kleinnner/

camus-06-eigenvector-analysis

Eigenvector Analysis of LLM Embedding Spaces

LLM embedding spaces are high-dimensional (1536 dimensions). Most dimensions encode noise or non-semantic artifacts. Retrieval in full-dimensional space is computationally expensive and less accurate than retrieval in a reduced semantic subspace.