self-preference

LLM judges show position bias, favoring earlier answers over content

Large language models used as judges in evaluation processes exhibit a significant "position bias," where they favor answers presented earlier in a list, regardless of content quality. This bias, along with verbosity an…