Explore ability to improve generation using Self-Rewarding

[pramitchoudhary]

From understanding there r 2 parts of it,

1. LLM-as-a-judge, to generate reward score. As first step evaluate this strategy can help during re-generation request with exhaustive beam search.

2. Self-training/modification on preference pairs

Reference: https://arxiv.org/abs/2401.10020