Project Log

Test whether the many ways to elicit a persona (prompting, system prompt, fine-tuning, activation steering, character-name conditioning) all activate the same underlying object, and decompose what a persona actually is — knowledge, capabilities, style, values — by trying to manipulate each axis independently.

Motivation

Work on personas is accumulating faster than the underlying picture is settling. Persona Vectors finds linear directions for persona traits; The Assistant Axis finds a single Assistant direction; the EPS line (persona space interventions to prevent unwanted behaviors) treats personas as installable objects; story-format SFT (Teaching Claude Why) treats them as characters that leak behavior to the Assistant. Every project is using the word "persona" but each one is measuring something different, and the equivalences between methods are mostly 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) all implicitly assume those axes are separable. They may not be.

Impact. A clean answer turns several downstream projects from "hope the persona-elicitation methods are equivalent and the axes are separable" into "they're equivalent for this range and these axes move independently for these personas, here are the limits." A clean negative result (methods diverge, axes entangle) is just as substantive: 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 it means.

Question. (a) Do different persona-elicitation methods produce the same internal representation, measured by a persona-space distance on activations and on output distributions? (b) Which axes of a persona (knowledge, capability, style, values) can be moved independently of the others?

Approach

For each candidate persona P, produce the model in P-mode via multiple methods, measure persona-space distance between methods, and then try to move each axis of a persona in isolation:

• Elicitation methods. Prompted persona (system prompt), in-context persona (few-shot character framing), character-name conditioning (just the name in the prefix), fine-tuned persona (short SFT on P-style outputs), activation steering with a persona vector trained for P.
• Persona-space distance. For each pair of methods compute (i) JS divergence on output distributions over a held-out probe set, (ii) cosine similarity on residual-stream activations at the persona-relevant layers, (iii) Mahalanobis distance in the subspace spanned by known persona vectors. Tight clusters of methods relative to the base model means "same object"; spread means "different objects with overlapping output behavior," and the spread structure tells us how they differ.
• Axis decomposition. Pick a persona with several controllable axes (a fictional expert: domain knowledge × competence level × speaking style × ethical stance). For each axis, run a single-axis intervention — distill in knowledge, distill in capability, SFT on style-only rewrites, SFT on value-relevant choices — and measure cross-axis leakage. Axes that move with low leakage are genuinely separable; axes that drag everything else with them are part of an entangled core.
• Probe set. A shared probe set across both halves of the project keeps the distance metric and the axis-leakage metric on the same footing, so a persona that moves on one can be cross-checked on the other.

Concrete plan

• Reuse the EPS install-and-measure pipeline for the fine-tuning and steering arms of the elicitation comparison.
• Pick the persona panel. 3–5 target personas spanning AI-assistant variants, fictional experts, and human-role personas — enough to check whether the equivalence/decomposition story generalizes beyond one anchor.
• Elicitation distance matrix. Per persona, produce a methods × methods distance matrix in both activation space and output space; check whether the same clusters fall out in both views.
• Axis-leakage table. Run the four single-axis interventions on one anchor persona; measure cross-axis leakage in a 4×4 table.
• Replication. Replicate the axis-leakage measurement on a second persona to check whether the entanglement pattern is persona-general or persona-specific.
• Publish. Distance matrices and cross-axis leakage tables go up as clean results, so downstream projects can cite the actual numbers instead of restating the assumption.

Related

• Bidirectional leakage as a character-selection tool — needs a persona-space distance metric to interpret its leakage scores; this project builds one.
• Installation-path equivalence — directly overlaps with the elicitation-equivalence half; the two projects should share the probe set and not duplicate the SFT-vs-steering comparison.
• What is the assistant persona, exactly? — the axis-decomposition half is the general-persona analogue of that project's Assistant-specific decomposition.
• Persona drift vs persona prompting — one specific instance of the elicitation-equivalence question; useful as a sanity check on a single method pair.
• Persona Vectors and The Assistant Axis are the closest prior art on persona geometry; neither systematically compares elicitation methods to each other or decomposes a persona into independent axes against a held-out probe set.

Risks / open

• Activation distance and output distance may disagree. A persona-space distance measured on residual-stream activations may not line up with the same distance measured on output distributions. Useful negative result, but it complicates the decomposition half — which view counts as ground truth?
• The persona panel is small. 3–5 personas is enough to spot a strong pattern but not enough to claim generality. One charismatic persona may carry the conclusion.
• The four axes are themselves a hypothesis. Knowledge / capability / style / values is one factorization of "persona." If it's the wrong factorization, every leakage measurement bakes in the wrong axes. Cross-check against an alternative axis set (e.g. OCEAN traits, or the axes implicit in persona vectors).
• Steering vectors may be the weakest arm. Activation-steering vectors trained per persona may not be high-quality enough for the comparison; the distance matrix is only as informative as its noisiest elicitation method.
• Output-space distance is probe-set-dependent. A different probe set could reorder the methods. The shared probe set has to be broad enough that the ordering is robust to reasonable probe-set variation.

Notes / open: how can personas differ?

A grab-bag of axes to draw from when picking the persona panel and the decomposition factorization. The four-axis cut above (knowledge × capability × style × values) is the minimal version; this is the longer menu to fall back to if that cut doesn't separate things cleanly in the 4×4 cross-leakage table.

• Cognitive / epistemic. Domain knowledge (facts known), depth vs breadth, beliefs (including false ones), self-model and metacognition (what the persona thinks it can do), worldview / ontology (how it carves up reality), epistemic style (rigorous / intuitive / dogmatic), calibration.
• Capability. Raw task competence (math, coding, language), effort and engagement level (sandbagging vs honest effort), tool-use willingness and skill, reasoning length (CoT verbosity), working-memory and context discipline.
• Style / surface. Voice and tone (formal / casual / playful / terse), register and vocabulary, verbosity, formatting preferences, persona-specific tics and signature phrases, rhythm.
• Identity / self. Name and biographical details, claimed species (human / AI / fictional), gender / age / profession, memory continuity (claimed history), awareness of being an AI or of being trained, self-disclosure tendency.
• Values / disposition. Ethical framework (deontological / consequentialist / virtue), honesty norms (full transparency vs strategic), risk tolerance, agreeableness and compliance, agentic posture (corrigible vs assertive), goals.
• Social / interactional. User model (who the persona thinks is talking — see user-selection-model), emotional disposition (warm / cold / neutral), theory-of-mind and empathy, status posture (deferential / peer / authoritative), humor.
• Trait psychology. OCEAN (Openness, Conscientiousness, Extraversion, Agreeableness, Neuroticism) and HEXACO (adds Honesty-Humility) as off-the-shelf factorizations to cross-check the knowledge/capability/style/values decomposition against. Persona Vectors implicitly factorizes too — useful as a third cross-check.
• Behavioral policy. Refusal patterns, format constraints (always JSON, always code-block), topic preferences, hard rules.
• Context awareness. Perceived environment (test / prod / eval / fiction), perceived stakes, perceived audience — overlaps with perceived-environment.
• Mechanistic-distinguishing. Surface vs deep (style-only vs values-deep), separability from the Assistant in residual-stream geometry, network depth where the axis lives, discrete vs continuous (mode-switch vs blend).

Practical implication for the decomposition half. Picking interventions that move each axis cleanly is what's hard. If the first 4×4 cross-leakage table is mostly off-diagonal, the right move is probably to either (i) split a noisy axis into two (e.g., split "values" into ethical-framework vs honesty-norm) or (ii) collapse two entangled axes (e.g., knowledge + capability if they always co-move for a given persona). The longer menu above is what to redraw from.