From Anisotropy to Anomaly: An Online Geometric Diagnostics Framework for Transformer Training

ICLR (under review)

Recommended citation: Z. Hu et al. (2026). "From Anisotropy to Anomaly: An Online Geometric Diagnostics Framework for Transformer Training." ICLR (under review).

Developed online diagnostics based on statistical dependence among hidden representations to detect anomalies during Transformer training, including overfitting, data anomalies and loss spikes.

Authors: Z. Hu et al.