The continual learning approach yields gains comparable to those typically seen across multiple successive model generations

This yields gains comparable to those typically seen across multiple successive model generations, at compute and personnel budgets substantially lower than commonly thought

http://arxiv.org/abs/2608.27147v1