generative recommendation

A Better Way to Seed New LM Tokens

GTI grounds new vocabulary tokens before fine-tuning, aiming to preserve distinctions that mean initialization tends to collapse.