Bilkent University
Department of Computer Engineering
CS 590/690 SEMINAR
Interpretable Automated Patch Correctness Assessment Using LLM-Based Code Clone
Sahand Moslemi Yengejeh
Master Student
(Supervisor: Asst.Prof.Anıl Koyuncu)
Computer Engineering Department
Bilkent University
Abstract: Automated Program Repair (APR) has seen significant advancements, but overfitting patches remain a critical challenge, threatening the reliability and practical application of these techniques [4]. Plausible patches generated by APR tools may pass all test suites yet fail to capture the intended semantics, requiring expert validation. Automated Patch Correctness Assessment (APCA) approaches aim to mitigate this burden, especially in oracle-based APCA, where the goal is to determine semantic equivalence between an APR-generated patch and a human fix. However, according to Rice’s theorem [2], achieving a generalizable and efficient solution remains computationally infeasible, making human evaluation indispensable [3]. Building on the potential of large language model (LLM)-based code clone detection [1], we introduce a novel method to enhance APCA by identifying clone relationships between patches. By categorizing clones into four types, we capture semantic equivalences (type-1 and type-4) and subtle variations (type-2 and type-3) that reveal patch characteristics. Our approach employs a rule-based representation learner [5], trained on agent clone labels, to determine patch correctness. The proposed method offers explainable agent labels, equipping APR users with actionable insights for informed decision-making. This enables more accurate assessments of patch validity while reducing dependency on exhaustive human evaluations.
DATE: December 2, Monday @ 15:30 Place: EA 502