Video-DeepResearch: Towards the Next-Generation Multimodal Deepresearch Agent
AI Summary
The paper extends multimodal deep research from static images to continuous video, identifying two bottlenecks: modality bias toward textual search and parametric knowledge leakage from internal memory. It introduces a decoupled perception-exploration pipeline, stage-wise tool unlocking, and a two-stage SFT plus GRPO training recipe. The authors also release Video-DR-Bench, a human-AI collaborative benchmark with 200 complex multi-hop video VQA instances. Their Video-DeepResearch-35B-A3B reports 64.0% average accuracy, above the reported scores for Claude-4.5-Sonnet, GPT-5, and Gemini 2.5 Pro.
Why it's worth reading
Video research agents are moving from clip understanding to cross-frame evidence gathering and web verification. This work is timely because it combines a training recipe, a dedicated benchmark, and released code, making both the claims and evaluation design available for scrutiny.
Deep Read
1. What happened
Original fact: The paper introduces Video-DeepResearch (Video-DR), extending multimodal deep-research agents from static images to continuous video, and presents Video-DR-Bench. The supplied metadata lists an arXiv publication date of August 4, 2026.
2. Core technology
Original fact: The system uses a decoupled perception-exploration pipeline with stage-wise tool unlocking. Agents are required to perform cross-frame visual grounding before web retrieval. Training combines supervised fine-tuning (SFT) with Group Relative Policy Optimization (GRPO). Analysis: The design separates visual evidence extraction from open-web exploration, aiming to reduce direct reliance on textual search and latent parametric knowledge.
3. Key evidence and numbers
Original fact: Video-DR-Bench contains 200 complex multi-hop video VQA instances. The paper reports 64.0% average accuracy for Video-DeepResearch-35B-A3B, compared with 59.0% for Claude-4.5-Sonnet, 52.5% for GPT-5, and 57.5% for Gemini 2.5 Pro. The 30B-A3B variant reports 59.3%. Caveat: These figures are taken from the supplied abstract and have not been independently verified here.
4. Why it matters
Analysis: Video research requires temporal localization, spatial grounding, evidence aggregation, and web retrieval at the same time. By naming modality bias and parametric knowledge leakage as explicit failure modes, the paper distinguishes answering from executing an auditable research procedure.
5. Practical impact
Original fact: The authors provide code at https://github.com/Osilly/Vision-DeepResearch. If the repository includes sufficient data processing, evaluation, and tool-interface details, researchers may be able to reproduce the perception, retrieval orchestration, and GRPO components. Analysis: The staged design could be useful for video monitoring, long-form media analysis, documentary fact-checking, and multimodal search where the reasoning trace must expose which visual evidence preceded web lookup.
6. Limitations and uncertainty
Original fact: The abstract characterizes the evaluations as preliminary, and the benchmark has 200 instances. It does not specify the item distribution, video durations, retrieval-source controls, human-rating protocol, statistical significance, or component ablations. Analysis: A relatively small, potentially author-curated benchmark may not represent open-world video research. Prompting, tool access, context limits, and inference budgets can also materially affect comparisons with proprietary systems. Unverified inference: Whether the reported advantage of the 35B-A3B model persists across video domains, retrieval environments, and random seeds requires the full paper and independent replication.