Zhengbo Zhang, Changtao Miao, Jinbo Su, Zhaowen Zhou, Chunxia Zhang, Xukai Wang, Ruiqi Liu, Kaiyuan Zheng, Jiansheng Cai, Bo Zhang, Zhe Li, Shiming Xiang, Ying Yan
4 min
Abstract
Multimodal large language models (MLLMs) have demonstrated impressive capabilities in many visual tasks, but they often struggle with factual grounding when confronted with complex, open-world scenarios. While recent multimodal deep search agents attempt to address this issue by utilizing external tools, the visual-native search paradigm remains underexplored. Existing methods primarily rely on simple images with explicit semantics and text-only evidence trajectories, limiting the agent's ability to perform multi-hop, cross-modal reasoning and search. To address these limitations, we propose Visual-Seeker, a visual-native multimodal deep search agent via active visual reasoning. Rather than treating vision as a static input, our agent actively attends to fine-grained visual details, dynamically harvests visual evidence throughout the search process. To unlock its visual-native potential, we design an active visual reasoning data pipeline and synthesize 5K high-quality multimodal trajectories for model training. Extensive experiments demonstrate the state-of-the-art performance across five challenging multimodal search benchmarks, even surpassing several proprietary models, validating robust visual-native reasoning and search in real-world web environments. The code and data can be accessed at: https://github.com/ZhengboZhang/Visual-Seeker.
Alex: There was quite a bit of engineering. The pipeline starts by identifying what they call "seed entities" — the key, recognisable objects in a given photo. A person, a building, a logo. From there, they use a method they describe as a "dual-strategy random walk" to generate a branching path of questions that the model has to follow to reach an answer.
Sam: A "dual-strategy random walk" — can you unpack that? It sounds technical.
Alex: Sure. Imagine you're following a treasure map, but instead of one straight path, the map keeps branching. At each fork, you have to make a decision based on what you've just found. The "random walk" part means the path isn't predictable — it can go in different directions depending on what the image contains. The "dual-strategy" part means the system can branch based on either visual evidence or text-based context, whichever is more useful at that step. The point is that the model can't shortcut its way to an answer. It has to gather evidence at every stage.
Sam: So by the end of training, the model has practised hundreds of these branching investigations. It's learned that a single glance isn't enough — it needs to keep asking questions until it's actually built a complete picture.
Alex: That's the core idea, yes. And the paper refers to this as "multi-hop reasoning" — each "hop" is one step in the chain, where a visual clue leads to a search, which leads to a new clue, which leads to another search. The model learns to chain those hops together rather than stopping after the first one.
Sam: How did it actually perform? Did the approach hold up when tested?
Alex: The paper reports that Visual-Seeker performs better than existing models on these kinds of tasks — including some proprietary systems that are generally considered strong benchmarks. The researchers suggest this supports their core claim: that treating vision as an active, multi-step process produces meaningfully better results than passive, single-pass image reading. That said, this is one paper, and independent replication would strengthen the case.
Sam: It's a real shift in how we think about what AI "seeing" actually means. Not a single glance, but a genuine investigation. Thanks for walking me through it, Alex.
Alex: Thanks for listening to ResearchPod.