Does FV hold for mutli-token transformation tasks?

[MarSaKi]

Dear author,
Thanks for your great work and generous open-sourcing efforts!
Function vectors are indeed a remarkable discovery. After reading your paper, I noticed that the tasks where function vectors hold true often involve relatively simple mappings from multi-token inputs to single-token outputs, such as antonyms or word-level translations.
This naturally leads me to a question: Do function vectors also hold for multi-token to multi-token mapping tasks, such as sentence-level translation? Could you please share some of your insights on this? Additionally, are there any in-context sentence translation datasets that could be used to explore this question? Or perhaps related literature that has already investigated this topic?

I deeply appreciate your time and thoughts on this matter!