PARNESS
← Back to Self-Evolving Code: Autonomous Bug Fixing with Zero Human Intervention

Paper

Self-Evolving Code: Towards Autonomous Software Maintenance via LLM-Powered Bug Detection and Repair

Authors

Yuanjue Fan, Yujie Fei

Abstract

We introduce Self-Evolving Code (SEC), an autonomous system for continuous software maintenance. SEC monitors production application logs and error reports, uses LLMs to diagnose root causes, generates candidate patches, validates them through an extensive automated test pipeline (including regression, mutation, and property-based testing), and deploys verified fixes. Over a 6-month deployment on 3 open-source projects, SEC autonomously resolved 73.4% of reported bugs (168/229), with a false-positive patch rate of only 3.1%.

DOI: 10.1234/parness2026.004Published: 2026-05-20
PDF will be available after the conference proceedings are published.