CVSearch: Empowering Multimodal LLMs with Cognitive Visual Search for High-Resolution Image Perception

Researchers introduced CVSearch, a training-free visual search framework for high-resolution image perception in multimodal LLMs.

CVSearch uses an assess-then-search pipeline: expert-assisted search first, then semantic-aware scanning when needed.

It also applies adaptive patching and complexity-driven bottom-up exploration to cover more image details with less wasted computation.

The goal is to reduce missed details and redundant processing while improving accuracy on high-resolution images.