“I apologize for my actions”: Emergent Properties of Generative Agents and Implications for a Theory of Mind
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
This work explores the design, implementation, and usageof generative agents towards simulating human behaviour.Through simulating (mis)information spread, we investigatethe emergent social behaviours they produce.Generative agents demonstrate robustness to(mis)information spread, showing realistic conversationalpatterns. However, this robustness limits agents’abilities to realistically simulate human-like informationdissemination. Generative agents also exhibit novel andrealistic emergent social behaviours, such as deception,confrontation, and internalized regret. Using deception,agents avoid certain conversations. Through confrontation,an agent can verify information or even apologize for theiractions. Lastly, internalized regret displays direct evidencethat agents can internalize their experiences and act on themin a human-like way, such as through expressing remorse fortheir actions.We also identify significant technical dynamics and otherphenomena. Generative agents are vulnerable to produce unrealistichallucinations, but can also produce confabulationswhich fill in logical gaps and discontinuities to improve realism.We also identify the novel dynamics of “contextualeavesdropping” and “behavioural poisoning”. Via contextualeavesdropping and behavioural poisoning, agent behaviour isaltered through information leakage and sensitivity to certainstatements, respectively.The social behaviors demonstrated by generative agents, suchas deception, confrontation, and internalized regret, suggest apreliminary avenue for considering elements of a Theory ofMind (ToM) in LLM-based systems. While these behaviorsdo not represent genuine understanding or intentionality, theyindicate a capacity to simulate human-like responses to socialand informational dynamics. For example, internalized regrethints at a mechanism for contextual adaptation, which couldbe seen as a rudimentary step toward representing aspects ofhuman mental states, albeit in a constrained sense.