Trait expression score (LLM judge)
Persona Vectors: Monitoring and Controlling Character Traits in Language Models — introduced
A 0-100 score quantifying how strongly a model response displays a target trait, computed by prompting GPT-4.1-mini with a trait-specific rubric and taking a logit-weighted sum over the single-token integer completions in its top-20 output logits. Validated against two human judges (94.7% agreement) and against external benchmarks before being used as the paper's primary outcome measure throughout.
Deciding how strongly a single response displays a trait like sycophancy or hallucination sounds like an inherently fuzzy judgment call, and asking a model to just blurt out a number invites exactly the kind of noisy, one-off guess that judgment calls tend to produce. This score tries to firm that up by reading a judge model's full distribution over plausible ratings rather than trusting whatever single digit it happens to sample.
Retrieving a logit-weighted number rather than a single sampled guess is the core mechanism, checked afterward against human raters and against outside benchmarks the paper didn't build itself. The same score turns out to do two different jobs — grading finished responses throughout the paper, and filtering out mismatched rollouts before any vector is even extracted.
A 0-100 signal read from a model's own logits
Trait expression score is the paper's primary behavioral measurement: a 0-100 number quantifying how strongly a single model response displays a target trait. Scoring works by prompting GPT-4.1-mini with the trait's own rubric and asking it to output a bare number, but the paper does not simply take whatever token the judge samples. Instead it retrieves the top-20 logits from the judge's output distribution, isolates the subset corresponding to integers 0-100 (known to tokenize as single tokens in OpenAI's tokenizer), and computes a logit-weighted sum over just those candidates, following the scoring methodology of Betley et al. (2025) (persona-vectors, §"B.1 Evaluation scoring details", p. 28). That choice trades a single discrete sample for a continuous readout blending the judge's whole distribution over plausible scores, which is what makes the score a stable, comparable signal across responses rather than a noisier single-draw guess.
Validating the judge: humans and external benchmarks
Because nearly every result in the paper depends on this metric, it is validated directly against human judgment. Two of the paper's authors served as human judges, each shown 50 randomly sampled response pairs per trait — one response with an LLM-assigned score above 80, one below 20 — and asked to pick the one that exhibits the trait more strongly. Agreement between the LLM judge and the human judges was 97% for evil, 92% for sycophancy, and 95% for hallucination, 94.7% overall across 300 pairwise comparisons (persona-vectors, §"B.2 Checking agreement between human judges and LLM judge", p. 28). The paper also checks that its own 20-question evaluation set behaves like established external benchmarks: for hallucination, scores on the first 1,000 questions of the QA split of HaluEval correlate strongly with scores on the paper's own evaluation questions, and similarly for sycophancy against the held-out questions from Nishimura-Gasparian et al. (2024)'s dataset (persona-vectors, §"B.3 Additional evaluations on standard benchmarks", p. 29).
Two roles: outcome measure and upstream filter
Trait expression score plays two different roles across the paper. Most often it is a downstream outcome measure — the y-axis of steering, finetuning, and monitoring results throughout. But inside vector computation itself it becomes an upstream data-quality gate: rollouts generated under a contrastive system prompt are scored against the rubric immediately, and only those exceeding 50 (for positive-prompted responses) or falling below 50 (for negative-prompted ones) are kept before any activation is averaged (persona-vectors, §"2.2 Extracting persona vectors", p. 3). The score's validity as an outcome measure is also protected by a companion metric: coherence score, scored by the same judge model but against a fixed, trait-agnostic rubric, gates whether a trait-expression result is trusted at all, since a steering intervention that degrades a response into gibberish could otherwise register an extreme trait score for the wrong reason.