Video-DeepResearch¶

Development line: project:video-deepresearch · thread video-deepresearch
Last event: 2026-08-10 · 1 dated since 2026-08-10 · Researched: 2026-09-05 · confidence: high

What it is¶

Video-DeepResearch is an open research stack and two models for questions grounded in video objects and events before checking external sources.

• Keyframe selection: selects video frames.
• Crop search: inspects visual subregions.
• Text search: queries external sources.
• Evidence extraction: pulls structured proof from web pages.

Running the documented setup requires separate inference, judge, and extract services; raw checkpoints and API keys do not replace them. The stack fits reproducible evaluation and fine-tuning of video-grounded agents, not a standalone chat service.

Development line¶

• 2026-08-10 — Video-DeepResearch public project and source references were recorded. The video task splits from earlier image-centric Vision-DeepResearch and enforces visual grounding before web search.

What changed¶

2026-08-10 — Video-DeepResearch appeared as a project page and a section in the Vision-DeepResearch repository: the video task splits from earlier image-centric Vision-DeepResearch and enforces visual grounding before web search.

How to use this¶

As of 2026-08-10, evaluate Video-DeepResearch from the linked project page and source repository, and verify capabilities and release status directly.

1. Clone the repository and work inside the Video-DeepResearch subfolder; it contains separate preprocessing, evaluation, SFT, and RL modules. — https://github.com/Osilly/Vision-DeepResearch/blob/main/Video-DeepResearch/README.md
2. Extract keyframes from the video; the documentation recommends CLIP similarity filtering, with pixel-difference only as a rough fallback. — https://github.com/Osilly/Vision-DeepResearch/blob/main/Video-DeepResearch/README.md
3. Run a VLM for inference, a separate OpenAI-compatible judge, and the required extract server for Visit; the example uses Qwen3-VL-30B-A3B-Instruct for judge and extract. — https://github.com/Osilly/Vision-DeepResearch/blob/main/Video-DeepResearch/README.md
4. Launch an evaluation script for SGLang, vLLM, or OpenAI-compatible MaaS, and pick tool, direct, or both mode. — https://github.com/Osilly/Vision-DeepResearch/blob/main/Video-DeepResearch/README.md

Best practices¶

• Collect evidence from frames and crops before opening text search: this order keeps the agent from answering from memory or a single text query. — https://costaliya.github.io/Video-DeepResearch/
• Use CLIP cosine-similarity for keyframes; use pixel-difference only when CLIP is unavailable and account for rougher selection. — https://github.com/Osilly/Vision-DeepResearch/blob/main/Video-DeepResearch/README.md
• Keep the extract server running: without it, Visit returns raw HTML instead of structured evidence, and documentation warns that agentic accuracy drops sharply. — https://github.com/Osilly/Vision-DeepResearch/blob/main/Video-DeepResearch/README.md
• Compare tool and direct modes through an independent judge to separate search gains from answers based solely on keyframes. — https://github.com/Osilly/Vision-DeepResearch/blob/main/Video-DeepResearch/README.md

Superseded by this¶

• Nothing marked obsolete yet.

Still unknown¶

• Public materials confirm code, paper, and checkpoint announcements, but we did not run inference or verify the current availability, license, and hardware requirements of each checkpoint.
• VideoDR-Bench500 is listed as coming soon; we cannot treat it as released or use it for comparison.

Sources¶

Agent brief¶

• Subject: project:video-deepresearch, thread video-deepresearch, 1 dated events 2026-08-10 → 2026-08-10.
• Practical note: As of 2026-08-10, practitioners should use the linked project page and source repository as the starting points for evaluating Video-DeepResearch, while independently verifying its capabilities and release status.
• Confidence: high. Dated supersedes above are the authority for what is obsolete.