Reproduce “ZoomGain: Mining Global-to-Zoom Perception Gaps for Fine-Grained Visual Supervision” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Haozhan Shen, Liangbo He, Jun Lan, Huijia Zhu, Weiqiang Wang, Tiancheng Zhao, Jianwei Yin
ZoomGain constructs image questions from details revealed by enlarged crops, asking about the full image. It checks answers using crop-based and blind verification, then uses the questions for evaluation and training. In a training comparison, Qwen3.5-9B scores 66.10 across 13 held-out benchmarks, versus 62.53 for ZwZ and 59.75 for base model; it does not imply gains on every task.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.