High-resolution visual question answering often fails because models cannot see small, localized details. Standard methods that zoom sequentially must choose a region before getting an overview. This paper introduces VPS, a framework where a main agent first inspects image tiles in parallel using question-specific sub-agents, then adaptively zooms into a precise or merged region.
Across three model sizes and five benchmark splits, VPS improved accuracy over dedicated zoom-only search in 14 of 15 comparisons, with gains up to 8.0 points and a consistent 3.2-point advantage on ZoomBench at every tested size. The method works both as a training-free inference scaffold and with post-training (supervised fine-tuning and reinforcement learning).
The key insight is that decoupling parallel tile inspection from adaptive zoom makes the search process both more effective and more amenable to learning. This addresses a fundamental bottleneck in high-resolution visual reasoning without requiring expensive sequential decisions.