Proceedings of the Annual Meeting of the Cognitive Science Society

Individuals frequently endorse views they privately reject to align with a perceived majority. This dynamic, known as pluralistic ignorance, is challenging to model because traditional scalar agents lack the semantic capacity to distinguish between private belief and public performance. To address this, we propose the Generative Cognitive-Social Simulation (GCSS) framework. We introduce the Grounding-Evaluation-Masking (GEM) architecture, which utilizes Large Language Models to simulate the strategic suppression of dissent based on Theory of Mind (ToM) risk assessment. Virtual classroom simulations suggest that ToM-guided risk evaluation is important for the emergence of the "spiral of silence." Crucially, we identify a computational mechanism of belief internalization: under prolonged masking, agents gradually adjust their private beliefs toward their publicly maintained stances. These findings quantify how social pressure can transform strategic conformity into more durable, collectively reinforced belief.