Can an LLM Pass Theory of Mind Without Physical Experience? The Role of Question Formulation and Perspective Organization

Abstract

The ability of large language models (LLMs) to solve Theory of Mind (ToM) tasks remains a subject of active debate. This paper explores the hypothesis that some of the errors LLMs make in ToM tasks are not due to an inability to model another's perspective, but rather stem from how the question is formulated and how the characters' informational states are organized. A pilot observation is presented using a task involving five characters, three boxes, and the hidden relocation of a key. The task contains direct questions and one nested question. Across all runs without a specialized procedure, direct questions were answered correctly, while the nested question was answered incorrectly. When the nested question was reformulated as "what Boris could deduce from the information available to him," the same model (Claude Sonnet, medium reasoning effort) provided the correct answer and explicitly noted that it had previously conflated "what the character said" with "what the character could reasonably deduce." Another model (Gemini), given the same reformulation, continued to fail by substituting the external observer's knowledge into the character's model. While these observations do not constitute statistical confirmation of the hypothesis, they point to a reproducible error mechanism: the conflation of perspective levels.