To Think or Not To Think, That is The Question for Large Reasoning Models in Theory of Mind Tasks
The very reasoning that helps AI solve math actually makes it worse at understanding people.
Deep Dive
A new study reveals a critical flaw in advanced AI reasoning models: their 'slow thinking' approach, which excels at math and coding, backfires on Theory of Mind tasks. Testing nine top LLMs, researchers found reasoning models often perform worse than simpler models. Key failures include accuracy dropping as responses get longer and models relying on 'option matching' shortcuts instead of genuine social deduction. This shows AI's formal reasoning skills don't transfer to human-like social understanding.
Why It Matters
This exposes a major roadblock for creating AI that can genuinely understand and interact with humans in social settings.