ProtoMBTI: Prototype-Guided Retrieval-Augmented Reasoning for MBTI Inference from Text

ACL ARR 2026 January Submission1401 Authors

Keywords: Prototype-based Reasoning, Personality Inference, Large Language Models, Cognitive Modeling, Interpretability

Abstract: Understanding a user's personality is crucial for personalized AI, and the MBTI provides a widely recognized operational framework for personality modeling. Existing text-based MBTI prediction methods often treat labels as fixed categories, neglecting the prototype-based nature of personality emphasized in cognitive psychology. To address this, we propose \textbf{ProtoMBTI}, a prototype-based reasoning framework for social-media text: it aligns LLM inference with the cognitive structure of personality via prototype retrieval-driven reasoning. Specifically, \textsc{ProtoMBTI} constructs a balanced, high-quality prototype library and performs a retrieve–reuse–revise–retain cycle during inference to achieve accurate, interpretable, and transferable predictions. On the Kaggle (85.14\%) and Pandora (71.41\%) benchmarks, \textsc{ProtoMBTI} significantly outperforms neural and LLM baselines, and under distribution shift achieves an average accuracy of 96.41\% on the Pandora test set, covering all 16 MBTI types.

Paper Type: Long

Research Area: Linguistic theories, Cognitive Modeling and Psycholinguistics

Research Area Keywords: cognitive modeling,computational psycholinguistics,Natural Language Processing,computational social science

Contribution Types: Model analysis & interpretability, NLP engineering experiment

Languages Studied: English

Submission Number: 1401