VistaHop: Benchmarking Multi-hop Visual Reasoning for Visual DeepSearch

2026-06-02

AI Summary

Existing benchmarks struggle to evaluate the capabilities required for visually grounded deep search, particularly iterative image inspection, visual anchoring, and multi-hop evidence integration. To address this gap, this work proposes VistaHop—the first multi-hop visual reasoning benchmark specifically designed for visual deep search—and introduces the accompanying evaluation environment, VistaArena. This framework integrates high-resolution images, multi-hop question answering, tool-augmented reasoning, and evidence-based answer verification, enabling systematic assessment of vision-centric search and long-chain reasoning abilities. Experiments across seven leading multimodal large language models reveal that even the best-performing model, SenseNova-MARS-32B, achieves only a 24.31% Pass@1 accuracy, highlighting significant deficiencies in visual grounding and cross-path information fusion.

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.

Problem

• Visual DeepSearch
• multi-hop visual reasoning
• visual grounding
• evidence integration
• benchmarking

Innovation

• multi-hop visual reasoning
• Visual DeepSearch
• visual grounding
• evidence chain
• tool-augmented reasoning