How can VLM with a small number of parameters, like under 1B, improve visual reasoning ability?

[zhouyizhuo]

@tanhuajie @Ramyyang
Hello authors, I am trying to improve the reasoning ability of Qwen3-0.8B.

Do you have any suggestions on using your project?
Is there anything to pay attention to compared to the Qwen2VL-2B/7B model? ）.
I am very grateful for any suggestions!