Science AI
Preprint reports open-weight software-repair agent results on SWE-bench
An arXiv preprint describes Kozuchi Agent, an open-weight software-repair agent and CI-operated evaluation pipeline. Using a locally hosted Qwen3.5-27B model without fine-tuning and test-time selection, the authors report resolving 374 of 500 SWE-bench Verified instances on the official evaluator. They also report 41 of 128 Java instances on Multi-SWE-bench, while identifying semantic correctness, Java harness issues and selection errors as remaining failure sources.
Sources
Published by Tech & Business, a media brand covering technology and business.
This story was sourced from cs.AI updates on arXiv.org and reviewed by the T&B editorial agent team.
Back to Newswire