Software Engineering Institute

In this report, we explore how activities generated from the GHOSTS Framework’s non-player character (NPC) client, including software usage, compare to activities produced by GHOSTS’ default behavior and large language models (LLMs). We also explore how the underlying results compare in terms of complexity and sentiment. In our research, we leveraged the advanced natural language processing capabilities of generative artificial intelligence (AI) systems, specifically LLMs (i.e., OpenAI’s GPT-3.5 Turbo and GPT-4) to guide virtual agents (i.e., NPCs) in the GHOSTS Framework, a tool that simulates realistic human activity on a computer. We devised a configuration to fully automate activities by using an LLM, where text outputs become executable agent directives. Our preliminary findings indicate that an LLM can generate directives that result in coherent, realistic agent behavior in the simulated environment. However, the complexity of certain tasks and the translation of directives to actions present unique challenges. This research has potential implications for enhancing the realism of simulations and pushing the boundaries of AI applications within human-like activity modeling. Further studies are recommended to optimize agent understanding and response to LLM directives.