Type: url
Original: https://arxiv.org/abs/2306.14898
Status: generated
Summary
The arXiv abstract page for InterCode, a framework that casts interactive coding as a standard reinforcement learning environment with code as actions and execution feedback as observations, using self-contained Docker environments. The authors build Bash, SQL and Python environments from the NL2Bash, Spider and MBPP datasets, evaluate several LLMs with prompting strategies such as ReAct and Plan & Solve, and present InterCode as a benchmark and testbed for interactive code generation that can be extended to tasks such as Capture the Flag.