Evaluating steering: behavior scores and capability floors
Showing that a steering vector "works" actually requires answering two different questions, and it is easy to conflate them: did the targeted behavior really move, and did anything else quietly break in the process? This theme keeps those two questions apart and follows the paper's instruments for each, since a vector that shifts sycophancy but wrecks the model's general competence would not be much of a result.
The page covers the behavior-side evidence first, a multiple-choice sweep across layers and a GPT-4-judged open-ended writing test, both showing where and how strongly a vector moves its target. It then turns to the capability-side evidence, a broad knowledge benchmark used as a floor that steering leaves essentially untouched, and a narrower truthfulness check that moves in exactly the direction the paper's own explanation of sycophancy predicts.
Showing that a steering method works takes two separate kinds of evidence, and this theme groups the paper's instruments for each. Multiple-choice behavioral evaluation is the primary test that the behavior actually moved: for each behavior, a steering vector is swept across every layer at multipliers of -1 and +1, and effect size is scored as the average token probability given to the behavior-matching answer on 50 held-out questions, a sweep that surfaces a consistent optimal layer -- 13 in the 7B model, 14 or 15 in the 13B model. Open-ended generation evaluation extends that test beyond contrived multiple-choice questions to free-form answers, with GPT-4 rating each response 1-10 on how strongly it displays the targeted behavior using a rubric prompt written per behavior, a rating whose reliability the authors cross-check by hand against manual inspection. GPT-4 is the rater doing that scoring in both settings. The second kind of evidence is that nothing else broke: MMLU, a 57-subject multiple-choice benchmark, is used unmodified as a capability floor, and the paper reports no significant effect on the model's average probability of giving the correct answer. TruthfulQA supplies a targeted truthfulness cross-check specific to one behavior, finding that subtracting the Sycophancy vector mildly improves performance and adding it mildly worsens it, in the direction the paper's own misgeneralization hypothesis about sycophancy predicts. The 2025 persona-vectors paper builds its own instruments on this same two-part pattern, one behavior score and one capability floor, but automates the rater itself. Trait expression score is that paper's behavior-score instrument: a 0-100 rating of how strongly a response displays a target trait, produced not by sampling a rating from the judge but 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, a method validated against two human judges at 94.7% agreement before being used as the paper's primary outcome measure throughout. Coherence score plays MMLU and TruthfulQA's role as the paper's capability floor: a 0-100 general-coherence rating, following Betley et al. (2025)'s methodology, that steering and finetuning experiments are required to keep above roughly 75-80 as evidence an intervention has not degraded the model's coherence rather than genuinely moved a trait. GPT-4.1-mini, OpenAI's smaller GPT-4.1 variant, is the rater behind both scores, and doubles in the paper's data-screening appendices as an LLM-based filter for samples that overtly exhibit a target trait.