PICTURE: Enhancing Theory-of-Mind in Large Language Models by Revealing, Not Hiding, Characters’ Lack of Knowledge

Large Language Models (LLMs) often struggle with Theory-of-Mind (ToM) tasks, specifically failing to distinguish between their own knowledge of reality and a character's limited perspective. Previous approaches attempted to solve this by 'event hiding'—manually removing information unknown to a character before asking the model to reason. However, these methods impose strict output format constraints that often degrade the model's overall reasoning performance. This paper investigates whether LLMs can instead be prompted to inhibit their own knowledge by explicitly stating what a character does not know, thereby maintaining a more natural, free-form reasoning process.

The authors introduce PICTURE (Perspective-taking with Generated Lack of Knowledge in Chain-of-Thought Reasoning). Instead of filtering the input, PICTURE instructs the LLM to perform a two-step process: first, it generates an explicit account of what the character knows and, crucially, what the character does not know. Second, it uses this generated perspective as a foundation for a standard 'think step by step' Chain-of-Thought (CoT) reasoning process. By explicitly verbalizing the character's ignorance, the model is better able to 'inhibit' its own access to the full story, preventing the common error of attributing the model's own knowledge to the character.

PICTURE consistently outperforms existing perspective-taking methods across multiple benchmarks, including BigToM, ToMi, and FANToM. On average, the method achieves a 7.3% improvement in accuracy on false-belief tasks. The authors demonstrate that this approach is effective across various model sizes, from Llama2-7B to GPT-4o, suggesting that the ability to inhibit irrelevant information is a generalizable skill that can be unlocked through proper prompting rather than just architectural changes or rigid input filtering.