In plain words: Instead of prompting or retraining to change a chatbot's personality, this finds slices of the model's connections that already carry each persona, by seeing which parts light up on example texts. The slices matched personas more strongly than prompting or fine-tuning, with less computing.
Abstract · Your Language Model Secretly Contains Personality Subnetworks
Humans shift between different personas depending on social context. Large Language Models (LLMs) demonstrate a similar flexibility in adopting different personas and behaviors. Existing approaches, however, typically adapt such behavior through external knowledge such as prompting, retrieval-augmented generation (RAG), or fine-tuning. We ask: do LLMs really need external context or parameters to adapt to different behaviors, or do they already have such knowledge embedded in their parameters? In this work, we show that LLMs already contain persona-specialized subnetworks in their parameter space. Using small calibration datasets, we identify distinct activation signatures associated with different personas. Guided by these statistics, we develop a masking strategy that isolates lightweight persona subnetworks. Building on the findings, we further discuss: how can we discover opposing subnetwork from the model that lead to binary-opposing personas, such as introvert-extrovert? To further enhance separation in binary opposition scenarios, we introduce a contrastive pruning strategy that identifies parameters responsible for the statistical divergence between opposing personas. Our method is entirely training-free and relies solely on the language model's existing parameter space. Across diverse evaluation settings, the resulting subnetworks exhibit significantly stronger persona alignment than baselines that require external knowledge while being more efficient. Our findings suggest that diverse human-like behaviors are not merely induced in LLMs, but are already embedded in their parameter space, pointing toward a new perspective on controllable and interpretable personalization in large language models.
Ruimeng Ye, Zihan Wang, Zinan Ling, Yang Xiao, Manling Li, Xiaolong Ma, Bo Hui
arXiv:2602.07164 · cs.CL, cs.AI · submitted Feb 6, 2026
abstract · pdf · html · ICLR 2026
Psychological instruments and concepts (like MBTI) are constructed from the semantics of everyday language. Personality models (being based on self-report, and not actual behaviour) are not models of actual personality, but the correlation patterns in the language used to discuss things semantically related to "personality". It would be thus extremely surprising if LLM-output patterns (trained on people's discussions and thinking about personality) would not also result in learning similar correlational patterns (and thus similar patterns of responses when prompted with questions from personality inventories).
The real and more interesting part of the paper is the use of statistical techniques to isolate sub-networks which can then be used to emit outputs more consistent with some desired personality configuration. There is no obvious reason to me that this couldn't be extended to other types of concepts, and it kind reads to me like a way of doing a very cheap, training-free sort of "fine-tuning".