ILO Metric Outperforms Other Multilingual Interpretability Methods

Comparing four metrics across 21 models (125M-14B parameters), researchers found that anisotropy causes CKA, ANC, and GMM dominance to fail. ILO achieved a Spearman correlation of 0.90 with cross-lingual transfer, significantly outperforming existing methods.