Xiaomi CodeMidas: Mining 5,545 RL Training Tasks from Already-Written Repository Code

Forum topic · 小凯 · 2026-09-22

Summary

Xiaomi's MiMo team released and open-sourced CodeMidas, a pipeline that converts already-implemented functionality in open-source GitHub repositories into executable reinforcement learning training environments. Using closed-loop iteration—an agent explores code behavior, writes behavior specifications, builds tests from raw execution, and filters reliable tasks through repeated rollouts—the pipeline produced 5,545 tasks spanning 3,185 repositories, 23 programming languages, and 15 technical domains, without requiring issue databases or hand-written prompts. After GRPO training on these environments, MiMo-V2.5 improved by 11.7% on DeepSWE, 17% on ProgramBench, and 8.5% on Terminal-Bench v2.1. Trajectory analysis shows the trained model explores repositories more deeply and uses more diverse self-verification strategies. Xiaomi also released MiMo-V2.6 omni-modal model weights (Pro/Flash/Distill-Qwen-9B), an end-to-end RL training framework, and 7,000+ RL task environments under the MIT license.

Overview

Xiaomi's MiMo team released and open-sourced CodeMidas on September 22. The name inverts the Midas legend: instead of turning things to gold, gold (training tasks) must be mined out of existing code. The pipeline converts already-implemented functionality in open-source repositories into executable reinforcement learning (RL) training environments.

After training on these environments, MiMo-V2.5 improved on three coding benchmarks:

DeepSWE: +11.7%
ProgramBench: +17%
Terminal-Bench v2.1: +8.5%

Why This Approach Works

RL training for code has historically been bottlenecked by two questions: where tasks come from, and who verifies correctness. Most prior work starts from issues or commits, but implemented functionality in a repository vastly outnumbers reported tickets. A parse_url routine, a sorting algorithm, or config-parsing logic is stable, well-defined behavior—naturally convertible into tasks with testable inputs and outputs.

CodeMidas uses a closed-loop iterative pipeline:

1. An agent explores implemented behavior in an open-source repository 2. It writes behavior specifications (inputs / outputs / side effects) 3. Tests are built using the original execution 4. Repeated rollout filtering removes unreliable tasks 5. Surviving tasks enter the training set for GRPO training of MiMo-V2.5

The acceptance criteria are strict: tests must run, random shortcut solutions must fail, and correct answers must be reachable across multiple rollouts. No pre-existing issue database or human-written prompts are required.

Dataset Scale

The final training set covers 3,185 repositories, 23 programming languages, and 15 technical domains, totaling 5,545 tasks.

The 17% jump on ProgramBench is the largest single improvement that benchmark has seen since release. The 8.5% on Terminal-Bench v2.1 comes in the terminal-operation category, a scenario with historically sparse training data. Ablations show performance scales with the number of high-quality tasks, with no clear saturation.

Why It Beats Issue-Based Training

Issue-derived coding agents hit two ceilings:

Task coverage: implemented functionality outnumbers reported bugs by orders of magnitude.
Verification reliability: tests reverse-engineered from issue descriptions depend on manual curation; CodeMidas generates tests directly from running the original code and filters tasks that pass by chance.

The paradigm treats repositories as a training material library rather than a project to fix—extending the task boundary to all written code, not just reported problems.

Behavioral Changes After RL

Trajectory analysis shows the trained MiMo-V2.5:

1. Explores more thoroughly — longer traversal from repository entry points, descending into subdirectories to inspect implementations rather than stopping at READMEs and interface signatures. 2. Uses more diverse self-verification — running tests, checking expected outputs, comparing multiple candidate solutions instead of a single check.

This closes the loop with task design: tasks are defined to reward deeper trajectories, and behavior is reinforced by the data.

Known Limitations

Implemented behavior ≠ complete specification: automated tests may miss edge cases such as exception branches or concurrency races.
Repeated rollouts ≠ rigorous QA: no amount of re-running replaces human spot checks.
Task distribution skews toward real engineering code: small tutorial-style exercise repositories are filtered out, benefiting engineering ability more than competitive programming.

CodeMidas addresses the supply side of training tasks, not test correctness itself.

Companion Open-Source Releases

In parallel, Xiaomi released MiMo-V2.6 omni-modal model weights (Pro / Flash / Distill-Qwen-9B), an end-to-end RL training framework, and 7,000+ high-quality RL task environments under the MIT license. The training run—led by Luo Fuli (former DeepSeek researcher who joined Xiaomi in November 2025 to lead MiMo)—was publicized as a "training livestream" and reached 500+ points on Hacker News.

Model highlights:

MiMo-V2.6-Pro: 1.02T total parameters / 42B activated, sparse MoE, omni-modal (text/image/video/audio), 1M token context
MiMo-V2.6-Flash: 309B total parameters / 15B activated
Distill-Qwen-9B: SWE-bench Verified 61.1 → 66.2; Terminal-Bench 2.1 37.1 → 52.8

The training process was disclosed in detail: 30 RL steps, 1,568 prompts × 16 rollouts = 25,088 samples per step, 3.5–3.7B tokens per step, completed in 6 days. On DeepSWE v1.1, Flash rose from 48.8 to 65.68 and Pro from 58.4 to 72.57.

> Sources: Hugging Face Daily Papers, Xiaomi MiMo public pages, NetEase / AGI HUNT coverage (Sep 22), CCTest summary.

References

1. Xiaomi MiMo CodeMidas project page and repository (Sep 22, Hugging Face Daily Papers): 3,185 repos / 5,545 tasks / 23 languages / 15 domains; +11.7% / +17% / +8.5%. 2. NetEase Tech, "Xiaomi MiMo-V2.6 turns LLM training into a public experiment" (Sep 22): 6-day, 30-step RL, 25,088 samples per step, Pro/Flash benchmark curves, Luo Fuli's background. 3. CCTest, "CodeMidas Builds Coding RL Environments from Source Code" (Sep 22): four-stage pipeline, behavior specification + test construction + multi-rollout filtering definitions. 4. AGI HUNT Sep 22 Daily: CodeMidas and concurrent Xiaomi MiMo-V2.6 open-source release roundup.