Dispersion loss counteracts embedding condensation in small language models

Dispersion loss counteracts embedding condensation and improves generalization in small language models (ICML 2026).