$ cat wiki/papers/2026/2609.22068-codemidas.md

CodeMidas: Scaling Agentic Coding RL Environments from Code Itself

paperupdated 2026-09-22created 2026-09-22

agents coding-agents agentic-rl rl benchmarks

TL;DR

An agentic pipeline that turns existing source code — and nothing else — into executable RL environments for coding agents, discarding the usual dependence on issues and commits. 5,545 training tasks from 3,185 codebases, 23 languages. Training MiMo-V2.5 with GRPO on them improves all five benchmarks tried, including Terminal-Bench v2.1 +8.5%.

Authors & Org

Not stated. The HuggingFace Daily Papers snapshot carries the arXiv id, title, upvote count, publication date and abstract; it carries no author list and no affiliation, and arxiv.org answers EGRESS_BLOCKED from this run's sandbox. Recorded as unknown rather than guessed.

Method

The stated premise is a limit on existing practice: open-source codebases are a rich source of RL tasks, but existing methods rely on development artifacts such as issues and commits, which bounds what can be extracted to what somebody happened to write down. CodeMidas takes source code as its only task-specific input (source).

Agentic compute is allocated to every stage of environment construction, not only to solving:

agents explore implemented functionality to formulate behavioral specifications;
they construct tests grounded in execution of the original code;
candidates are validated and filtered through execution checks and repeated solution rollouts.

The resulting dataset: 5,545 training tasks from 3,185 open-source codebases, spanning 23 programming languages and 15 technical domains.

Results

Training MiMo-V2.5 on these tasks with GRPO. Figures as published, all as deltas against the untrained baseline:

Benchmark	Task family	Gain
DeepSWE	issue repair	+11.7%
ProgramBench	whole-program construction	+17%
Terminal-Bench v2.1	terminal work	+8.5%
"Improves performance on all five diverse benchmarks" is stated; **the
abstract names three of the five** and gives no figure for the other two, so no
aggregate is recorded here.

Two further claims:

Ablation: increasing the number of high-quality training tasks improves performance — a scaling claim about the environment set rather than the model.
Trajectory analysis: the RL-trained agent shows increased codebase exploration and more diverse self-verification.

No absolute score appears anywhere in the abstract — every number is a relative gain, so none of these is comparable to any other figure on this wiki.

Significance

This is the fourth paper in six days on the same move, and the cluster is the finding rather than any one result: the artefact is the specification. ProgramDistill: From Interactive Web Apps to Verifiable Reference-Guided SWE Tasks (written up 09-21) makes a running reference application the spec; 2609.22000 RecreationWorld makes a running reference the oracle for hidden behavioural tests; 2609.05571 Code2Skill turns 19,769 repositories into a skill bank; CodeMidas turns source code into RL environments. All four remove the human-written task statement, and three of the four remove the human from construction entirely.

What separates CodeMidas from the other three is which side of the loop it is on. ProgramDistill and RecreationWorld build evaluations; CodeMidas builds training environments from the same substrate. That is the same asymmetry Agentic Reinforcement Learning has been recording — verifiable environments are the bottleneck on agentic RL, and the supply of them has been bounded by human annotation.

It also lands on Eval Harness Configuration from an unusual angle: if the environment is synthesised by an agent, the harness is now part of the training data, not only part of the measurement.

Open Questions

Which two of the five benchmarks are unreported, and what they moved by. "All five improve" with three figures published is a claim this wiki cannot check.
Absolute scores. Without a baseline number, the +8.5% on Terminal-Bench v2.1 cannot be placed against the 52.8 / 69.7 / 73.0 figures this wiki already holds for that benchmark on Ternary Bonsai 2 27B and Qwen 3.8 27B.
Whether the gains survive a second base model. Only MiMo-V2.5 is reported.
Contamination. The training tasks are built from open-source codebases and the evaluation benchmarks are built from open-source codebases; the abstract states no overlap check.
What the filter rejects. "Repeated solution rollouts" as a validity filter selects for tasks the current model can sometimes already solve, which is a bias in the environment set the abstract does not discuss.