In Language Models Represent Beliefs of Self and Others, Wentao Zhu and colleagues delve into the often-overlooked facet of social cognition in LLMs, significantly Theory of Mind (ToM). They unveil how the internal representations of self and other agents’ beliefs can be decoded from these models, pointing to an innate potential for sophisticated social reasoning.

Key insights include:

• Evidence of internal representations in LLMs for different perspectives
• Influence on ToM performance through manipulation of these representations
• Potential generalizability to various social reasoning tasks

The findings suggest a deeper cognitive-like processing within LLMs that had been speculative until now. The implications for AI’s future role in social interactions and how we understand machine understanding are profound. It hints at a future where AI could be expected to empathize and contextualize social dynamics akin to humans.

Further reading at: ArXiv Link