Unifying the structures of language in a neural population code.

Large language models encode linguistic diversity into a high-dimensional space that closely mirrors neural population codes in the human brain.

• Why it matters: Understanding how language emerges from neural activity is a major challenge in cognitive neuroscience, and current symbolic models fall short of capturing natural language's complexity.
• What they did: The study compares LLMs' statistical, distributed representations with neural population codes, using insights from artificial intelligence and neuroscience to assess their cognitive plausibility.
• The result: Findings suggest that language processing in humans may rely on similar high-dimensional, distributed codes as in LLMs, prompting a shift toward new theoretical frameworks beyond algebraic-symbolic models.