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.
VGEG-Centered Data Engine
Our VGEG-based data engine constructs and verifies multi-image and video questions and filters expert trajectories. Starting from raw visual sources, the pipeline identifies localized visual anchors, connects them to source-supported web facts, constructs task-level VGEGs, and synthesizes verified question–answer pairs and expert trajectories. VGEG links localized visual anchors to real-world entities, records multi-hop relational paths and source-supported facts, and specifies the operations that compose these facts into an answer. Based on VGEG, the data engine extracts visual entities, retrieves and verifies web facts, generates compositional questions, and synthesizes expert tool-use trajectories while retaining visual locations, source provenance, and evidence dependencies. These retained dependencies support question-answer construction and verification, expert-trajectory filtering, fine-grained trajectory rewards for RL, and operation-level evaluation.
The pipeline begins with visual source curation, reducing an initial pool of 2.5M web videos to 70k candidates through metadata screening, task-suitability assessment, category balancing, and duration balancing. Dense visual anchor discovery then converts each retained video into a hierarchical visual representation, sampling and grouping frames into local clips, generating clip- and frame-level descriptions, removing redundant key frames, aggregating temporally related clips into events, and localizing searchable objects in representative frames. Web evidence graph construction connects visual anchors to external knowledge through image search, OCR, and text retrieval, verifying candidate identities against the visual context and retrieving source-supported relations and attributes. These results form an input-level evidence graph that preserves the correspondence among visual anchors, real-world entities, external facts, and supporting sources.
VGEG-based task construction derives task-level VGEGs from the input-level graph through a task-conditioned projection that selects required visual anchors and facts and augments them with explicit answer-producing operations. The generator jointly outputs questions, answers, and VGEGs, which are verified against selected evidence and rewritten with visually grounded references. Single-image tasks reuse existing data and convert Wikipedia entity-sampling paths into VGEGs.
Expert trajectory synthesis has an expert model interact with the tool environment using only the visual input, question, and available tools until producing a final answer or reaching the interaction budget. Retained trajectories pass both answer-correctness and process-quality evaluation and are combined with single-image trajectories in a unified multi-turn format to construct OneSearch-VL-SFT-110K. For RL, OneSearch-VL-RL-10K is constructed from task pools spanning the same three visual input types.
Evidence-aware Visual-Grounded Rubric Reward
We further derive the Evidence-aware Visual-Grounded Rubric reward (EVGR) from VGEG annotations to supervise evidence traceability and visual grounding during RL. EVGR evaluates complete trajectories along two complementary dimensions: evidence traceability, which verifies whether the required facts are supported by tool observations; and visual grounding, which verifies whether the relevant objects, regions, or frames are correctly identified and used in the solution. The rubric records the answer core, required claim facts, supporting sources and snippets, and applicable frame or region annotations. Two separate judge calls evaluate these dimensions, and their combination forms the trajectory-level process reward for policy optimization.
Operation-Oriented Benchmarks
For fine-grained evaluation, we construct OneSearch-MI-Bench and OneSearch-Video-Bench. These benchmarks organize questions according to the research operations encoded in their corresponding VGEGs, which capture the entity-relation paths and fact-composition dependencies required to derive the answer. This organization emphasizes how an agent retrieves and composes evidence rather than what the visual input contains. Both benchmarks use the same operation taxonomy, enabling consistent analysis across multi-image and video inputs. Six principal research operations are defined, covering entity lookup, relation tracing, conditional filtering, and fact composition across single-image, multi-image, and video inputs.
OneSearch-MI-Bench uses deduplicated evidence frames from the same visual source forming an unordered image set, with each example depending on at least two images. OneSearch-Video-Bench preserves temporal structure and links visual anchors to their events, key frames, and web facts, requiring agents to locate relevant video content before retrieving and composing external knowledge.
Experimental Results
OneSearch-VL-8B outperforms Qwen3-VL-8B with tool access by 20.2 and 17.6 percentage points on OneSearch-MI-Bench and OneSearch-Video-Bench, respectively, and by 27.0 points on VideoDR. On seven single-image benchmarks, OneSearch-VL-8B obtains an average score of 58.3, outperforming the same-scale Qwen3-VL-8B Agent by 16.3 points and OpenSearch-VL-8B by 1.7 points. It improves over OpenSearch-VL on all seven benchmarks, with gains of 2.8, 2.1, and 2.0 points on VDR, InfoSeek, and MMSearch, respectively, and achieves the best result among listed agentic methods on six of the seven benchmarks.
Joint SFT on all three visual input types achieves a six-benchmark average of 55.8, versus 51.7–55.1 for single-type training. Adding either multi-image or video trajectories to the single-image data raises the average score, and joint training on all three types gives the best result on SimpleVQA, OneSearch-MI-Bench, and OneSearch-Video-Bench. EVGR further raises the average from 57.3 to 61.1 over answer and query rewards alone.
From Dataset to Deployment
The OneSearch-VL project repository is available, and the constructed benchmarks and models provide a unified framework for multimodal deep research across single-image, multi-image, and video inputs. The VGEG-centered data engine and operation-oriented benchmarks offer new capabilities for evaluating evidence traceability and visual grounding, supporting the development of more capable and interpretable multimodal research agents.