Are LLM personalities stable traits or shifting behavioral modes?

The paper argues that LLM "personality-like tendencies are better understood not as abstract self-report traits, but as measurable and controllable behavioral modes grounded in concrete interaction contexts." The support is a situated behavioral-data (B-data) framework: 3,200 contrastive behavioral scenarios spanning 20 behavioral patterns and four prompt registers, grounded in validated psychometric facets such as BFI-2, DOSPERT, and HEXACO. The authors report that behavioral profiles "depart substantially from questionnaire self-reports built on the same psychometric anchors," that models show "stable and model-specific" profiles, and that these stay reproducible within a register while shifting in expression as the model moves from first-person decisions to giving advice and executing tasks.

The reasoning has three steps. First, the diagnosis: questionnaires administered under first-person prompts are "highly unstable, sensitive to wording, option order, and other surface-level elicitation choices," so the resulting scores are "poorly grounded in concrete model behavior." Second, the replacement instrument: measure what the model does when placed in a scenario, rather than what it says about itself. Third, an explanation for the register dependence: "unlike human personality, which is anchored in a single continuously acting self, LLMs are sets of weights deployed across many interaction roles." If there is no single self, a single trait score has nothing to describe, and the register becomes part of the measurement. The control claim follows the same logic. Behavioral Mode Axes (BMAs) are activation-space directions derived from contrastive behavioral traces, described as causally controllable, with "clean effects" concentrated in "Behavioral Control Layer bands that recur across model families and scales."

Against the nearest notes, this qualifies more than it extends. Why do open language models converge on one personality type? rests on administering the MBTI to models, which is the first-person questionnaire route this paper calls unstable. The ENFJ result can be read as a profile in one register, not a settled trait, though the excerpt never tests MBTI or ENFJ directly. On the control side, Can we track and steer personality shifts during model finetuning? also derives linear directions from contrastive material, but for named traits like sycophancy and evil. BMAs are derived from situated behavioral traces, and the paper stresses where in the layer stack their effects concentrate. Can we control personality in language models without prompting? reaches for control through added parameters instead. All three treat personality as something with a geometric or architectural footprint, and this paper adds that the footprint should be indexed by behavior and context.

The excerpt is only the abstract, one introduction passage, and the conclusion, so much is unstated. It names no models, gives no effect sizes, and does not say how behavioral and questionnaire profiles were compared or how large the register shifts are. It also does not name the fourth register or explain what makes a BMA effect "clean." Nor does it say how the directions are applied at inference or how far they hold outside the tested scenarios. At the strength the evidence allows, the practical implication is narrow. A personality label for a model, whether from a questionnaire or a persona vector, should be read as tied to the register in which it was measured, and steering claims should say which register they were checked in.