VistaHop benchmark tests multi-hop visual reasoning

Arxiv CS.CV (computer vision)

VistaHop is a new benchmark for evaluating multimodal large reasoning models (MLRMs) on multi-hop visual reasoning in Visual DeepSearch tasks. It requires agents to inspect image regions, ground intermediate reasoning in visual evidence, and connect fine-grained clues across long contexts.

multimodal

Arxiv CS.CV (computer vision)·Hang HeChuhuai YueChengqi DongChengcheng WanTing SuHaiying SunJiajun ChaiXiaohan WangGuojun Yin

8 days ago