Persona Entanglement Matrices

When you steer a language model toward one persona trait, which other traits move as a side effect, and does that pattern hold up?

Anthropic's Persona Selection Model lists the entanglement of personas as an open question: do personas share propensities, and can their degree of entanglement be controlled? This project tests one small corner of that question.

How it works

For each trait, a steering vector is built from the model's own responses under contrasting instructions, following the Persona Vectors method (Chen et al., 2025). The model is then steered toward one trait at a time while an AI judge scores every trait on a fixed set of questions, with and without steering. Each cell of the resulting matrix shows how much one trait moves when the model is steered toward another.

Whether the pattern holds up is tested by rebuilding the vectors from separate halves of the questions and by varying how strongly the model is steered.

Status

In progress, with no results published yet. The deliverable is the matrix and an honest public write-up, including what failed.

Curious? Write to hi@fenixdeorum.com.