What is shared classification head?

[douyh]

In paper, you metioned:

Did you mean llm hidden states were not fed into lm_head, but a single 256-way mlp? In this way, we don't need to add special tokens to the vlm?

Thanks.