Fetching the paper…

Transformers generalize differently from information stored in context vs in weights · Around