SR Values Significantly Lower Than Reported (24 and 19 vs. 38.3±0.6)

[LiLingFeng-ydmy]

Dear Author,

Thank you for your work!

While reproducing your paper, I have encountered some issues. I used LA_MODEL_NAME="gemini-3.1-pro-preview" together with VA_MODEL_NAME="qwen3.5-27b", without changing any other parts. I conducted two experiments in total, and the SR values were 24 and 19, respectively, which are far below the SR of 38.3±0.6 reported in your paper.

Could you please let me know whether this situation is normal?

Thank you for your reply!

[carottX]

Hi @LiLingFeng-ydmy,

Thank you for your interest in our work! We haven't tested this exact model combination ourselves, but the drop in SR could be due to some configuration mismatches. Could you please check the following:

• Is the thinking effort for gemini-3.1-pro-preview set to the maximum?

• Is the Chain of Thought (CoT) turned off for qwen3.5-27b?

• Crucial check: Qwen3.5 outputs relative coordinates [0, 1000]. Have you converted these into absolute coordinates (for a 640x480 image) before computing the world coordinates?

We suggest using the simple visualization website to verify if everything goes well.

Hope this points you in the right direction!

[LiLingFeng-ydmy]

Thanks for your reply！
I completely forgot to convert those relative coordinates [0, 1000] to absolute ones for the 640×480 image before computing world coords. Oops.
Really appreciate your help！