“Research from UC Berkeley found that LLMs used as judges exhibit several biases, including a positional bias (preferring the first answer) and a verbosity bias (preferring more verbose answers).”