Several questions regarding the paper

Dear authors,

First of all, very inspiring and novel work to the Graph LLM community. Yet, I have several questions regarding the paper details, especially the first stage of model training, which I hope can be clarified.

1. What is the [DEC] token in this case. Does it represent the overall token sequence of the textual input? or it is just a self-designed token at the start of the sequence to signal the decoding task?
2. The third objective in the first-stage training is to compute whether the query representation matches the textual representation through classifier, yet how do you define the two representations match or not? Does the output score the classifier indeed reflect the golden truth of whether they match?

I also try to find the answers of above questions via the original paper of 'BLIP', but they neither provide a clear illustration in their paper.

It would be greatly appreciated if you can help me understand these questions. Thank you!

Start by reading the paper's first-stage training description and the cited BLIP paper, focusing on the [DEC] token and the third classifier objective. The issue is done when the authors clarify the token's role and how matching labels and classifier scores are defined.

Written by the indexing model from the issue text.