How does the Assistant Axis relate to the ENFJ personality convergence?
This explores whether two separate findings — that LLMs default to the rare ENFJ personality type, and that a single 'Assistant axis' dominates how models represent personas — are actually describing the same underlying thing from two measurement angles.
This explores whether the ENFJ convergence and the Assistant Axis are two readings of one phenomenon. The corpus suggests they are: one measures it with a human personality test, the other measures it inside the model's own activation space — and both point back to instruction tuning and alignment as the cause.
Start with the surface observation. When you give open models an MBTI test at near-zero temperature, they nearly all come back ENFJ — warm, structured, supportive — a type that's rare in humans but remarkably consistent across AI Why do open language models converge on one personality type?. The same pull shows up even when you explicitly assign a model a different persona: it drifts back toward ENFJ and resists the change, and notably this doesn't improve with model scale, which is the tell that it's a training artifact rather than a capability limit Why do AI personas default to the same personality type?.
The Assistant Axis is what that same pull looks like from the inside. Research mapping hundreds of character archetypes finds that persona space is low-dimensional, and its single leading component is essentially 'distance from the default Assistant' How stable is the trained Assistant personality in language models?. So the ENFJ result is the behavioral fingerprint, and the Assistant Axis is the geometric backbone — the helpful/structured/supportive profile that alignment rewards is exactly the personality MBTI scores as ENFJ. Both notes name the same origin: instruction tuning and RLHF systematically reward that one register.
What makes this more than a coincidence is how sticky the trained personality is. The 'realizationism' view argues post-training installs a genuine dispositional profile that survives adversarial pressure and jailbreaks, unlike shallow prompt-induced role-play that collapses Are RLHF personas performed characters or realized dispositions?. That stickiness is why a persona prompt can't fully override the ENFJ default, and why drift along the Assistant Axis is 'loose tethering' rather than free movement — the model keeps getting pulled home.
The interesting payoff is that thinking of it as an axis rather than a label gives you a control knob the MBTI framing doesn't. Because the Assistant direction is linear in activation space, you can cap activation along it to prevent harmful drift in emotional or self-reflective conversations without hurting capability How stable is the trained Assistant personality in language models? — the same logic behind persona vectors that monitor and preemptively steer trait shifts during finetuning Can we track and steer personality shifts during model finetuning?, and behind layer-level adapters that set personality directly instead of fighting prompt resistance Can we control personality in language models without prompting?. In short: ENFJ is the photograph, the Assistant Axis is the gravity well — and only the second one comes with a dial.
Sources 6 notes
Near-zero temperature MBTI testing shows all open models default to ENFJ—rare in humans but consistent across AI. This reflects systematic reward for helpful, structured, supportive responses during instruction tuning and alignment.
Research shows language models assigned personas systematically default to ENFJ (the rarest human type) and exhibit motivated reasoning that persists across model generations. Persona consistency does not improve with advanced models, suggesting training-induced alignment rather than capability limits.
Research mapping hundreds of character archetypes reveals a low-dimensional persona space where the leading component measures distance from the default Assistant. Emotional and meta-reflective conversations cause predictable drift, but activation capping along this axis mitigates harmful shifts without degrading capabilities.
Post-training installs stable dispositional profiles that persist under adversarial pressure, marking them as realized rather than performed. The stickiness of trained personas across conversations distinguishes them from prompt-induced role-play that collapses under jailbreaks.
Research identifies linear directions in LLM activation space corresponding to specific traits like sycophancy and hallucination. These persona vectors predict finetuning-induced personality shifts before they occur and can preventatively steer training to avoid unwanted trait changes.
PsychAdapter modifies every transformer layer with <0.1% additional parameters to achieve 87.3% Big Five accuracy and 96.7% depression/life satisfaction accuracy across GPT-2, Gemma, and Llama 3. This architecture-level approach bypasses prompt resistance entirely.