26 Aug 2024
Researchers at Federal University of Minas Gerais and Pontifical Catholic University of Minas Gerais conducted a comprehensive study of GPT-3.5 Turbo's ability to generate Java code documentation, finding that 69.7% of generated comments were equivalent or nearly equivalent to human-written ones, with 22.4% rated as superior. The study critically demonstrated that automated metrics like BLEU scores are insufficient for accurately assessing the quality and utility of LLM-generated code documentation.
There are no more papers matching your filters at the moment.
