A newly posted paper on Hugging Face presents OneSearch-VL, a unified multimodal agent designed for deep research over single images, multiple images, and video. According to the authors, these tasks demand different visual operations yet share a common workflow: visual grounding, external retrieval, and fact composition.
The paper identifies a central challenge in preserving the dependencies that link localized visual anchors, entity relations, source-supported facts, and the operations that produce answers. To address this, OneSearch-VL is centered on a Visually Grounded Evidence Graph (VGEG), which encodes those dependencies as a shared task-level reference used for data construction, process supervision, and operation-level evaluation.
Using a VGEG-based data engine, the authors construct and verify multi-image and video questions and filter expert trajectories. This yields OneSearch-VL-SFT-110K and OneSearch-VL-RL-10K for supervised fine-tuning and reinforcement learning respectively. The team also derives an Evidence-aware Visual-Grounded Rubric reward (EVGR) from VGEG annotations to supervise evidence traceability and visual grounding during RL.
For evaluation, the paper introduces OneSearch-MI-Bench and OneSearch-Video-Bench, which organize questions by the research operations encoded in their VGEGs. Experiments report that OneSearch-VL-8B improves over Qwen3-VL-8B with tool access by 20.2 and 17.6 percentage points on the two new benchmarks, while also achieving substantial gains across seven image benchmarks and VideoDR.
The project repository and code are available at github.com/appletea233/OneSearch-VL, with data hosted at huggingface.co/OneSearch-VL.