Ascertain optimal prompting fidelity for LLM-enabled avatar generation and simulation

Ascertain the optimal prompting fidelity—the level of detail and specificity in prompts—required to reliably generate stakeholder digital avatars and participatory simulation outputs using Large Language Models for a given SAEMS decision context, balancing realism, controllability, and computational cost.

Background

The baseline simulations demonstrate that avatar profiles and outputs can vary with prompt specificity. Determining an optimal prompt design would improve consistency and validity of synthetic stakeholder behavior and planning results.

Establishing prompting guidelines would help ensure that LLM-generated avatars adequately represent stakeholder preferences and improve the interpretability of multi-criteria evaluations.