Abstract
Training multimodal search agents for multi-hop reasoning is challenging due to a fundamental structural disconnect: existing pipelines construct training data, search environments, and reward signals independently, leading to the discarding of synthesized structural metadata, reliance on irreproducible external engines, and sparse RL rewards at the trajectory level.
We present SearchEyes, which uses a typed knowledge graph as the backbone of a simulated search world that unifies all three components. We propose Perception-Knowledge Chains (PKC) to sample constrained multi-hop paths over the visual-knowledge intersection of Wikidata5M, retaining hop-level entity metadata that simultaneously defines a self-contained search world and step-level reward anchors.
Furthermore, we introduce Hop-Anchored Policy Optimization (HaPO), which reuses these anchors for step-level credit assignment without a separately trained process reward model. Experiments on six multimodal knowledge-intensive benchmarks show that SearchEyes achieves state-of-the-art performance among open-source multimodal search agents, with SearchEyes-27B improving over the strongest open-source baseline by an average of 6.2 points.
Blogger's Review: SearchEyes effectively addresses the structural disconnect in multimodal search by integrating training data, environments, and reward signals, providing a fresh perspective for multi-hop reasoning. This innovative approach not only enhances performance but also offers a significant reference framework for future research.