Automated concept-description-to-direction pipeline (Wu et al., 2025)
Persona Vectors: Monitoring and Controlling Character Traits in Language Models — inherited
Prior work developing an automated pipeline that uses an LLM to generate synthetic contrastive pairs of generations from an arbitrary natural-language concept description, and extracts linear directions from them. Cited as the most closely related prior work to this paper's own extraction pipeline.
Before this paper built its own recipe for turning a trait description into a steering direction, another line of work had already tried something similar for concepts generally, not personality traits specifically — automating away the hand-built datasets such methods used to require.
What the pipeline does opens the account, along with why the paper singles it out as the closest thing in prior work to its own extraction method. Left unattributed to this earlier pipeline are three things the paper's own version adds: a quality filter on generated responses, an empirically chosen point in the response to read activations from, and a validation of the resulting judge against actual human raters.
What the pipeline does
Wu et al. (2025) built an automated pipeline for turning an arbitrary natural-language description of a concept into a linear direction in a language model's activation space, without any hand-curated dataset. An LLM is used to generate synthetic contrastive pairs of text directly from the description — generations that express the target concept and generations that do not — and a linear direction is then extracted from the resulting activations. Persona vectors' own introduction to its extraction pipeline credits this as "an automated pipeline for translating natural language concept descriptions into contrastive pairs of generations, and eventually into linear directions" (persona-vectors, §"2 An automated pipeline to extract persona vectors", p. 3).
The paper's own closest relative
Persona vectors names Wu et al. (2025) directly as its nearest prior work, in a footnote introducing its own extraction method: "Most similarly, Wu et al. (2025) also developed an automated pipeline for translating natural language concept descriptions into contrastive pairs of generations, and eventually into linear directions" (persona-vectors, §"2 An automated pipeline to extract persona vectors", p. 3). Its Related Work section repeats the comparison and draws the contrast that matters most: unlike prior contrastive-pair methods that "generally required bespoke data curation to obtain contrastive pairs" (persona-vectors, §"7 Related work", p. 12), citing Turner et al. (2024), Panickssery et al. (2024), and Zou et al. (2025), both Wu et al.'s pipeline and the paper's own Automated persona vector extraction pipeline generate that contrastive data automatically from a description alone, with no dataset assembled by hand at either end.
What the paper adds beyond it
The paper does not describe Wu et al.'s internal steps beyond this framing, so the comparison it licenses is architectural rather than step-by-step. What it does specify, and does not attribute to Wu et al., are three further stages built into its own pipeline: a Trait expression score (LLM judge) filter applied before any activation is averaged, keeping only "responses with trait scores greater than 50 for positive prompts and less than 50 for negative prompts" (persona-vectors, §"2.2 Extracting persona vectors", p. 4); an empirically chosen extraction position — averaging over response tokens rather than prompt tokens, a choice validated by a dedicated comparison (persona-vectors, §"A.3 Extracting persona vectors at different token positions", p. 27); and a validation of the resulting judge against two human raters, reaching 94.7% agreement (persona-vectors, §"B.2 Checking agreement between human judges and LLM judge", p. 28). The contribution reads as generation plus a quality-control apparatus wrapped around an architecture the paper credits as already closest to its own.