implement REFINER

[sandys]

https://github.com/debjitpaul/refiner

https://arxiv.org/abs/2304.01904

LLMs Getting Fine-Grained Reasoning Feedback Improves Performance

-Critic model trained to provide feedback reasoning LLM uses to iteratively improve intermediate arguments -Easily beats baselines -Even w/ GPT3.5, critic significantly improves reasoner