@ZeyiAndyLiu on Backlist

Spectral Lens: looking past loss curves in LLM training (t.co)

Activation and gradient spectra can expose representation geometry, forecast token efficiency early, and separate real learning gains from throughput gains