VistaHop: A New Challenge for Visual DeepSearch Models

Visual DeepSearch is entering a new phase with the introduction of VistaHop. This benchmark aims to evaluate vision-centric search and multi-hop visual reasoning, a much-needed advancement in understanding and improving multimodal large reasoning models (MLRM).

Why VistaHop Matters

VistaHop isn't just another benchmark. It contains 300 high-resolution images, 25 visual search scenarios, and 350 multi-hop QA tasks. These tasks require models to follow complex evidence chains in images or combine information from multiple reasoning paths.

Most benchmarks focus on single-step visual understanding or static image-question answering. VistaHop, however, challenges models to inspect images iteratively, ground their reasoning in visual evidence, and connect clues across extended reasoning chains.

The Current State of MLRMs

How are current models performing against this new standard? Seven representative MLRMs were tested, with the best, SenseNova-MARS-32B, achieving a mere 24.31% Pass@1. It's a clear indicator of the existing gaps in the capabilities of these models.

These results highlight significant limitations in areas like visual grounding, evidence revisiting, long-chain reasoning, and multi-anchor information fusion. If current MLRMs can't keep up, what's the point of developing more complex models?

The Path Forward

Enter VistaArena, a unified evaluation environment that supports enhanced reasoning with tools like text search, image search, image cropping, and evidence-based answer validation. It's a step in the right direction, but there's still a long way to go.

For developers and researchers, VistaHop is a call to action. We need stronger benchmarks and more effective training methods. It's time to rethink how we approach multi-hop visual reasoning. The question is: will the community rise to the challenge?