SSEToM: Spatially Guided Reasoning Enhances Theory of Mind in Large Language Models

ACL ARR 2025 February Submission7384 Authors

Abstract: Theory of Mind (ToM) refers to an individual's ability to understand and infer the mental states of others. While this capability develops naturally in humans, equipping Large Language Models (LLMs) with similar abilities remains challenging. Some chain-of-thought (CoT) methods, such as SimToM, have improved LLM performance in ToM reasoning. However, existing methods often overlook the spatial dimension perception that humans utilize when solving ToM problems. To address this limitation, we propose SSEToM, a method inspired by the Event Segmentation Theory (EST) in psychology, which posits that individuals in different spatial locations perceive information about events within their respective environments. SSEToM segments ToM stories into discrete events based on spatial dimensions, enhancing the LLM's ability to perceive and reason about the mental states of the discussed characters. Experiments conducted on three datasets demonstrate that SSEToM significantly enhances LLMs' reasoning capabilities in ToM tasks, achieving state-of-the-art performance.

Paper Type: Long

Research Area: Human-Centered NLP

Research Area Keywords: values and culture, Theory of Mind, Event Segmentation Theory, spatial dimension

Contribution Types: NLP engineering experiment

Languages Studied: English

Submission Number: 7384