OneSearch-VL: Unified Multimodal Deep Research Agent for Image and Video

This work tackles the problem of keeping track of how visual parts, object relationships, source‑based facts, and answer‑producing steps depend on each other in tasks that use a single image, several images, or video. We introduce a single multimodal agent system called **OneSearch‑VL** that makes visual grounding, external retrieval, and fact composition easier. At the heart of the system is the **Visually Grounded Evidence Graph (VGEG)**. VGEG acts as a common reference for building data, supervising processes, and evaluating individual operations. The framework uses a VGEG‑based data engine to create and verify multi‑image and video questions. It also filters expert solution paths, producing dedicated datasets for supervised fine‑tuning (SFT) and

reinforcement learning (RL). In addition, we define an **Evidence‑aware Visual‑Grounded Rubric (EVGR)** reward, derived from VGEG annotations, to guide the model in tracking evidence and staying visually grounded during reinforcement learning. For thorough testing, we set up special benchmarks that group questions by the types of operations they require. Experiments show that this unified agent outperforms baseline models with tool access on many image and video benchmarks, demonstrating a strong method for advanced multimodal reasoning and retrieval tasks.

Message: Hi FE, I need help with “OneSearch-VL: Unified Multimodal Deep Research Agent for Image and Video” in “Electronics & Communication Engineering”