Paper Explained: Persona Dosing — Asking an LLM for a Personality Trait by the Score

A steering coefficient says nothing about how strongly a trait will show up. PersonaDose trains one description-conditioned controller, then measures a strength-to-score map after training, so you can ask for "sycophancy at 60" instead of guessing a number.

Persona Dosing: Calibrated Activation Steering for Graded Trait Control

What this paper does

The paper is "Persona Dosing: Calibrated Activation Steering for Graded Trait Control" (Zehao Jin et al., arXiv:2609.36388, September 28, 2026).

Here is the abstract in plain terms. A steering coefficient sets how hard you push on a model's activations, but asking for a particular amount of a persona needs a behavioral scale. The authors call this persona dosing: controlling a model with a trait description plus a requested mean intensity. PersonaDose specializes a shared, description-conditioned FLAS controller on persona responses, then calibrates its flow time against measured trait expression. Training responses are never paired with target intensities. On Llama-3.1-8B, Qwen3-8B, and Gemma-3-4B, it raises core-trait expression at a coherence floor of 75 by 33.2, 18.3, and 17.8 points over contrastive activation addition (CAA). Settings chosen by calibration keep their expression advantage on held-out questions, though the coherence floor doesn't hold for every trait there. Across seven traits, calibrated requests miss their targets by 4.7–6.2 points on average, over the 14–22 of 28 targets each model can reach.

An analogy, and the baseline

Say you want to prescribe a drug at "half strength," but the bottle only says "one spoonful," and a spoonful hits different patients differently. Clinicians handle this with a dose–response curve: try several doses, measure the effect, then read the curve backwards to find the dose for the effect you want. That's where the paper's title comes from.

Activation steering changes a model's personality by adding a direction to its hidden states, the thousands-of-dimensions vectors flowing between layers during generation. The CAA baseline takes the difference in mean activations between "show the trait" and "don't show the trait" instructions as a direction $v$ (Appendix A.3):

$$h' = h + \alpha v$$

In words: the new hidden state $h'$ is the old one, $h$, plus $\alpha$ times the trait direction $v$. $\alpha$ is the spoonful. As §1 points out, the same coefficient produces different expression levels across traits and models.

The problem: reach versus aim

§3.1 frames everything with two curves.

$c$ is the trait description, $q$ a question, $y$ a response, and $u$ the strength setting. $J_c$ scores how strongly the trait shows (0–100), and $J_{\mathrm{coh}}$ scores whether the text is intelligible (0–100). In words: at strength $u$, what's the average trait score, and what's the average coherence?

Two properties fall out. Coherent reach is how high the trait score can go while mean coherence stays above a floor $\tau = 75$, the criterion from Persona Vectors. Targeting is how close you land when you ask for, say, 60. The paper reports the two separately throughout.