Sparse autoencoders (SAEs) — offers an unsupervised alternative to → Steering vector

Sparse autoencoders and steering vectors are two different answers to the same question — how do you find a direction in activation space that means something — and Persona Vectors is the paper in this corpus that states the contrast directly. A steering vector, whether CAA’s Mean Difference over hundreds of contrast pairs or the persona-vector-extraction-pipeline’s automated version, requires picking a target concept in advance and constructing labeled positive/negative examples of it before any direction can be computed. A sparse autoencoder needs neither: it is trained on unlabelled activations alone, discovering a dictionary of thousands of directions with no concept specified up front, which is exactly why Persona Vectors’ own related-work section describes SAEs as "another approach to obtaining interpretable directions... which find directions in an unsupervised way" (persona-vectors, §"7 Related work", p. 12), and its limitations section frames that unsupervised property as a genuine advantage over its own supervised pipeline: SAEs "may therefore enable unsupervised discovery of persona-relevant directions, including specific traits that cannot be easily elicited through prompting" (persona-vectors, §"8 Limitations", p. 13). From the sparse-autoencoder paper’s own 2023 standpoint this framing did not exist yet — its dictionaries were built purely to decompose activations at scale, with no notion of a "steering vector" to be an alternative to — so the contrast is legible only in hindsight, once a later paper had a supervised, labeled-direction method for the unsupervised method to be compared against.