The paper analyzes the effectiveness of self-repair in large language models (LLMs) for code generation, specifically GPT-3.5 and GPT-4, on a challenging dataset called APPS. The study finds that self-repair is only effective in GPT-4 and is bottlenecked by the feedback stage. Using expert human programmers to give feedback on the programs generated by GPT-4 unlocks significant performance gains.

https://arxiv.org/abs//2306.09896

YouTube: https://www.youtube.com/@ArxivPapers

PODCASTS:

Apple Podcasts: https://podcasts.apple.com/us/podcast/arxiv-papers/id1692476016

Spotify: https://podcasters.spotify.com/pod/show/arxiv-papers