AA-PersonalityBench

AA-PersonalityBench measures model personality through writing style. It rates how prominent each trait is, not how good the answer is: the same trait can be welcome or not depending on the reader.

• The set has 30 questions per trait, 900 in total. Some gently encourage the trait, others are neutral. It stays private apart from a few samples.
• Each model answers every question once, as a single user message with no system prompt, so the answer shows how it writes by default.
• For each trait, a judge model from a panel compares two blinded answers to the same question and picks the one showing the trait more, or a tie.
• A Firth-penalized Bradley-Terry model fits the verdicts, a tie counting as half a win for each side. GPT-6.1 Sol (max) anchors every trait at 1,000 Elo; a 400-point gap means 10-to-1 odds.
• An Elo can be negative. For "Sycophantic" and "Slop", a higher Elo means more flattery or more filler.
• Artificial Analysis groups the traits into eight indexes, each the mean of its traits' z-scores, percentile ranked from 1 to 100; "Sycophantic" counts inversely toward Transparency. It publishes those indexes only for the models on its charts, so BenchSift does not recompute them and shows the per-trait Elos only.