Bridging the Gap between Theory of Mind and Action in LLMs

While Large Language Models (LLMs) demonstrate impressive capabilities in Theory of Mind (ToM) tasks, translating this social understanding into strategic action remains a critical challenge. In this paper, we address the Knowing Doing Gap in social intelligence, where models correctly infer the mental states of others yet fail to select actions consistent with those inferences. To rigorously evaluate this discrepancy, we introduce the Theory of Mind-to-Action (ToMA) benchmark, a unified framework designed to analyze the causal link between social reasoning and decision making in diverse cooperative and competitive scenarios. Our empirical analysis reveals that reasoning capabilities do not guarantee strategic behavior, identifying specific failure modes where reasoning decouples from action. To bridge this gap, we propose alignment mechanisms that enforce a sequential dependency between perspective taking and execution. By prioritizing the explicit analysis of others' mental states before decision making, social reasoning strictly conditions strategic actions.