“UC Berkeley's research showed that LLMs used as judges have a self-preference bias, where a Llama model will favor answers from other Llama-family models.”