Probing

Train a small classifier on intermediate activations and see what it can recover. It is the oldest tool here and the easiest to over-read: a probe finding information does not establish that the model uses it.