Class prediction in a zero/few-shot setting at inference time

Hi everyone,

let’s say I have fine-tuned a model to classify text for 7 classes because I have a lot of examples for them.
Now I know that in reality there are 3 more classes for which I have very limited to no data.

This is where few-shot/zero-shot comes into play and makes it (kind of) possible for me to predict these additional 3 classes. Great.

Now to my question:
If I do few/zero-shot learning/classification on these classes the model makes no updates to its weights, so it actually does not “learn” them (This term always bothers me). It just transfers its knowledge of what the task is, what it knows about language and the additional training examples or class definitions I provide to these 3 new classes. (Just to clarify).

So how would I implement this at inference time?
The model is supposed to predict all 10 classes. As far as I understand the “zero-shot-classification” pipeline will only focus on the new classes?

I would really appreciate it if someone could explain this to me, since I seem to be missing an important point.
Best wishes