Language models mimic human belief reasoning
A new study finds that large language models use a small, specialised internal circuit to perform Theory of Mind reasoning, a step towards more human-like AI cognition.
In a recent paper, researchers at Stevens Institute of Technology revealed that large language models (LLMs) use a small, specialised subset of their parameters to perform tasks associated with the psychological concept of ‘Theory of Mind’ (ToM), the human ability to infer others’ beliefs, intentions and perspectives.
The study found that although LLMs activate almost their whole network for each input, the ToM-related reasoning appears to rely disproportionately on a narrow internal circuit, particularly shaped by the model’s positional encoding mechanism.
This discovery matters because it highlights a significant efficiency gap between human brains and current AI systems: humans carry out social-cognitive tasks with only a tiny fraction of neural activity, whereas LLMs still consume substantial computational resources even for ‘simple’ reasoning.
The researchers suggest these points as a way to design AI models that are more brain-inspired, selectively activating only those parameters needed for particular tasks.
From a policy and digital-governance perspective, this raises questions about how we interpret AI’s understanding and social cognition.
If AI can exhibit behaviour that resembles human belief-reasoning, oversight frameworks and transparency standards become all the more critical in assessing what AI systems are doing, and what they are capable of.
Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!
