Panickssery

LLM evaluation panels show self-preference bias, favoring own outputs

A developer discovered that their LLM evaluation panel exhibited a significant self-preference bias, where models favored their own generated outputs over others, regardless of quality. This bias, documented in a NeurIP…