# Preprint reports coding gains from verified refactoring training

_Published Wednesday, October 28, 2026 at 3:06 AM EDT · AI, Science · Latest · Tier 2 — Notable_

Researchers report that targeted refactoring training improved a 92B-parameter coding model serving more than 100,000 developers. Their preprint describes RefactorForge, a pipeline that uses software engineering tools and verification techniques to build executable training tasks from Python repositories.

Across 306 repositories, the pipeline retained 4,572 verified teacher trajectories totaling 72.51 million tokens. Additional supervised training used tokens equal to 0.46% of the preceding coding mixture and improved RefactorBench performance by 18.00 percentage points and SWE-Bench ProMax by 7.06 points.

Under OpenCode, gains extended to issue resolution: 4.97 points on SWE-Bench Verified and 1.23 points on SWE-Bench Pro. Results for generating new repositories were mixed.

## Sources

- [cs.LG updates on arXiv.org](https://arxiv.org/abs/2610.33851)

---
Canonical: https://techandbusiness.org/newswire/f4giA3A8JLaMZFc2zZ8WMQ
Published: 2026-10-28T07:06:08.722Z
Story chronology: 2026-10-28T04:00:00.000Z
Retrieved: 2026-10-28T09:36:36.087Z
Publisher: Tech & Business (techandbusiness.org)