Mastering HumanEval with Reflexion

This is a spin-off project inspired by the paper: Reflexion: an autonomous agent with dynamic memory and self-reflection. Noah Shinn, Beck Labash, Ashwin Gopinath. Preprint, 2023

Read more about this project in this post

Check out an interesting type-inference implementation here: OpenTau

Check out the code for the original paper here

Check out a new superhuman programming agent gym here

If you have any questions, please contact noahshinn024@gmail.com

Note

Due to the nature of these experiments, it may not be feasible for individual developers to rerun the results due to limited access to GPT-4 and significant API charges. Due to recent requests, both trials have been rerun once more and are dumped in ./root with a script here to validate the solutions with the unit tests provided by HumanEval.

To run the validation on your log files or the provided log files:

python ./validate_py_results.py <path to jsonlines file>

Warning

Please do not run the Reflexion agent in an unsecure environment as the generated code is not validated before execution.

Cite

Note: This is a spin-off implementation that implements a relaxation on the internal success criteria proposed in the original paper.

@article{shinn2023reflexion,
  title={Reflexion: an autonomous agent with dynamic memory and self-reflection},
  author={Shinn, Noah and Labash, Beck and Gopinath, Ashwin},
  journal={arXiv preprint arXiv:2303.11366},
  year={2023}
}