Preprint: EGA beats PCA for estimating dimensionality in LLM embeddings

A new preprint by Golino et al. finds PCA unreliable for estimating the dimensional structure of LLM item embeddings, while exploratory graph analysis (EGA) achieved 97.6–100% recovery across tests on three LLMs including GPT-4o.

2026-09-30 ~ 2026-09-30 · 2 related posts

• Preprint: stop using PCA on LLM item embeddings — EGA recovers structure 97.6-100% of the time — GolinoHudson · 2026-09-30

1 near-duplicate retellings: GolinoHudson