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

2026-08-02

AI Summary

This work addresses a critical limitation in current large language models (LLMs) when performing theory of mind (ToM) reasoning: their reliance on event-hiding strategies, constrained output formats, and tendency to generate responses based on information unknown to the characters, which degrades performance. To overcome this, the authors propose PICTURE, a novel prompting method that explicitly models a character’s ignorance of specific events within free-form chain-of-thought (CoT) reasoning, replacing conventional event-hiding mechanisms. Without altering model architecture, PICTURE leverages prompt engineering to guide LLMs to accurately articulate knowledge gaps, thereby avoiding irrelevant or hallucinated responses. Evaluated on standard false-belief tasks, PICTURE outperforms existing prompting approaches by an average of 7.3%, significantly enhancing ToM reasoning accuracy.

Abstract

Simulating human-like Theory of Mind (ToM) has been a longstanding problem in natural language processing (NLP). To address this, existing works introduce a reasoning step of event hiding (a.k.a. perspective-taking), where events unknown to a character are removed before question answering. However, resorting to event hiding for ToM reasoning presents a performance degradation issue due to the strict output format constraints involved in event hiding. To mitigate this issue, we propose generating perspective-taking outputs as free-form explanations without event hiding, but this poses a notable yet underexplored challenge: LLMs need to inhibit responses to events unknown to characters, because the absence of event hiding exposes LLMs to these events throughout reasoning. To address this challenge, we hypothesize and empirically verify that LLMs can achieve such inhibition if a character's lack of knowledge about events is made explicit during reasoning. Based on this finding, we introduce PICTURE, a new prompting method that enables LLMs to generate a character's lack of knowledge within free-form Chain-of-Thought (CoT). Experimental results show that PICTURE outperforms existing prompting methods by an average of 7.3% on false-belief tasks.

Problem

• Theory of Mind
• event hiding
• large language models
• false-belief tasks
• perspective-taking

Innovation

• Theory of Mind
• event hiding
• free-form Chain-of-Thought
• knowledge inhibition
• prompting method