 8.5%. The process uses agents in the loop to evaluate task difficulty via pass@n metrics and test robustness by allowing agents to attempt to cheat the environment.