AI

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

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 09/10/2026, 04:18:24 UTC.

Briefings

OneSearch-VL introduces a unified multimodal agent framework using the Visually Grounded Evidence Graph (VGEG) to manage visual grounding, external retrieval, and fact composition. The system constructs and verifies multi-image and video questions, assembling SFT and RL models with 110K and 10K parameters respectively. It derives the Evidence-aware Visual-Grounded Rubric reward (EVGR) to supervise evidence traceability during RL. Experiments show OneSearch-VL-8B outperforms Qwen3-VL-8B by 20.2 and 17.6 percentage points on new benchmarks while improving performance across 7 image benchmarks and VideoDR.

Key points

• OneSearch-VL introduces the Visually Grounded Evidence Graph (VGEG) to encode dependencies linking visual anchors, entity relations, and source facts.
• OneSearch-VL-8B improves over Qwen3-VL-8B by 20.2 percentage points on the OneSearch-MI-Bench benchmark.

Context

This research addresses the challenge of unifying visual operations for single-image, multi-image, and video deep research. The VGEG framework provides a shared task-level reference for data construction, process supervision, and operation-level evaluation in multimodal agents.

Limitations

The study relies on internal benchmark metrics rather than external validation. Results are limited to the specific benchmarks tested and do not generalize to other domains or real-world applications without further testing.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 14/11/2026, 11:04:40 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeickbtqvjpv3rkskegnbodkliv574dw3usrifskzpc2ehoirwjyvzi