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

arXiv 2026 49.1 method

TLDR

Introduces OneSearch-VL, a unified multimodal deep research agent using Visually Grounded Evidence Graph for image/video research tasks, with data engine, reward model, benchmarks, and substantial gains.

Reasoning

The paper presents a strong unified framework with a novel evidence graph representation, data construction pipelines, and new benchmarks, showing large empirical gains on multimodal research tasks. However, its scope is confined to image/video deep research rather than broader scientific discovery or autonomous experimentation, and evaluation is benchmark-based rather than real-world deployment.

Read-first score

Read-first score 49.1, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 31.

Recency 8%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Methodology quality 25%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=benchmark,evaluation,experiment

Reproducibility 25%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=code,github,repository

Topical relevance 42%

25.8

Uses existing LLM keyword relevance scores normalized to 0-100. AI scientist,automated scientific discovery,autonomous research agent,automated research,literature review agent,survey generation,automated experimentation,experiment design agent,AI for scientific research,paper writing agent,research automation,scientific discovery agent

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 110.

Keyword Scores

automated research

research automation

autonomous research agent

AI for scientific research

AI scientist

automated scientific discovery

literature review agent

survey generation

automated experimentation

experiment design agent

paper writing agent

scientific discovery agent

Deep Analysis

Innovations

Proposes OneSearch-VL, a unified multimodal deep research agent for single-image, multi-image, and video tasks.
Introduces the Visually Grounded Evidence Graph (VGEG), which encodes dependencies among localized visual anchors, entity relations, source-supported facts, and answer-producing operations as a shared task-level reference.
Develops a VGEG-based data engine that constructs and verifies multi-image and video questions and filters expert trajectories, yielding OneSearch-VL-SFT-110K and OneSearch-VL-RL-10K.
Derives the Evidence-aware Visual-Grounded Rubric reward (EVGR) from VGEG annotations to supervise evidence traceability and visual grounding during reinforcement learning.
Constructs OneSearch-MI-Bench and OneSearch-Video-Bench, organizing questions by the research operations encoded in their VGEGs.

Methodology

OneSearch-VL centers on the Visually Grounded Evidence Graph (VGEG) as a shared task-level reference for data construction, process supervision, and operation-level evaluation. A VGEG-based data engine constructs and verifies multi-image and video questions and filters expert trajectories, producing OneSearch-VL-SFT-110K for supervised fine-tuning and OneSearch-VL-RL-10K for reinforcement learning. During RL, the Evidence-aware Visual-Grounded Rubric reward (EVGR) supervises evidence traceability and visual grounding; evaluation compares OneSearch-VL-8B against Qwen3-VL-8B with tool access on two new benchmarks, 7 image benchmarks, and VideoDR.

Key Results

OneSearch-VL-8B improves over Qwen3-VL-8B with tool access by 20.2 percentage points on OneSearch-MI-Bench and 17.6 percentage points on OneSearch-Video-Bench, while also achieving substantial gains across 7 image benchmarks and VideoDR.

Resources

Paper