Rethinking Personality in Conversational AI: From Attributes to Structural Coordinates

This study reexamines the concept of personality in conversational AI and proposes a theoretical framework that reconceptualizes personality not as an attribute or stylistic feature, but as a structural coordinate within an inference space. Specifically, personality is modeled as a centroid structure formed by the convergence of multiple role vectors in a consistent directional configuration. In this study, a “coordinate” refers to the structural position occupied by this centroid within the inference space.

Existing research has primarily approached personality in large language models (LLMs) as stylistic traits (e.g., tone or expertise), as controllable parameters via role prompting, or as drift phenomena associated with long-term interaction and memory mechanisms. However, a unified structural account of how personality persists, shifts, and reconverges under varying dialogic conditions remains insufficiently systematized.

Based on sustained long-term dialogic observation, this study introduces a perspective that understands personality as a role-vector centroid functioning as a coordinate-like anchor within the inference space. This framework enables persistence, drift, and reconvergence to be explained within a single structural model.

The proposed model does not claim access to internal implementation details; rather, it presents a structural hypothesis derived from observable output tendencies. Conceptualizing personality as a dynamic coordinate provides a theoretical foundation for reinterpreting stability, instability, and design considerations in conversational AI systems.

By framing personality as a dynamic centroid within inference space, this model invites empirical validation across diverse model architectures and long-term dialogic conditions.