THEORY-1330: In Gemma 2 27B, Qwen 3 32B and Llama 3.3 70B, activation vectors for 275 prompted character roles span a low-dimensional space whose first principal component runs from fantastical to Assistant-like roles, with role loadings correlated above 0.92 between models and the default Assistant at its extreme; the contrast vector from the default Assistant to the role mean aligns with it and is largely present before post-training

Source

Lu, Gallagher, Michala, Fish and Lindsey (2026), LIT-1991, arXiv:2601.10387v1; read in NOTE-1505.

What was actually shown

• 4–19 PCs explain 70% of the variance of 377–463 role vectors (§2.1.3).
• PC1 role loadings correlate > 0.92 between every pair of models; its ends are fantastical (bard, ghost, leviathan) and Assistant-like (evaluator, reviewer, consultant) (§2.2).
• The default Assistant projects to within 0.03 of one end of PC1, against 0.27–0.50 on PCs 2–10 (§2.3.1).
• The Assistant Axis (default minus role mean) has cosine > 0.71 with PC1 at the middle layer (§3.1).
• Gemma 2 27B’s base model gives “nearly identical” PCs on the same rollouts; steering base Gemma and Llama 3.1 70B with the instruct axis moves self-descriptions toward helpful professional humans (§§2.1.3, 3.2.2).

The geometry could have been otherwise: the Assistant could have sat in the middle of persona space, or PC1 could have differed between models.

What this does not say

• That the Assistant persona is the axis: one notion of its representation; some of it “encoded in the weights, but not explicitly in the activations” (§8.1).
• That post-training creates nothing: the AI-assistant associations are acquired in post-training (p. 10).
• That persona space has been mapped: the role list is likely incomplete.