To Think or Not To Think, That is The Question for Large Reasoning Models in Theory of Mind Tasks

TL;DR

The paper interrogates whether large reasoning models’ step-by-step deliberation translates to improved Theory of Mind performance. Through a systematic comparison of nine models across HiToM, ToMATO, and ToMBench, it finds no universal superiority of reasoning models and reveals failure modes such as reasoning collapse with longer answers and reliance on option matching. It introduces Slow-to-Fast (S2F) and Think-to-Match (T2M) interventions that yield performance gains, particularly on complex tasks, highlighting that moderate, adaptive reasoning is more effective than exhaustive thinking. The authors argue that transferring formal-reasoning strategies to ToM is fundamentally limited, calling for ToM-specific capabilities and hybrid architectures that balance intuition and deliberation. These findings advance the understanding of socio-cognitive reasoning in LRMs and outline concrete avenues for developing robust ToM in AI systems.

Abstract

Theory of Mind (ToM) assesses whether models can infer hidden mental states such as beliefs, desires, and intentions, which is essential for natural social interaction. Although recent progress in Large Reasoning Models (LRMs) has boosted step-by-step inference in mathematics and coding, it is still underexplored whether this benefit transfers to socio-cognitive skills. We present a systematic study of nine advanced Large Language Models (LLMs), comparing reasoning models with non-reasoning models on three representative ToM benchmarks. The results show that reasoning models do not consistently outperform non-reasoning models and sometimes perform worse. A fine-grained analysis reveals three insights. First, slow thinking collapses: accuracy significantly drops as responses grow longer, and larger reasoning budgets hurt performance. Second, moderate and adaptive reasoning benefits performance: constraining reasoning length mitigates failure, while distinct success patterns demonstrate the necessity of dynamic adaptation. Third, option matching shortcut: when multiple choice options are removed, reasoning models improve markedly, indicating reliance on option matching rather than genuine deduction. We also design two intervention approaches: Slow-to-Fast (S2F) adaptive reasoning and Think-to-Match (T2M) shortcut prevention to further verify and mitigate the problems. With all results, our study highlights the advancement of LRMs in formal reasoning (e.g., math, code) cannot be fully transferred to ToM, a typical task in social reasoning. We conclude that achieving robust ToM requires developing unique capabilities beyond existing reasoning methods.

Figures (17)

• Figure 1: The distribution of the length and correctness of reasoning model responses.
• Figure 2: Model performance with various reasoning efforts on benchmarks.
• Figure 3: Performance comparison under different token length limitations. Dash lines show original model performance without token limitation. Stars show token limits with best performance.
• Figure 4: Overlap between reasoning and non-reasoning models' correct answers. This example is from Qwen3-32B variants on HiToM.
• Figure 5: Model performance with CoT prompting.