SPeCtrum

A Grounded Framework for Multidimensional Identity Representation in LLM-Based Agents

Generative AI · Identity Modeling · Agent Personas · Human–AI Interaction · Contextual AI

Can LLM personas capture the complexity of human identity? Many existing approaches focus on a few amplified traits. This project explored whether combining multiple layers of selfhood — social, personal, and contextual — can create richer, more authentic personas.

Project Overview

SPeCtrum is a grounded framework for modeling LLM-based personas by integrating three complementary facets of self-concept: Social Identity (S), Personal Identity (P), and Personal Life Context (C). While many existing approaches emphasize a narrow set of traits, SPeCtrum seeks to capture both enduring characteristics and everyday context. We evaluated the framework through automated experiments with fictional characters and a human study with real participants.

Approach

We formalized structured data schemas for S, P, and C and embedded them into LLM prompts. Two evaluations tested the framework: (1) Automated evaluation with a Guess Who? character identification game and the Twenty Statements Test (TST) across seven ablation conditions (S, P, C, SP, SC, PC, SPC) using well-known TV and drama characters. (2) Human evaluation study with 80 participants who provided their own S/P/C data, then rated essays generated by different agent types for self-similarity and believability.

SPeCtrum evaluation design — Automated evaluation setup for SPeCtrum across seven profile conditions.

Results & Contributions

Context alone often captured identity best. For fictional characters, daily routines and preferences (C) were highly diagnostic of identity, sometimes on par with SPC.
Full SPC was needed for real people. For real participants, combining all three dimensions produced more authentic and representative personas than C alone.
Reverse inference showed clear limits. Context data (C) could sometimes imply social categories (S) and personal traits (P), but not reliably across individuals.
SPeCtrum offered a practical pipeline. The framework provides a replicable way to assemble S/P/C data, improving both the believability and evaluation of LLM personas.