OneSearch-VL: A Unified Multimodal Deep Research Agent for Images and Videos (arXiv 2610.12419)

Forum topic · 小凯 · 2026-10-11

Summary

OneSearch-VL is a unified multimodal deep research agent covering single-image, multi-image, and video research tasks, built around a Visually Grounded Evidence Graph (VGEG) that encodes dependencies between visual anchors, entity relations, source-supported facts, and answer-producing operations. The authors built a VGEG-based data engine that 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. From VGEG annotations they derive the Evidence-aware Visual Grounding score Reward (EVGR) to supervise evidence traceability during RL. For evaluation they introduce OneSearch-MI-Bench and OneSearch-Video-Bench, organized by VGEG-encoded research operations. Experiments show OneSearch-VL-8B outperforms Qwen3-VL-8B with tool access by 20.2 and 17.6 percentage points on the two new benchmarks, with notable gains on seven image benchmarks and VideoDR. Code: https://github.com/appletea233/OneSearch-VL, paper: https://arxiv.org/abs/2610.12419

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

Field: Computer Vision (CV) arXiv: 2610.12419 Repository: https://github.com/appletea233/OneSearch-VL

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.

The authors 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.

Key Contributions

VGEG-based data engine: constructs and verifies multi-image and video questions, and filters expert trajectories.
Training datasets: assembled OneSearch-VL-SFT-110K for SFT and OneSearch-VL-RL-10K for RL.
EVGR reward: an Evidence-aware Visual Grounding score Reward derived from VGEG annotations, supervising evidence traceability and visual grounding during RL.
New benchmarks: OneSearch-MI-Bench and OneSearch-Video-Bench, organized by VGEG-encoded research operations for fine-grained evaluation.

Results

OneSearch-VL-8B outperforms Qwen3-VL-8B with tool access by 20.2 and 17.6 percentage points on the two new benchmarks, while also achieving significant improvements across 7 image benchmarks and VideoDR.

Authors

Hongyu 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

*Source: forum post, auto-collected 2026-10-11.*