
Paul Zhou
@zhiyuan_zhou_ • 1,770 subscribers
RL & robots. phd @berkeley_ai, prev @physical_int
Shorts
Videos

Do you ever find finetuning VLA overfits to the target task, to the point where generalist ability is lost and even minor deviations beyond the SFT data break the policy? We found an extremely simple solution: directly merge the base and finetuned policy in weight space 🤯 👇🧵
Paul Zhou128,624 просмотров • 10 месяцев назад
Больше нет контента для загрузки