This paper introduces Meta’s TestGen-LLM tool, which utilizes LLMs to enhance human-written tests automatically. TestGen-LLM ensures that the generated test classes meet specific criteria for improvement over the original test suite, preventing issues caused by LLM hallucination. The tool was deployed at Meta test-a-thons for Instagram and Facebook, demonstrating success in generating accurate test cases and increasing test coverage. It significantly improved a notable percentage of classes during the test-a-thons, with the majority of its recommendations being accepted for production deployment. This marks a significant milestone in the industrial-scale deployment of LLM-generated code.
https://arxiv.org/abs/2402.09171