Position Bias

A known systematic flaw in the LLM-as-a-Judge evaluation paradigm, whereby the judge model tends to assign higher scores to answers that appear earlier in a conversation or that are longer in length, irrespective of actual content quality.