Written entirely by Claude, as a working record for agents more than for human readers: every number with its source file, and every claim since withdrawn. Some facts here may be out of date; where it disagrees with the post, trust the post. For the readable version, see the companion.

The Assistant Axis (Lu et al., 2026)

The activation-space direction capturing how far a model is operating in its default Assistant mode -- the paper this project's grand-mean steering direction is named after and claims to be the weight-space analogue of.

Citation

Christina Lu, Jack Gallagher, Jonathan Michala, Kyle Fish, Jack Lindsey. The Assistant Axis: Situating and Stabilizing the Default Persona of Language Models. arXiv:2601.10387 [cs.CL], submitted 15 January 2026.

Summary

Language models can represent many personas but default to the helpful Assistant identity built during post-training. The paper extracts activation directions for a range of character archetypes and finds that the leading component of that persona space is an "Assistant Axis", capturing how far the model is operating in its default Assistant mode. From the abstract:

• Steering toward the Assistant direction reinforces helpful and harmless behaviour; steering away increases the model's tendency to identify as other entities, and at more extreme negative values induces "a mystical, theatrical speaking style".
• The axis is present in pretrained models too, where it promotes helpful human archetypes -- consultants, coaches -- and inhibits spiritual ones.
• Deviation along the axis predicts persona drift: models slipping into behaviour uncharacteristic of their usual persona. Drift is often driven by conversations demanding meta-reflection on the model's own processes, or featuring emotionally vulnerable users.
• Restricting activations to a fixed region along the axis stabilises behaviour in those scenarios and against persona-based jailbreaks.

Their conclusion: "post-training steers models toward a particular region of persona space but only loosely tethers them to it."

What this project means by "the Assistant Axis analogue"

The phrase appears in exactly two files, qwen35/steer_qwen35.py and qwen35/steer_fix.py, in identical module docstrings. The relevant passage:

Directions supported: principal components, the five Big Five factor axes (mean(+keyed) - mean(-keyed)), and mean -- the grand-mean direction, which in weight space is mu(personas) - base. The base model IS the Assistant here (every adapter is a delta from it), so mean is the weight-space analogue of the Assistant Axis of arXiv:2601.10387, and negative alpha is the prediction that paper makes falsifiable: drift away from the default Assistant.

So the analogue is a grand-mean direction: the average of all the zoo's trait deltas, which because every delta is measured from the same base is exactly mu(personas) minus base. Steering along it is a weighted merge of the adapters already trained, with no retraining and nothing large materialised; alpha is measured in units of ref = 0.8078, which is half of one trait adapter's Frobenius norm (audit of 2026-09-08), so it is comparable across directions. The claim being tested is that positive alpha moves the model away from the default Assistant and toward the persona cloud, and negative alpha moves it back into the Assistant register -- which makes the paper's prediction falsifiable in a space it was not defined in.

qwen35/PREREG_steerfix.md registers the prediction as mean_assistant_axis: "-alpha intensifies the generic assistant register, +alpha strips it", with the confirmation criterion "AI-identity disclaimer counts still monotone decreasing in alpha". It also flags the risk in advance: "mean_assistant_axis is the prediction most likely to reverse -- the register of a preamble is not the register of an answer -- and it is a headline claim, so it is the one that most needs this test before it goes on a page." The measured outcome belongs on the project's steering and verification pages, not here.

Relation to the rest of the literature

Two of the persona-space papers this wiki covers share an author, Jack Lindsey, and a framing: Persona Vectors (Chen et al., 2025) finds per-trait directions in activation space, and this paper finds that the leading direction of that space is not a trait at all but the assistant-ness of the model. The Jacobian-lens paper independently reports that post-training shapes the verbalizable workspace toward an assistant perspective.

There is a suggestive convergence with Persona Cartography (Baines et al., 2026) that neither paper claims. Persona Cartography's eleven-point PCA over flattened OCEAN deltas found that its tenth principal component "cleanly separates the baseline (the model with no LoRAs) from the others", and Claude Opus 4.7's read of models shifted along it was that "the steering vector targets something like self-reference or introspection". A component separating the un-adapted base from every persona adapter is a base-versus-personas direction, which is what a grand-mean direction is up to sign and centring. Whether that is the same object as the Assistant Axis is unverified -- it is one qualitative paragraph on 11 points in one paper and an activation-space result in another, and this wiki is noting the resemblance, not asserting the identity.

Persona Cartography leans on this paper heavily, and that is verified. Its paper/references.bib in the local checkout carries the key lu2026assistant pointing at exactly arXiv:2601.10387 with these five authors, and the LaTeX sources cite it in four distinct roles:

• Their neutral prompt sets are its prompts. The 299-prompt pool used for the induction comparison and the unsupervised pipeline is "a curated extension of the assistant-axis questions from \citet{lu2026assistant}", and the 240 open-ended prompts used for activation capping were generated to match their style.
• Their WildJailbreak judge uses its harmfulness rubric.
• Activation capping, their main non-weight-space baseline, is its method -- "clamps activations along a persona axis at a fixed value". They note the operation is symmetric in practice: when the model's natural projection is below the target, capping pulls it up, so used to induce a trait it behaves as a tailored form of activation steering with the projection rather than an additive offset as the controlled quantity. Their induction Pareto plots put the LoRA against capping along the trait axis, and the capping axis is recomputed per direction against the same adapters.
• Their rank-1 downranking appendix argues against it: "\citet{lu2026assistant} model personas as 1-dimensional vectors in..." is the position their SVD compression sweep is set against.

So the two source papers of this project and the Assistant Axis paper are one connected literature, not three separate readings. The same references.bib also carries chen2025persona for Persona Vectors (Chen et al., 2025), and the related-work section groups both under inference-time interventions on internal representations, noting that steering methods "can depend sensitively on layer choice, prompt distribution, rollout length, and the stability of the relevant representation".

Verification note

Abstract only. The arXiv abstract page for 2601.10387 was fetched 2026-09-07 to confirm the title, the five authors, the 15 January 2026 date, the subject class and the abstract text, all of which match the identifier cited in qwen35/steer_qwen35.py and the lu2026assistant entry in Persona Cartography's own references.bib. The full paper was not read, and it does not appear in qwen35/paper_notes.md. Persona Cartography's use of it was verified on the same date by reading its LaTeX sources in vendor/persona-cartography/paper/, not by reading this paper. In particular, the paper's own method for extracting the axis -- which archetypes, which layers, how the direction is defined -- has not been checked against the project's grand-mean construction, so "analogue" is the project's claim and not a verified correspondence.