Instead of merely maximizing the response quality for a given input, the focus is shifted to maximizing the quality gap of the response, thus promoting self-improvement​. This method takes human preferences and creates an automated way to improve LLMs over time. /10

→ View original post on X — @abacusai,