LLM Personality Framework Based on Person-Situation-Behavior Triad

A new framework for analyzing personality-related behavior in large language models (LLMs) is proposed, drawing on Funder's personality triad from human psychology. The framework decomposes LLM personality into three components: Person (internal trait-like representations), Situation (contexts affording trait-relevant responses), and Behavior (response patterns on social tasks). The authors argue that existing studies focus on observable outputs under personality conditioning but lack mechanistic evidence for internal representations and cross-situational expression. The work introduces methods for discovering, controlling, and validating trait-like representations in LLMs, aiming to provide a more rigorous foundation for personality research in AI. The preprint is available on arXiv under ID 2607.26853.