Presentation Information
[B-14-14]A Study on Reviewability metrics for AI Code Review Based on Static Analysis and Test Results
〇huayang gong1, hideo makino1, toyoki yamauchi1 (1. SoftBank)
Keywords:
Eevelopment Methodology,Machine Learning
With the widespread adoption of LLM-based coding assistants, the amount of automatically generated source code in software development has rapidly increased. However, the quality of AI-generated code is not always consistent, creating a need for a framework that can quantitatively evaluate and continuously improve code review quality. This study proposes Reviewability metrics, a quality evaluation metric that integrates quantitative indicators obtained from static analysis and test results. The proposed method normalizes and combines multiple quality indicators into a single reviewability score, which is used as feedback for AI-based code review and code refinement. By iteratively performing quality evaluation, code review, code modification, and re-evaluation, the proposed framework aims to establish a continuous quality improvement loop for AI-assisted code review. This approach provides a quantitative and systematic framework for evaluating and improving the quality of AI-assisted code review.