Preprint: stop using PCA on LLM item embeddings — EGA recovers structure 97.6-100% of the time

GolinoHudson · x · 2026-09-30

A new preprint by Garrido, Russell-Lasalandra, Rodríguez-Montoya and Golino delivers a blunt recommendation: stop using PCA to estimate dimensional structure from LLM item embeddings.

The team tested PCA vs. Exploratory Graph Analysis (EGA) across 3 LLMs (GPT-4o, GPT-5.4, Claude Sonnet 4.6), 2 embedding models (OpenAI text-embedding-3-small, Jina v3), 6 known personality dimensions, hundreds of Monte Carlo replications, plus an empirical replication on the Multidimensional Schizotypy Scale.

Key findings:

EGA recovered the correct 6-dimensional structure in 97.6-100% of conditions; PCA + Kaiser and PCA + Parallel Analysis achieved essentially 0% exact recovery
For a true 6-factor structure, PCA frequently estimated 8-24 dimensions. The cause: few items create a long tail of "nuisance" dimensions, so eigenvalue-based methods massively overextract
In their AI-GENIE package, UVA removes redundant items while bootEGA flags items with unstable dimensional placement — redundancy ≠ structural instability, and different LLMs showed different failure profiles
On the real psychological instrument, EGA came substantially closer to the human-response benchmark than PCA under both embedding models

Original post →

Preprint: stop using PCA on LLM item embeddings — EGA recovers structure 97.6-100% of the time

More from Research