Large Language Models (LLMs) play a key role in social simulations but most existing virtual agents retain little if any dynamic adaptability and do not convincingly express different emotional states, limiting the fidelity of the simulation despite advancements in generation population-aligned persona. We present EvoPersona, a design to enhance population-aligned persona with both dynamic contextual awareness and emotional dynamics. EvoPersona has two components: a Contextual Awareness Module, built on an instruction-tuned small LLM, that allows the agent (persona) to adapt its behavior and language style in response to situational cues in real-time; and an Emotional Dynamics Module based on an evolving internal emotional state that is driven by real-time emotional input analysis and subject to Reinforcement Learning supervised feedback to ensure that the emotional state is naturalistic and, importantly, consistent with the context. We report results from five extensive experiments showing continued strong population-level psychological alignment along with significantly better contextual coherence and emotional realism than baseline state-of-the-art models.
Martin et al. (Thu,) studied this question.