From trait name to vector: extraction becomes automatic
Building a steering vector for a new behavior used to mean hand-assembling a fresh set of contrastive examples every time. This theme covers what changes once that step is automated: give a pipeline nothing but a trait's name and a short description, and it manufactures its own contrastive prompts, test questions, and scoring rubric, then extracts a working direction from the model's own activations.
The page walks through both stages, an LLM writing the raw materials from one template, then a second stage generating rollouts under those materials, filtering them by how strongly each one shows the trait, and averaging the activation difference between what passes. It also places the pipeline against its closest prior-work relative, and points to what automation actually buys: the same machinery running unchanged across dozens of traits.
The automated persona vector extraction pipeline is the 2025 paper's answer to a problem CAA left unsolved: building a steering vector for a new behavior required a hand-curated contrastive dataset for each one. This pipeline needs only a trait name and a short description. Its first stage, trait-specific artifact generation, hands that description to Claude Sonnet, used throughout the paper as a pipeline and data-generation tool rather than an experimental subject, which writes five pairs of contrastive positive/negative system prompts, forty elicitation questions split into extraction and evaluation sets, and a scoring rubric, all from one generic prompt template. The rubric feeds the trait expression score, an LLM judge (GPT-4.1-mini) that rates how strongly a response displays the trait on a 0-100 scale and is used both to filter the pipeline's own rollouts and, throughout the rest of the paper, as its primary outcome measure. The second stage, response-conditioned contrastive vector computation, generates rollouts under the positive and negative system prompts, filters them by trait expression score, and takes the mean difference of their response-token activations to produce the persona vector itself. The pipeline's closest relative is Wu et al.'s automated concept-description-to-direction pipeline, which similarly turns an arbitrary natural-language concept description into contrastive generations and a linear direction; the paper positions its own pipeline as the more thoroughly validated, trait-specific descendant of that approach. What the automation buys is scale: because extraction no longer needs bespoke data curation, the same machinery runs over dozens of traits without new human effort, which is the precondition for everything downstream -- monitoring, screening, mitigation -- being trait-general rather than built one behavior at a time.