Why does the model I use for evaluation prefer the model it came from?
機会
LLM-as-judge has become the default way to run quality evaluations at scale, but judges from the same model family rate their own outputs 10 to 25 percent higher than equivalent outputs from a different family. Multiple 2025 and 2026 papers confirm the bias holds across position, verbosity, and rubric-constrained formats, and simple mitigations like panel voting reduce but do not eliminate it. Existing measurement frameworks cannot separate genuine quality differences from stylistic self-matching, so the effect cannot be corrected without first being cleanly isolated. Any team using the incumbent model to evaluate a replacement is running an evaluation with a structural thumb on the scale.
重要な理由
The entire AI improvement loop depends on honest quality signals, and a biased evaluator corrupts that signal at the source without raising any alert.
機会をどう評価するか
Opportunity Scoreは測定値ではなく、私自身の見解です。どれほど痛みを伴うか、どれほど頻繁に影響を与えるか、そして今日時点で解決策がいかに少ないか。スコアが高いほど、構築する価値が高いと私は考えています。
それが現れたときにどれほどの痛みをもたらすか。
実際にどれほど頻繁に人々がそれに直面するか。
今日時点で、それに対する優れたツールがいかに少ないか。