Visual-Seeker: Towards Visual-Native Multimodal Agentic Search via Active Visual Reasoning

## The Core Problem

Current multimodal AI agents—systems that combine vision and language—treat images like static snapshots. When they search for information online, they rely heavily on text clues and miss the rich visual details that could guide their investigation. This limits their ability to answer complex questions that require piecing together evidence from multiple sources and reasoning across both images and text.

## What Visual-Seeker Does

The authors introduce Visual-Seeker, an agent that actively looks at fine-grained visual details as it searches, rather than passively consuming images upfront. Instead of a search process that goes: "read text → look at image → conclude," the agent continuously revisits and examines different parts of images as new questions arise during multi-hop reasoning (reasoning that requires multiple steps to connect evidence).

The key insight is treating vision as an active, dynamic tool. When the agent encounters a visual clue—say, a product detail in a webpage screenshot—it can zoom in, compare details across images, or highlight regions relevant to the current search goal. This mirrors how a human would investigate a complex claim by repeatedly examining visual evidence from different angles.

## The Training Innovation

To teach the agent this behavior, the authors built an active visual reasoning data pipeline and synthesized 5,000 high-quality training trajectories—step-by-step examples showing how an agent should reason through a multimodal search task. Each trajectory includes the agent's decisions about *where to look* visually, what text to extract, and how to connect findings across sources. This dataset is the scaffolding needed to train a model that naturally performs visual-native reasoning.

## Why It Matters

The benchmark results are striking. Visual-Seeker achieves state-of-the-art performance across five challenging multimodal search benchmarks—datasets designed to test real-world web search scenarios with complex, ambiguous queries. Notably, it outperforms several proprietary models, suggesting the visual-native approach genuinely unlocks capabilities that text-dominant methods miss.

Real-world impact: fact-checking misinformation, e-commerce product verification, or academic research all benefit from an agent that can cross-reference visual details with textual claims. A system that asks "Wait, let me look more closely at that image" often catches contradictions a text-first agent would overlook.

## The Takeaway

Rather than bolting vision onto a language-first architecture, the authors designed an agent where vision and reasoning are woven together from the start. The synthesis of 5K training trajectories and the active visual reasoning pipeline represent a shift in how we should think about multimodal agents: not as language models that occasionally glance at images, but as systems that actively harvest visual evidence as they think.