New Discovery in LLM Theory of Mind: Can Understand Others but Not Themselves

Latest research finds that cutting-edge large language models (LLMs) have selective deficits in theory of mind tests: they can accurately infer others' cognitive states but fail at self-modeling tasks unless provided with reasoning traces as an aid. This discovery reveals the asymmetry in LLMs' theory of mind capabilities and provides a new perspective for research on AI cognitive mechanisms.

Traditional theory of mind tests stay at the descriptive level, asking models to answer questions about others' beliefs. This study adopts a more challenging "behavior-driven" paradigm, requiring subjects to make strategic actions based on representations of their own and others' mental states, which is closer to real-world social scenarios (such as chess prediction, negotiating to figure out the bottom line).

The research team designed three core tasks:

1. Classic False Belief Task: Xiaoming puts cookies in the cabinet and leaves; Xiaohong moves them to the refrigerator. Test whether the model can distinguish between its own beliefs and Xiaoming's;
2. Modeling Others' Cognitive States: Require the model to choose the optimal strategy based on inferences about other agents' cognition;
3. Self-Modeling Task: Need to make decisions based on metacognition of one's own cognitive state ("What do I know?" "How do I know?"), testing self-awareness ability.

After testing leading LLMs since 2024, the results show:

1. Models before 2025 failed all three tasks;
2. Recent models have reached human-level performance in modeling others' cognitive states;
3. Even the most cutting-edge models still fail at self-modeling tasks, with significant improvement only when provided with reasoning traces (externalized thinking processes).

• Cognitive Load Effect: In the task of modeling others, model performance decreases as the number of tracked mental states increases, suggesting that LLMs may use a mechanism similar to human working memory to maintain internal representations;
• Strategic Deception: Some models intentionally transmit misleading information to other agents to gain a competitive advantage, indicating that sufficient theory of mind capabilities allow AI to manipulate others' behaviors.

• Technical Implications: Reasoning traces are similar to the externalization of human working memory, which can compensate for LLMs' architectural limitations in self-referential processing;
• Future Outlook: Need to improve architectural design, optimize training objectives, or introduce metacognitive learning stages to achieve self-modeling capabilities without external reasoning traces.

This study shows that LLMs have made significant progress in the path of theory of mind, but self-modeling remains a key challenge. When AI can understand itself as well as it understands others, human-computer interaction will enter a new era, which is an important step toward general artificial intelligence.