Visual-Seeker Agent Enhances Multimodal Search with Active Visual Reasoning

Researchers have introduced Visual-Seeker, a novel multimodal deep search agent designed to operate with a "visual-native" approach. While existing multimodal large language models (MLLMs) excel in many visual tasks, they often struggle with factual grounding in complex, real-world scenarios. Visual-Seeker aims to overcome this by actively engaging with fine-grained visual details and dynamically extracting visual evidence throughout the search process. Unlike previous methods that primarily rely on simple images and text-only evidence, Visual-Seeker is built to perform multi-hop, cross-modal reasoning and search. To achieve its visual-native potential, the team developed an active visual reasoning data pipeline and synthesized 5,000 high-quality multimodal trajectories, which were used to train the model. Extensive experiments demonstrate that Visual-Seeker achieves state-of-the-art performance across five challenging multimodal search benchmarks. It even surpasses several proprietary models, validating its robust visual-native reasoning and search capabilities in real-world web environments. This advancement signifies a step towards more sophisticated and visually intelligent AI agents.