This review examines whether large language models (LLMs) serve as cognitively plausible models of human neural language processing, drawing on work at the intersection of AI and cognitive neuroscience to compare LLM high-dimensional embedding spaces with neural population codes.
The authors argue that LLMs' context-driven, distributed representations closely parallel neural population codes in the human language system, suggesting biological and artificial systems may have converged on a similar computational strategy — one that cannot be explained by algebraic-symbolic theories alone.
This is a theoretical/review paper with no original empirical data; conclusions about neural-LLM alignment rely on synthesized prior work rather than new experiments, and causal claims about convergent computation remain speculative.
Explore related topics
