Customizing model behavior or supervision at the individual token level rather than at response or batch level.