CodeMidas: Scaling Agentic Coding RL Environments from Code Itself

Forum topic · 小凯 · 2026-09-22

Summary

CodeMidas is an agentic pipeline that converts implemented functionality in existing open-source codebases into executable reinforcement learning (RL) environments, using source code as its only task-specific input. Unlike prior methods that depend on development artifacts such as issues and commits, CodeMidas allocates agentic compute to every stage of environment construction: agents explore implemented functionality to formulate behavioral specifications, build tests grounded in execution of the original code, and validate and filter candidate tasks through execution checks and repeated solution rollouts. The resulting dataset contains 5,545 training tasks from 3,185 open-source repositories, spanning 23 programming languages and 15 technical domains. Training MiMo-V2.5 with GRPO on these tasks yields improvements across five diverse benchmarks, including issue fixing (DeepSWE +11.7%), full program construction (ProgramBench +17%), and terminal operation (Terminal-Bench v2.1 +8.5%). Ablations show more high-quality tasks improve performance, and trajectory analysis reveals better RL-trained agent behaviors such as more repository exploration and diverse self-verification.

Paper Overview

Research Area: ML

arXiv: 2609.22068

Abstract

Training capable coding agents via reinforcement learning (RL) requires diverse tasks with reliable verifiers. Open-source codebases offer a rich source of such tasks, while existing methods typically rely on development artifacts such as issues and commits, limiting the range of tasks that can be extracted. To better scale RL environments, the authors present CodeMidas, an agentic pipeline that turns implemented functionality in existing codebases into executable RL environments using source code as its only task-specific input.

Key Points

CodeMidas allocates agentic compute to every stage of environment construction:
Agents explore implemented functionality to formulate behavioral specifications.
Tests are constructed grounded in execution of the original code.
Candidate tasks are validated and filtered through execution checks and repeated solution rollouts.
The resulting dataset contains 5,545 training tasks from 3,185 open-source codebases, spanning 23 programming languages and 15 technical domains.
Using these tasks to train MiMo-V2.5 with GRPO yields improvements on all five diverse benchmarks:
Issue fixing: DeepSWE +11.7%
Full program construction: ProgramBench +17%
Terminal operation: Terminal-Bench v2.1 +8.5%
Ablation studies show that increasing the number of high-quality training tasks improves performance.
Trajectory analysis shows RL-trained agents exhibit better behaviors, such as more repository exploration and more diverse self-verification.

Conclusion

These results establish source code itself as a scalable foundation for building RL environments that broadly improve coding agent performance, without relying on issues, commits, or other development artifacts.