Visual-Seeker: visual-native multimodal search surpasses rivals

Visual-Seeker, a paper submitted to arXiv on 13 Jun 2026 by Zhengbo Zhang and 12 other authors, presents a visual-native multimodal deep search agent that performs active visual reasoning. The paper frames the problem as a shortfall of current multimodal large language models, which the authors say "often struggle with factual grounding when confronted with complex, open-world scenarios." Visual-Seeker treats vision not as a static input but as a dynamic source of evidence gathered through active attention.

What Visual-Seeker does

The agent actively attends to fine-grained visual details and dynamically harvests visual evidence throughout a search session. The paper positions this approach against existing multimodal deep search agents that "primarily rely on simple images with explicit semantics and text-only evidence trajectories," which the authors argue limits multi-hop, cross-modal reasoning and search. Visual-Seeker is described as a "visual-native multimodal deep search agent via active visual reasoning." The code and data are linked in the paper and can be accessed at a URL provided by the authors.

Training, data and results

To train the agent, the authors designed an active visual reasoning data pipeline and synthesized 5K high-quality multimodal trajectories for model training. The paper reports "extensive experiments" and claims state-of-the-art performance across five challenging multimodal search benchmarks. It also states the model even surpasses "several proprietary models." The submission lists Zhengbo Zhang as first author alongside 12 co-authors and includes an arXiv-issued DOI for reference.

Why it matters

Multimodal models often fail to ground facts in open-world visual contexts, the paper argues. By making the search process visual-native and by harvesting visual evidence actively, Visual-Seeker aims to close that gap and enable multi-hop, cross-modal reasoning that static visual inputs and text-only evidence traces cannot. If the paper's claims hold under broader scrutiny, the approach could shift how multimodal search agents are trained and evaluated, particularly on complex, real-world web tasks.

What to watch

Watch for the authors' linked code and dataset at the URL in the paper and for independent evaluations of the claimed state-of-the-art performance on the five multimodal search benchmarks. The 5K synthesized multimodal trajectories and how they are used in the active visual reasoning pipeline will be the key materials to inspect when reproducing the reported results.