[WithinReason]

They are different but transformers don't have fixed windows, you can extend the context or make it smaller. I think you can extend a positional encoding if it's not a learned encoding.