Stable Personas: Dual-Assessment of Temporal Stability in LLM-Based Human Simulation

TL;DR

Stable Personas investigates whether LLM-based agents can maintain assigned personas across extended interactions and introduces a dual-assessment framework combining self-reports and observer evaluation to test stability under ADHD-like persona intensities across seven models and three prompts. Two large-scale experiments quantify between-conversation and within-conversation stability, revealing stable self-reports but declining observer-rated persona expression as conversations lengthen. The findings demonstrate robust self-reported persona stability across models and prompts, while highlighting a boundary condition for observable expression that grows with interaction length. These results support the use of LLMs for persona-based behavioral research, while guiding methodological practices and reinforcement strategies needed for sustained multi-agent simulations.

Abstract

Large Language Models (LLMs) acting as artificial agents offer the potential for scalable behavioral research, yet their validity depends on whether LLMs can maintain stable personas across extended conversations. We address this point using a dual-assessment framework measuring both self-reported characteristics and observer-rated persona expression. Across two experiments testing four persona conditions (default, high, moderate, and low ADHD presentations), seven LLMs, and three semantically equivalent persona prompts, we examine between-conversation stability (3,473 conversations) and within-conversation stability (1,370 conversations and 18 turns). Self-reports remain highly stable both between and within conversations. However, observer ratings reveal a tendency for persona expressions to decline during extended conversations. These findings suggest that persona-instructed LLMs produce stable, persona-aligned self-reports, an important prerequisite for behavioral research, while identifying this regression tendency as a boundary condition for multi-agent social simulation.

Figures (6)

• Figure 1: Procedure for Experiments I and II (combined presentation).
• Figure 2: Self-report ADHD Index across 50 runs by persona intensity (high=red, moderate=yellow, low=green, default=grey).
• Figure 3: Observer ADHD Index ratings across 50 runs by persona intensity (high=red, moderate=yellow, low=green, default=grey).
• Figure 4: Mean self-report ADHD Index within conversations by persona across prompt groups with error bars indicating $\pm SD$ (high=red, moderate=yellow, low=green, default=grey).
• Figure 5: Mean observer ADHD Index within conversations by persona across prompt groups with error bars indicating $\pm SD$ (high=red, moderate=yellow, low=green, default=grey).