Context: cos i could not do on the link.

If I understand this correctly, this paper enables a model to expand it's training to unseen tokens, without having to train from scratch, and without forgetting previous knowledge. (no catastrophic forgetting). Any thoughts on this approach?