Sign InOpen Brain
arXivPaperNeeds Review

CodeMidas: Scaling Agentic Coding RL Environments from Code Itself

CodeMidas derives verified coding-agent RL tasks directly from working source, widening training data beyond issues and commits while improving repair, construction, and terminal benchmarks.

arXiv · Sep 18, 2026
Open Source Open MarkdownOpen JSON
Source Summary

CodeMidas turns existing implementations into executable RL environments using source code as the only task-specific input. Its agents derive specifications, build execution-grounded tests, and filter tasks through checks and repeated solution attempts, producing **5,545 tasks from 3,185 repositories**.

Practical Implication

For coding-agent training, treat working code as a source of behavioral tasks even when issue and commit histories are sparse. MiMo-V2.5 trained with GRPO improved across all five tested benchmarks, including **+17% on ProgramBench** and **+11.7% on DeepSWE**.

Agent-Ready Context
CodeMidas turns existing implementations into executable RL environments using source code as the only task-specific input. Its agents derive specifications, build execution-grounded tests, and filter tasks through checks and repeated solution attempts, producing **5,545 tasks from 3,185 repositories**.

For coding-agent training, treat working code as a source of behavioral tasks even when issue and commit histories are sparse. MiMo-V2.5 trained with GRPO improved across all five tested benchmarks, including **+17% on ProgramBench** and **+11.7% on DeepSWE**.

The abstract reports that more high-quality tasks helped and that trained agents explored and self-verified more, but it does not establish how well the pipeline handles underspecified, unsafe, or hard-to-execute repositories.
Connected Context · Feed7 Judgment

CodeMidas broadens executable task generation beyond repositories with usable issue or PR histories by deriving behavior from implementations themselves, and supplies training gains rather than only a data-pipeline proposal. It complements history-derived environments, while leaving task validity, unsafe repositories, execution feasibility, and supervision quality as unresolved filters rather than solved properties.

Change2Task: From Repository Changes to Executable Coding Agent Tasks and EnvironmentsChange2Task derives executable tasks from merged changes, whereas CodeMidas uses implementations alone, making the approaches complementary for repositories with rich versus sparse histories.MindForge: Teaching Small Language Models Whole-Life-Cycle Software Engineering via Source-Free Program SynthesisBoth report ProgramBench gains from constructed coding environments, but MindForge reconstructs programs from documentation and behavior while CodeMidas derives tasks from existing source.SWE-Prime: Fewer Trajectories, Better PerformanceSWE-Prime shows that successful trajectories still require quality filtering, an implementation consequence for CodeMidas after its task filters and repeated solution attempts produce training traces.State of Data — Sean Cai, Independent / State of DataThe real-trace thesis provides a useful contrast: CodeMidas demonstrates gains from manufactured source-derived tasks, but does not establish that they capture the behavior or scaffold dependence of live work.
Context Map
benchmarkcoding#coding-agents#agent-evals#harness-engineering
Uncertainty
The abstract reports that more high-quality tasks helped and that trained agents explored and self-verified more, but it does not establish how well the pipeline handles underspecified, unsafe, or hard-to-execute repositories.