{"ID":3005045,"CreatedAt":"2026-06-03T03:09:48.883664427Z","UpdatedAt":"2026-06-05T07:32:30.50480903Z","DeletedAt":null,"paper_url":"https://arxiv.org/abs/2606.03273","arxiv_id":"2606.03273","title":"VistaHop: Benchmarking Multi-hop Visual Reasoning for Visual DeepSearch","abstract":"Visual DeepSearch requires multimodal large reasoning model (MLRM) agents to answer complex visual queries by repeatedly inspecting image regions, grounding intermediate reasoning in visual evidence, and connecting fine-grained clues across long reasoning chains. However, existing benchmarks mainly focus on single-step visual understanding or static image-question answering, offering limited evaluation of iterative image inspection, visual-anchor grounding, and multi-hop evidence integration. In this work, we introduce VistaHop, a benchmark for evaluating vision-centric search and multi-hop visual reasoning in Visual DeepSearch. VistaHop contains 300 high-resolution images, 25 visual search scenarios, and 350 multi-hop QA tasks that require models to follow evidence chains from visual anchors or fuse information across multiple image-grounded reasoning paths. We further develop VistaArena, a unified evaluation environment that supports tool-augmented reasoning with text search, image search, image cropping, and evidence-based answer validation. Experiments on seven representative MLRMs show that current models remain far from solving VistaHop: the best model, SenseNova-MARS-32B, achieves only 24.31% Pass@1. These results reveal persistent limitations in visual grounding, evidence revisiting, long-chain reasoning, and multi-anchor information fusion, highlighting the need for stronger benchmarks and training methods for Visual DeepSearch.","short_abstract":"Visual DeepSearch requires multimodal large reasoning model (MLRM) agents to answer complex visual queries by repeatedly inspecting image regions, grounding intermediate reasoning in visual evidence, and connecting fine-grained clues across long reasoning chains. However, existing benchmarks mainly focus on single-step...","url_abs":"https://arxiv.org/abs/2606.03273","url_pdf":"https://arxiv.org/pdf/2606.03273v1","authors":"[\"Hang He\",\"Chuhuai Yue\",\"Chengqi Dong\",\"Chengcheng Wan\",\"Ting Su\",\"Haiying Sun\",\"Jiajun Chai\",\"Xiaohan Wang\",\"Guojun Yin\"]","published":"2026-06-02T07:37:23Z","proceeding":"cs.CV","tasks":"[\"cs.CV\",\"cs.AI\",\"cs.CL\"]","methods":"[]","has_code":false}