Reproduce “NAS-PO: Native Attention-Strategy Policy Optimization for Vision-Language Models” 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

Ke An, Shizhe Xiang, Yue Liu, Liang Chen, Pei Fu, Shaojie Zhang, Xiaoxing Hu, Yupei Wang

NAS-PO trains vision-language models to use native attention as a signal for visual grounding. It separates how much each response token attends to the image from where that attention falls, then uses these signals in trajectory-level credit shaping and positive-rollout attention alignment. On Qwen3-VL-8B, it reaches 68.04% overall Accuracy@8 versus GRPO’s 63.54%; patch-removal tests support region relevance, not a complete explanation.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.