State of Deep Learning for Code Generation (DL4Code)
There have been a set of recent interesting results that apply large language models (e.g., GPT-3) to code generation (a.k.a program synthesis, program generation). The results so far have been competitive and some models have been integrated into developer productivity tools such as GitHub Copilot - an AI pair programming tool that supports programmers live as they write code (hint: it works quite well).
This post seeks to curate a running list of interesting papers, associated models and datasets that have been used to address this task.
Several papers and models have been published.