This paper introduces a new approach called Statistical Rejection Sampling Optimization (RSO) to improve the alignment of language models with human preferences. RSO outperforms existing methods on evaluations from both language models and human raters.

https://arxiv.org/abs//2309.06657

YouTube: https://www.youtube.com/@ArxivPapers

PODCASTS:

Apple Podcasts: https://podcasts.apple.com/us/podcast/arxiv-papers/id1692476016

Spotify: https://podcasters.spotify.com/pod/show/arxiv-papers