TimeToM: Temporal Space is the Key to Unlocking the Door of Large Language Models' Theory-of-Mind

TL;DR

TimeToM introduces a temporal space around narratives to address gaps in Large Language Models' Theory of Mind (ToM). It constructs Temporal Belief State Chains (TBSC) for each character and splits beliefs into self-world and social-world components, using a Time-Aware Belief Solver to leverage belief communication windows and transform higher-order ToM into first-order reasoning. Across ToMI, BigToM, and FanToM benchmarks, TimeToM yields substantial improvements across multiple LLMs, with pronounced gains on higher-order ToM and in dialogue scenarios, and demonstrates improved robustness via All/All* metrics. Limitations include reliance on accurate TBSC construction for smaller models and a focus on textual modalities, suggesting avenues for multimodal extension and further refinement.

Abstract

Theory of Mind (ToM)-the cognitive ability to reason about mental states of ourselves and others, is the foundation of social interaction. Although ToM comes naturally to humans, it poses a significant challenge to even the most advanced Large Language Models (LLMs). Due to the complex logical chains in ToM reasoning, especially in higher-order ToM questions, simply utilizing reasoning methods like Chain of Thought (CoT) will not improve the ToM capabilities of LLMs. We present TimeToM, which constructs a temporal space and uses it as the foundation to improve the ToM capabilities of LLMs in multiple scenarios. Specifically, within the temporal space, we construct Temporal Belief State Chain (TBSC) for each character and inspired by the cognition perspective of the social world model, we divide TBSC into self-world beliefs and social world beliefs, aligning with first-order ToM (first-order beliefs) and higher-order ToM (higher-order beliefs) questions, respectively. Moreover, we design a novel tool-belief solver that, by considering belief communication between characters in temporal space, can transform a character's higher-order beliefs into another character's first-order beliefs under belief communication period. Experimental results indicate that TimeToM can dramatically improve the reasoning performance of LLMs on ToM questions while taking a big step towards coherent and robust ToM reasoning.

Figures (6)

• Figure 1: (A) and (B): The structure of story and dialogue, as well as ToM questions in reading comprehension and dialogue scenarios. (C): ToM reasoning errors made by LLMs. (D): Reasoning perspective of belief solver.
• Figure 2: Pipeline overview of TimeToM, which constructs a temporal space and uses it as the foundation to improve the ToM capabilities of LLMs. TimeToM does not require training, it works in a zero-shot setting.
• Figure 3: Performance comparison of TimeToM and baselines on first-order and higher-order ToM questions.
• Figure 4: Case 1: The benefit of constructing temporal space. Case 2: The comparison between the belief solver as prompt and as feedback. Case 3: The application of TimeToM in situations involving agent communication.
• Figure 5: Story structure of ToMI.