Consider citing Vision-OPD?

[YqiAnH]

@yzhbradoodrrpurp @shx2005

Hi great work!

I'd like to point you to a closely related prior paper that worth citing: Vision-OPD: Learning to See Fine Details for Multimodal LLMs via On-Policy Self-Distillation (arXiv:2605.18740, code: https://github.com/VisionOPD/Vision-OPD). It shares V-Zero's core setup—using on-policy distillation from a crop-conditioned teacher to a full-image student, is answer-label-free, needs no RL/reward verifier, and adds no inference-time tools.

Since crop-conditioned OPD is central to V-Zero, it'd be great to cite and discuss Vision-OPD in the paper. V-Zero's contribution is a clear addition on top of this line of work, so positioning it relative to Vision-OPD would help readers understand what's new.

[shx2005]

@YqiAnH Hi, thank you very much for pointing this out.

We became aware of Vision-OPD during the submission period and consider it a closely related concurrent work. We appreciate the reminder and will cite it in the next version.

At the same time, we believe V-Zero differs from Vision-OPD in several important aspects, particularly in its contrastive visual evidence design and its use of positive/negative evidence to gate on-policy distillation. We will add a dedicated discussion to clearly position V-Zero with respect to Vision-OPD, highlighting both the shared motivation and the key methodological differences.

Thanks again for the helpful suggestion!

[shx2005]

@YqiAnH By the way, we also want to say that we genuinely find Vision-OPD inspiring. It is a great contribution to this line of work, and we really appreciate the perspective it brings to on-policy distillation in vision domain.