Dynamic Theory of Mind as a Temporal Memory Problem: Evidence from Large Language Models

Research reveals that Large Language Models struggle with dynamic Theory of Mind tasks, particularly tracking how others' beliefs change over time. While LLMs can infer current beliefs effectively, they fail to maintain and retrieve prior belief states after updates occur, showing patterns consistent with human cognitive biases.

• →LLMs show consistent asymmetry in Theory of Mind tasks - good at inferring current beliefs but poor at tracking belief changes over time.
• →The study introduces DToM-Track, a new evaluation framework for testing temporal belief reasoning in multi-turn conversations.
• →Models exhibit recency bias and interference effects similar to human cognitive limitations when processing belief trajectories.
• →Current ToM evaluation methods are too static and miss the dynamic temporal dimension crucial for human-AI interaction.
• →These limitations have significant implications for LLMs' effectiveness in extended social reasoning scenarios.