OPD-V is a visual On-Policy Self-Distillation (OPSD) framework for improving multimodal reasoning under Modality Imbalance. Instead of relying on a single privileged teacher, OPD-V contrasts a Positive Teacher conditioned on a Zoom-In Image with a Negative Teacher conditioned on a Mask Image, and distills only the on-policy tokens selected by the resulting Modality-Balance Trust Region.
This checkpoint corresponds to OPD-V-Qwen3-VL-4B-Instruct trained with teacher-box positive images and random-mask negative images.
1@misc{opdv2026,
2 title={OPD-V: Self-Distillation Improves Visual Detail Perception in Vision-Language Models},
3 author={},
4 year={2026},
5 eprint={2608.05131},
6 archivePrefix={arXiv},
7 primaryClass={cs.CV}
8}
