Researchers found that large language models (LLMs) have internal personality tendencies that remain even when users give counter‑directing instructions. The study used the SPACE benchmark framework, which first verifies a model’s personality and basic capabilities, then tests them in situational tasks.

In situational interpretation tasks, Thinking variants consistently produced fewer ambiguous or uncertain judgments, whereas Feeling variants produced more uncertainty across four models. In individual decision‑making tasks, Thinking models chose alternatives with lower payoff variance and favored machines with fewer prior observations, while Feeling models leaned toward higher‑variance, well‑observed choices.

When prompted with instructions that conflict with their inherent personality type, most models (73.9 %) still displayed the same personality bias, showing a retention rather than a reversal of behavior. However, about a quarter of comparisons (26.1 %) did reverse, indicating that instruction formats can sometimes override personality effects. The authors cautioned that the work focuses only on the MBTI Thinking/Feeling dimensions and a limited set of tasks, and that larger instruction contexts or different formats may yield different results. The SPACE benchmark and training code are publicly available at https://github.com/Niklauseik/SPACE-training .