[5J2-OS-31a] OS-31

1:00 PM - 1:15 PM JST(4:00 AM - 4:15 AM UTC)

[5J2-OS-31a-05] Fault localization using LLM with automatic test oracle generation

〇Hiroki Futami¹, Yasuyuki Tahara¹, Yuichi Sei¹, Ryohei Orihara¹ (1. The University of Electro-Communications)

Keywords:

software,Large language model,fault localization

This paper propose a method to generate a test oracle from source code using LLM and identify the location of bugs. The proposed method generates a code specification from the source code, its calling relationships, comments, and method names. Based on the generated specification and source code, tests and expected output are generated. LLM is used to infer the actual output from the test input values, and by comparing this with the expected output, the probability that the method is buggy is inferred. This method takes test traces into account, but by not executing tests, it succeeds in reducing computational costs and addressing the problem of inexecutable generated test code.