ModalSearcher defines autonomous cross‑modal search, allowing an agent to choose among direct reasoning, text‑to‑text search, text‑to‑image search, and reverse‑image search, and to introduce visual evidence when appropriate. To enable this, the authors build the ModalSearch‑Pipeline, which constructs an interleaved multimodal knowledge graph by sampling paths of varying depth and modality composition. Backward question construction obscures entity names and image targets to prevent reasoning shortcuts while keeping essential disambiguation clues, yielding text‑only, image‑initialized, or arbitrarily interleaved multi‑hop problems. The ModalSearcher agent is trained with supervised fine‑tuning and reinforcement learning in a live web environment, and the ModalSearch‑Bench benchmark suite is released. Experiments show that ModalSearcher‑8B improves average accuracy by 16.2 points over Qwen3‑VL‑8B across five multimodal knowledge‑intensive QA benchmarks, with consistent gains on independently constructed active visual‑search benchmarks.
Review: This study demonstrates a practical framework for autonomous multimodal evidence acquisition, advancing the capability of large multimodal models in complex real‑world queries.