https://arxiv.org/pdf/2308.04592.pdf
https://github.com/facebookresearch/Shepherd
Simple methods can enhance the performance of LLMs (such as GPT-4) using iterative refinement. One approach involves having an LLM critique generated text and then self-improve based on that feedback. In this study, the authors developed a small (7B, thus fast and cheap to run) model which surpasses other competing models, including ChatGPT.