OneSearch-VL: Unified Multimodal Deep Research Agent for Image and Video
AuthorsHongyu Li, Manyuan Zhang, Kaituo Feng, Shu Chen, Dian Zheng, Hao Li, Hao Yu, Zhangquan Chen, Zoey Guo, Ray Zhang, Shaofei Huang, Tianrui Hui, Linjiang Huang, Si Liu
Affiliations[
OneSearch-VL uses an evidence graph to help a single agent visually ground, retrieve, and synthesize reliable answers across images and videos.
Key results
Percentage-point gain over the Qwen3-VL-8B agent.
Percentage-point gain over the Qwen3-VL-8B agent.
Percentage-point gain over the Qwen3-VL-8B agent.
Average rises from 57.3 when both evidence traceability and visual grounding rewards are used.
Expert trajectories used for supervised fine-tuning.
Tasks used for reinforcement learning.
What the paper found
OneSearch-VL unifies deep research across single images, image collections, and video with a shared tool-using policy. Its core contribution is the Visually Grounded Evidence Graph, or VGEG, which connects localized visual anchors to identified entities, source-supported facts, and the operations needed to compose an answer. The graph supports verified task construction, trajectory filtering, and operation-level benchmarks; the training pipeline produces 110K supervised trajectories and 10K reinforcement-learning tasks. Starting from Qwen3-VL-8B, the agent uses Group Relative Policy Optimization with an Evidence-aware Visual-Grounded Rubric reward, or EVGR, that scores both evidence traceability and visual grounding. On the new benchmarks, it improves over the Qwen3-VL-8B agent by 20.2 percentage points on OneSearch-MI-Bench and 17.6 points on OneSearch-Video-Bench, and gains 27.0 points on VideoDR. Across six benchmarks, adding both EVGR dimensions raises the average from 57.3 to 61.1. The pipeline also uses ByteDance Seed 2.0 Pro to help construct data and GPT-4o to judge answer correctness, linking visual recognition, web retrieval, and evidence-based answer composition in one research agent.
Original abstract
Single-image, multi-image, and video deep research require different visual operations but share a workflow of visual grounding, external retrieval, and fact composition. A key challenge is to preserve the dependencies linking localized visual anchors, entity relations, source-supported facts, and answer-producing operations. We introduce OneSearch-VL, a unified agent centered on the Visually Grounded Evidence Graph (VGEG), which encodes these dependencies as a shared task-level reference for data construction, process supervision, and operation-level evaluation. Our VGEG-based data engine constructs and verifies multi-image and video questions and filters expert trajectories. Using these data, we assemble OneSearch-VL-SFT-110K and OneSearch-VL-RL-10K for SFT and RL, respectively. We further derive the Evidence-aware Visual-Grounded Rubric reward (EVGR) from VGEG annotations to supervise evidence traceability and visual grounding during RL. For fine-grained evaluation, we construct OneSearch-MI-Bench and OneSearch-Video-Bench, organizing questions by the research operations encoded in their VGEGs. Experiments show 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, respectively, while also achieving substantial gains across 7 image benchmarks and VideoDR. Project repository: https://github.com/appletea233/OneSearch-VL
Read the original paperMore in Multimodal AI
Browse all 64 papers →ALIVE: Interaction-Aligned Object Insertion for First-Frame-Guided Video Editing
Zhenghong Zhou, Zhe Lin, Jiebo Luo, Yuqian Zhou
ALIVE makes inserted video objects behave as if they truly belong in the scene by learning how they should interact with existing actions.
DEPICT: Scoring Text-to-Image Alignment by Answer Agreement
Vasco Ramos, Sandra Godinho Silva, Joao Magalhaes, Ricardo Rei, Pedro Henrique Martins
DEPICT improves text-to-image evaluation by checking whether an image and its caption lead to the same answers, especially for tricky cases like negation, missing objects, and incorrect attributes.
Omni-Embed-Mini: Binding Modalities Without Forgetting via Dense Distillation
Mohammed Irfan Kurpath, Jaseel Muhammad Kaithakkodan, Sahal Shaji Mullappilly, Ivan Laptev, Hisham Cholakkal
A compact embedding model unifies text, speech, audio, images, video, and documents in one search space without sacrificing the original text capabilities.