Parameter Alignment Mitigates Catastrophic Forgetting in Multilingual Expert Language Models

In three linesStudy on preventing catastrophic forgetting during continual pretraining of multilingual language models. Authors propose five parameter alignment strategies (layer freezing, regularization, post-hoc reversion, model merging) tested across 32 languages and four evaluation axes. Parameter alignment substantially reduces forgetting while maintaining language acquisition.

Summary generated by Claude — human-verified