Is There a Way to Control the Behavior of Pre-Trained Models?

I feel like I've seen something kind of interesting related to that sort of continual learning where you can say, oh, OK, yeah, we trained this thing. And now we don't really need all those weights. Let's just keep these. Later on, the future ones don't mess with these ones too much because these are the only ones that really matter for the stuff we've worked so far. It will become potentially relevant with large pre-trained models as well. Pre-trained models have such a large store of different information. Right now, with just normal forward passes, you just get some weird combination of everything that the model knows. But maybe the model actually knows the individual