LLMs' Hidden States Mirror Human Semantic Associations in Feature Space
Researchers found that geometric relations between semantic features in LLMs' hidden states closely match human psychological associations. The study projects 360 words onto 32 semantic axes, showing high correlation with human ratings.