Let It Go or Control It All? The Dilemma of Prompt Engineering in Generative Agent‐Based Models

What the paper says

The integration of large language models into agent‐based models has created generative agent‐based models, offering powerful new approaches for simulating complex social dynamics. However, their implementation faces a critical prompt engineering dilemma concerning the balance between simulation control and behavioral authenticity. This paper examines generative agent‐based models as prompt‐driven information conduction networks across four components: profile, memory, planning, and action. Through systematic analysis of 22 recent studies, we identify patterns of potential “over‐control” where prompt‐based design choices can inadvertently predetermine simulation outcomes rather than allow genuine emergent social dynamics. Our analysis reveals component‐specific manipulation techniques across agent identity construction, memory architecture, decision‐making frameworks, and behavioral constraints, as well as collaborative control systems that apply prompt manipulation across multiple components simultaneously. These findings underscore a fundamental epistemological dilemma: distinguishing between authentic emergent phenomena arising from agent interactions and methodological artifacts produced by prompt engineering decisions. This dilemma may stem from an inherent tension where descriptive prompts, while aiming to foster agent autonomy through abstract concept activation, suffer from low internal validity and behavioral inconsistency that necessitate introducing explicit constraints and guidance, yet such instructional interventions may lead to over‐control problems, creating a methodological paradox for researchers.