Theory of Mind (ToM)—the capability to infer unobservable mental states of users based on their behavior—remains a fundamental challenge for Large Language Models (LLMs). Recently, pioneering studies have treated ToM as a static prediction task, directly generating final inferences. However, such approaches neglect the inherently dynamic nature of ToM, where mental states are not explicitly stated but are progressively inferred from contextualized behaviors. To bridge this gap, we first define the reasoning process underlying mental states as a crucial aspect of the scene about LLMs. Inspired by Goffman’s dramaturgical theory, we propose DMS-ToM, treating reasoning as a staged cognitive performance. DMS-ToM structures ToM inference into three coordinated stages: the Front Stage, which supports structured mental simulation; the Off Stage, where scene progression is guided by a director-like mechanism; and the Back Stage, responsible for mental-state attribution and consolidation. Through this staged organization, DMS-ToM progressively reconstructs mental states from the evolution of expressed reasoning. DMS-ToM demonstrates strong capability on challenging ToM reasoning tasks, achieving an overall accuracy of 79.5%, outperforming both structured reasoning baselines—BIP-ALM (76.7%) and AutoToM (75.5%)—and the prompt-based SimToM (50.7%), as well as LLMs including GPT-4o (51.2%) and DeepSeek V3.1 (52.8%).