Vibeleaderboard
← All Intel
Intel / article

Kozuchi Agent: A Language-Agnostic Open-Weight Agent for Software Repair

Source
arxiv.org
Author
Mehdi Bahrami, Kosaku Kimura, Satoshi Munakata, Satoshi Nakashima, Yu Ishikawa, Kosuke Maeda, Nao Soma, Kenichi Kobayashi, Keisuke Miyazaki, Keizo Kato, Shigeki Fukuta, Tatsuo Kumano, Nobutaka Imamura, Kevin Musgrave, Shahbaz Abdul Khader, Kwun Ho Ngan, Joe Townsend, Fayas Asharindavida, Matthieu Parizy, Akira Sakai, Yuma Ichikawa, Yang Zhao, Michiaki Takizawa, Taku Fukui, Hiroki Ohtsuji, Wei-Peng Chen, Hiromichi Kobashi
Date

Context

Kozuchi tackles software repair by generating several possible fixes and testing them against one another's checks. Using a locally hosted Qwen3.5-27B model without , eight independent runs each produce a patch and tests. A selector applies those tests across the candidate patches before choosing one, without consulting the hidden benchmark tests. Selection method.

The selected patches resolved 374 of 500 issues, or 74.8%, on the official evaluator. Across all eight runs, at least one successful patch existed for 408 issues. That gap shows why generating a correct fix and recognizing which fix to submit are separate problems, even with substantial repeated computation. Evaluation.

Terms in this piece · Glossary
  • AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
  • fine-tuning — Taking a trained model and training it a bit more on your own examples so it gets better at one specific job.
  • SWE-bench — The standard benchmark for AI coding agents: real GitHub issues from real repositories, scored by whether the agent's patch passes the project's own tests.
  • open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.
Recommended reads
Comments

Checking sign-in…

Loading comments…