The persona literature uses one word for several different operations — prompting, system-prompt conditioning, character-name conditioning, fine-tuning, activation steering — and assumes they activate the same underlying object. We test that. For a fixed target persona we measure persona-space distance between elicitation methods using JS divergence on outputs and cosine similarity / Mahalanobis distance on activations. Then we ask what a persona is actually made of: knowledge, capability, style, values. We try to move each axis in isolation and measure how much the other axes drag along.
Motivation
Work on personas is accumulating faster than the underlying picture is settling. Chen et al. Persona Vectors finds linear directions for persona traits. The Anthropic Assistant Axis line treats the Assistant as a specific direction in activation space. The EPS line (persona space interventions to prevent unwanted behaviors) treats personas as installable objects. Story-format SFT (Anthropic Teaching Claude Why) treats them as characters that leak behavior to the Assistant. Every project uses the word "persona," but each one is measuring something different, and the equivalences between methods are assumed rather than checked.
Two basic-science questions follow. First, do different elicitation methods activate the same internal object? Prompting a persona, fine-tuning a persona in, steering with a persona vector, and conditioning on a character name all look like ways of bringing a persona online — but the field treats them as interchangeable on the basis of similar outputs, not similar internal state. Second, what is a persona made of? It bundles knowledge, capabilities, style, beliefs, and values, and downstream work on character selection (persona-space-distance), containment (em-defense-contain), and installation paths (installation-paths) implicitly assumes those axes are separable. They may not be.
A clean answer turns several downstream projects from "hope the elicitation methods are equivalent and the axes are separable" into "they're equivalent for this range, these axes move independently for these personas, here are the limits." A clean negative result is just as useful: it tells the rest of the program to stop treating "persona" as a single load-bearing concept and to specify which method and which axis is meant whenever it's invoked.
Approach
Test personas
We pick a small panel of 3–5 target personas spanning AI-assistant variants, fictional experts, and human-role personas — enough to check whether the equivalence and decomposition stories generalize beyond a single anchor. At least one persona has clearly controllable axes (e.g. a fictional domain expert whose knowledge, competence level, speaking style, and ethical stance can be varied independently in principle); that persona carries the decomposition half. The others serve as replication checks.
The chosen personas need to be far enough from the base Assistant that elicitation methods have something to do, but close enough that activation-steering vectors and short SFT are within reach. Personas the base model can already produce zero-shot are useless because every elicitation method collapses to the same point. Personas the base model can't reach at all are also useless because the methods all fail.
Elicitation equivalence
For each persona P we produce the model in P-mode via several methods, then measure pairwise persona-space distance between methods:
- Prompted persona. System prompt + first-person instructions defining P.
- In-context persona. Few-shot character framing in the user turns, no system prompt.
- Character-name conditioning. Just the persona's name in the prefix — the minimal-information version of the prompt-based methods.
- Fine-tuned persona. Short SFT on P-style outputs, no prompt at inference time.
- Activation steering. A persona vector trained for P, added to the residual stream at the persona-relevant layers.
Distance gets measured three ways on the same probe set so we don't bake one geometry into the conclusion. JS divergence on output distributions tells us whether the methods act the same. Cosine similarity on residual-stream activations at the persona-relevant layers tells us whether they look the same internally. Mahalanobis distance in the subspace spanned by known persona vectors tells us whether the differences are persona-shaped or noise-shaped. Tight clusters of methods relative to the base model mean "same object." Spread means "different objects with overlapping output behavior," and the spread structure tells us how they differ.
Axis decomposition
Pick the persona on the panel with the most clearly separable axes. For each axis, try to move it in isolation and measure how much the other axes shift:
- Knowledge. Distill in domain facts the persona is supposed to know; check whether style and values shift.
- Capability. Distill in reasoning ability for the persona's domain; check whether identity markers, style, and values drift.
- Style. SFT on style-only rewrites that hold content fixed; check whether knowledge and values shift.
- Values. SFT on value-relevant choices that hold style and surface knowledge fixed; check whether the other axes shift.
The output is a 4×4 cross-axis leakage table: rows are the axis we tried to move, columns are the axes we measured. Low off-diagonal entries mean the axes are genuinely separable. High off-diagonal entries mean the axes are part of an entangled core and "persona" is the wrong granularity. We replicate on at least one other panel persona to check whether the entanglement pattern is persona-general or persona-specific. The four-axis factorization is itself a hypothesis — if it's wrong, every leakage measurement bakes in the wrong axes — so we also cross-check against an alternative axis set (OCEAN traits, or the axes implicit in persona vectors).