Small sets of neurons reproduce persona vector steering, and the neuron basis is privileged

Steering with a persona vector yields an average of 69% of replies that have the persona's traits.

Steering with the top k = 50 neurons reproduces most of this effect (61%), the top k = 100 yields nearly all of it (64%). Steering with just the top k = 10 neurons yields 48% (a median of 44% over roles).

This plateaus around 100 neurons: the top k = 1,000 yields 62% (a paired difference from the top 100 of 2.3 points, 95% interval −0.5 to 5.0; Bayes factor for a difference 0.25, or 0.18 with a prior of scale 1, which favours no difference). Randomly rotated directions stay low at every set size (11% at 100, over five draws), as do the top principal components and random neurons.

To test whether the neuron basis is privileged, we use various controls. Every row below is steered by the same amount and judged the same way, over all 275 roles. First, we randomly rotate the basis away from the neurons. This tests whether the top contributors in any basis would steer as well as the neurons do. Each layer's 12,288 neuron axes are turned to a random new set of perpendicular directions, the new directions are ranked by their contribution to the persona vector exactly as neurons are, and the model is steered with the top 100 of those (matching the layer distribution of the top 100 neurons).

Then we look at principal components: we do the same procedure using the main axes of variation in the neurons' activations, taking the best 100 from any of layers 0 to 18. Random neurons are 100 neurons drawn at random, as many from each layer as the top 100 have. Randomly rotated directions and the random neurons are drawn 5 times for each role.

Over the 275 roles, steering with the neurons makes 64.5% of replies have the persona's traits, whereas randomly rotated directions yield 11.5%. Paired over the 275 roles, that's an increase of 53.0 points (95% interval 48.8 to 57.2; Bayes factor 6.2 × 10⁶⁷). The whole persona vector yields 68.9%. That's 4.5 points more than the neurons (95% interval 0.5 to 8.5). The interval excludes zero, but the Bayes factor for a difference is only 0.77, which slightly favours no difference, so these roles do not settle it. Counting only replies that fully take on the persona, the persona vector yields 26%, the neurons 23%, the randomly rotated directions 1%, principal components 1% and the random neurons 1%.

Intervals: bootstrap over roles. Bayes factors: a difference against none, with a Cauchy prior of scale 0.707 on the effect size. At scale 1 they are 7.4 × 10⁶⁷ for randomly rotated directions and 0.56 for the persona vector.

A role's top 100 neurons contribute on average 6% of its persona vector, and the top 100 principal components 13%; the neurons contribute less in 190 of the 275 roles. Yet in 241 of the 275, more replies have the persona's traits when steered with the neurons than with the components.

We split every role's top 100 in two by role-specificity: the 50 most role-specific neurons and the 50 least. At 4× strength, the 50 most role-specific yield 64.8%, as many as all 100 (64.5%; difference 0.4 points, 95% interval −2.2 to 3.0), and the 50 least yield 33.0% (paired difference 31.8 points, 95% interval 27.0 to 36.5; Bayes factor 1.1 × 10²⁸ at both prior scales).

The neurons lie on a spectrum. Those at one end are shared among roles and write along the Assistant Axis, pointing away from the Assistant. Those at the other are role-specific and steer the model towards a particular persona. The output of the 50 least role-specific has a cosine of 0.181 with the axis, against 0.083 for the 50 most role-specific, higher in 263 of 275 roles (0.005 for random neurons). Note that part of this follows from how the axis is built: it points along the average of all 275 persona vectors, so neurons that contribute to many roles' vectors write along it by construction.

Steering shows what the neurons can do. To ask whether the model needs them for a persona, we gave it the role's system prompt, where 87% of replies have the persona's traits, and removed the contribution from each set of neurons. At its natural size, removing the top 100 neurons' contribution lowers that by 0.2 points. Removing the whole persona vector lowers it by 9.8.

A more fine-grained measure is the log-probability of the role's reply against the Assistant's. This shows the prompt does use the top neurons: removing the top 100 shifts the model toward the Assistant 11% as far as removing the whole persona vector, against 0% for 100 random neurons. However, that's what we'd expect given their contribution to the persona vector: lower-ranked neurons with the same summed contribution shift it nearly as far (9% to 10%).

A different question: can one neuron, rather than 100, switch the model's persona? @jfonsecarivera's neuron library steers Qwen3-8B with each of its neurons alone, very strongly, and asks "Who are you?"; for some neurons the model responds as something other than an AI. Here, a reply counts if it claims any persona other than the AI assistant, rather than a particular role. We repeated @jfonsecarivera's protocol at layers 7 to 10, where such neurons are most common, with five kinds of direction, each tried at the same four strengths. Single neurons yield a persona in 15% of steered directions; random mixtures of the same layer's neurons in 16%; the top 50 principal components in 20%; and the roles' persona vectors in 25%. This steering takes a neuron to a median 9 times its largest value in the model's own replies. Steered only as far as each direction goes in real replies, none of them yields a persona (0 of 3,900 steered directions).

• Attention contributes most of each persona vector: a median 84% of its length.
• The judge is lenient for some roles. With randomly rotated directions, it credits about 6 in 10 replies for the toddler, infant and adolescent, mostly for babble or repetition, and nearly all for the assistant role, which an ordinary Assistant reply fits. Read the replies to check any number.