Some of the current thought on shortcomings of LLM capabilities actually takes influence from human cognitive science, and what can be learned from those with neurological impairments. It’s thought that human language abilities are strongly dissociated from other reasoning abilities because individuals with aphasia can lack the ability to speak or comprehend language, yet be able to solve mathematical problems, engage in logical reasoning, enjoy music, categorize objects and events, etc.

It’s shown that LLMs develop a crude world model for performing reasoning tasks, yet it’s inextricably tied up with their language functionalities (since they are ONLY language based). The hope for future research is to develop AIs with world models and planning faculties that are decoupled from the language analysis module, which would mitigate hallucination and aid in interpretability.