CodeMidas: Scaling Agentic Coding RL Environments from Code Itself
CodeMidas builds coding-agent RL tasks directly from existing source code, without relying on issues or commits.
The paper describes an agentic pipeline that turns implemented behavior in open-source repositories into executable training environments. It produces 5,545 tasks from 3,185 codebases across 23 languages and 15 domains. Training MiMo-V2.5 on the resulting tasks improved scores on five benchmarks, including DeepSWE, ProgramBench, and Terminal-Bench v2.1. The authors say larger sets of high-quality tasks drove better results, with trained agents exploring codebases more and using more varied self-checks. ArXiv · AI/CL/LG's note
The paper describes an agentic pipeline that turns implemented behavior in open-source repositories into executable training environments. It produces 5,545 tasks from 3,185 codebases across 23 languages and 15 domains. Training MiMo-V2.5 on the resulting tasks improved scores on five benchmarks, including DeepSWE, ProgramBench, and Terminal-Bench v2.1. The authors say larger sets of high-quality tasks drove better results, with trained agents exploring codebases more and using more varied self-checks. ArXiv · AI/CL/LG's note
score 7