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

OneSearch-VL. Unified Multimodal Deep Research Agent for Image and Video. By Hongyu Li and thirteen others, from Beihang University, the Chinese University of Hong Kong, and three other institutions. Published on the eighth of October.

Deep research over a single image, over several images, and over video calls for different visual operations. Underneath, though, all three follow the same workflow: ground something in the picture, retrieve information from outside, and compose facts into an answer. The hard part is keeping track of how those pieces depend on each other. A localized visual anchor links to entity relations, those link to facts supported by sources, and those link to the operations that finally produce the answer. Lose a link and the chain breaks.

OneSearch-VL is a unified agent built around one structure, the Visually Grounded Evidence Graph, or VGEG. The graph encodes those dependencies and serves as a shared, task-level reference for three things at once: building data, supervising the process, and evaluating individual operations.

A data engine based on the graph constructs and verifies multi-image and video questions, and filters expert trajectories. From that data come two training sets: OneSearch-VL-SFT-110K for supervised fine-tuning, and OneSearch-VL-RL-10K for reinforcement learning.

The graph annotations also yield a reward, the Evidence-aware Visual-Grounded Rubric reward, or EVGR. During reinforcement learning it supervises whether evidence can be traced and whether answers stay visually grounded. For evaluation, two new benchmarks, OneSearch-MI-Bench and OneSearch-Video-Bench, organize their questions by the research operations recorded in each graph.

Against Qwen3-VL-8B with tool access, OneSearch-VL-8B improves by twenty point two percentage points on the multi-image benchmark. On the video benchmark the improvement is seventeen point six points. The authors also report substantial gains across seven image benchmarks and on VideoDR.

A single graph does the work that would usually take three separate pieces of machinery. The structure that generates the questions is the same one that rewards the agent and the same one that grades it, so every stage of the pipeline agrees about what counts as grounded evidence.