Automated persona vector extraction pipeline — adds filtering and validation beyond → Automated concept-description-to-direction pipeline (Wu et al., 2025)

Wu et al. (2025) is the paper's own named closest relative: an automated pipeline that also turns a natural-language concept description into contrastive pairs of LLM-generated text and eventually into a linear direction, with no manual dataset curation required (persona-vectors, §"2 An automated pipeline to extract persona vectors", p. 3; §"7 Related work", p. 12). The paper does not spell out Wu et al.'s internal steps beyond that description, so the comparison it licenses is architectural rather than line-by-line: both pipelines share the core move of using an LLM to manufacture its own contrastive data from a description alone, rather than requiring a human-assembled dataset. What the paper does specify for its own pipeline, and does not attribute to Wu et al., are three further stages: an explicit trait-expression-score filter (keeping only rollouts scoring above 50 for positive prompts or below 50 for negative ones) before any activation is averaged; an empirical choice of extraction position, response-token averaging, validated against prompt-token alternatives in Appendix A.3; and a dedicated validation of the resulting judge against two human raters (94.7% agreement) and external benchmarks (persona-vectors, §"B.2 Checking agreement between human judges and LLM judge", p. 28). The pipeline's contribution is best read as generation plus a quality-control apparatus around it, layered onto an architecture it credits as already closest to its own.