-
Notifications
You must be signed in to change notification settings - Fork 4
Description
Hi, and congratulations on the acceptance to NeurIPS25 for an Oral presentation!
I wonder how does MokA perform conversation-like inference. If my understanding is correct, with each question the modality tokens would have to interact in the model with the new question tokens, which would require to recompute the whole sequence for each new question. This is due to the novel cross-attention mechanism this work suggests. I would really appreciate your guidance in this regard, as this would help adoption of your method to real-world models.
Regarding NeurIPS, may I ask for the expected time when the updated manuscript would be published on arXiv? I'm interested about the (possible) changes to the paper due to the demanding rebuttal of this year's edition.
Best regards,
Antoni Kowalczuk