Glonce

Video-DeepResearch beats Claude Sonnet 4.5 by 5 points on new video benchmark

researchThe Glonce editorsHugging Face🔥55Glonce rating55Impact60Novelty65Relevance95Surprise55Players50Credibility85
Video-DeepResearch beats Claude Sonnet 4.5 by 5 points on new video benchmark

Multimodal agents built for deep research, combining visual tools with open-web search to answer complex questions, have mostly been developed and tested on static images. Video adds a much harder requirement: an agent has to track and ground what happens across a continuous stream of frames while still exploring the open web. The researchers behind Video-DeepResearch (Video-DR) ran preliminary evaluations of current models on video tasks and found two recurring bottlenecks. The first is modality bias: agents tend to bypass the visual tools available to them and fall back on textual search instead. The second is parametric knowledge leakage: models lean on knowledge they already memorized internally rather than genuinely using tools to look things up.

Video-DR is built to close both gaps. Its perception-exploration pipeline is decoupled and unlocks tools stage by stage, which forces the agent to exhaustively ground what it sees across video frames before it is allowed to search the web. The system is trained in two stages: supervised fine-tuning first, then Group Relative Policy Optimization (GRPO), a reinforcement learning method the researchers say lets the agent explore on its own and break past the ceiling that pure imitation learning imposes.

To measure the result, the researchers built Video-DR-Bench, a human-AI collaborative benchmark of 200 complex, multi-hop video question-answering instances designed for this kind of task.

On that benchmark, Video-DeepResearch-35B-A3B reaches 64.0% average accuracy, which the researchers describe as a new state of the art. That is 5.0 points ahead of the proprietary Claude Sonnet 4.5 (59.0%), and it clears GPT-5 (52.5%) and Gemini 2.5 Pro (57.5%) by even wider margins, 11.5 and 6.5 points respectively. A smaller 30B-A3B variant reaches 59.3%, which the researchers call competitive with Claude Sonnet 4.5 and take as evidence that their training approach still works at a more compact scale.

The researchers have published code for Video-DR on GitHub. The abstract does not name individual authors or institutions behind the work, does not give a submission or publication date, and does not say whether the trained model weights or the Video-DR-Bench dataset itself will be released publicly, only that code is available.

Key facts

  • Video-DeepResearch (Video-DR) extends multimodal deep-research agents from static images to continuous video, built to fix two bottlenecks the researchers found in current models: modality bias, where agents skip visual tools in favor of text search, and parametric knowledge leakage, where models answer from memorized knowledge instead of genuine tool use.
  • Its decoupled perception-exploration pipeline uses stage-wise tool unlocking to force exhaustive cross-frame visual grounding before web retrieval is allowed, trained through supervised fine-tuning followed by Group Relative Policy Optimization (GRPO).
  • On the researchers' new Video-DR-Bench, a human-AI collaborative benchmark of 200 multi-hop video question-answering instances, Video-DeepResearch-35B-A3B scores 64.0% average accuracy, a new state of the art.
  • That is 5.0 points ahead of the proprietary Claude Sonnet 4.5 (59.0%), 11.5 points ahead of GPT-5 (52.5%), and 6.5 points ahead of Gemini 2.5 Pro (57.5%).
  • A smaller 30B-A3B variant reaches 59.3%, which the researchers call competitive with Claude Sonnet 4.5; code is public on GitHub, though the abstract does not say whether model weights or the benchmark dataset itself will be released.

Why it matters

Most deep-research agents, systems that combine visual tools with web search to answer complex questions, have been built and tested on static images. Video adds a much harder requirement: an agent has to track and ground what happens across a continuous stream of frames as well as search the open web. The researchers' preliminary tests found that current models tend to sidestep this difficulty rather than solve it, either by leaning on text search instead of actually watching the video (modality bias) or by answering from memorized knowledge instead of using tools at all (parametric knowledge leakage). Video-DR is built specifically to close both gaps, and the result, on the researchers' own new benchmark, is a comparatively compact model beating three proprietary frontier systems, Claude Sonnet 4.5, GPT-5 and Gemini 2.5 Pro, at video-based deep research.

Who it affects

Researchers and engineers building multimodal or video-capable agents, especially deep-research style systems that pair visual tool use with open-web search, gain both a training recipe, stage-wise tool unlocking plus supervised fine-tuning then GRPO, and a fresh benchmark to measure against. Anyone comparing how well proprietary frontier models like Claude Sonnet 4.5, GPT-5 and Gemini 2.5 Pro actually use tools on video tasks, rather than answering from memory, now has a specific, if newly introduced, benchmark to point to.

How to use it

There is no consumer product here: this is a research paper describing a training method, two resulting model variants, Video-DeepResearch-35B-A3B and a smaller 30B-A3B variant, and a new benchmark, Video-DR-Bench. The researchers have published code for Video-DR on GitHub. The abstract does not say whether the trained model weights or the Video-DR-Bench dataset itself will be released publicly, or under what license, so hands-on access beyond the code repository is not established by this description.

How solid is it

This is a paper posted to Hugging Face's Papers page, and the source available here does not name individual authors or institutions behind the work. The headline result, that Video-DeepResearch-35B-A3B beats Claude Sonnet 4.5, GPT-5 and Gemini 2.5 Pro, is measured on Video-DR-Bench, a benchmark the same research team built for this paper, so there is no independent, third-party evaluation to check it against yet. The benchmark itself covers 200 multi-hop video question-answering instances, described as human-AI collaborative in construction, though the abstract does not detail how that collaboration worked. The two bottlenecks the paper diagnoses in existing models, modality bias and parametric knowledge leakage, are also the authors' own framing rather than an independently audited finding.

Risks and caveats

The core comparison rests on Video-DR-Bench, a benchmark built by the same team that built Video-DR, so there is no independent replication reported here. The benchmark itself is fairly small, 200 questions, which limits how far the result generalizes to open-ended, real-world video tasks. The abstract gives no training data volume, compute budget, or submission date beyond the '35B-A3B' and '30B-A3B' labels in the model names, and it does not say whether the model weights or the benchmark dataset will be released, only that code is public. Comparisons against Claude Sonnet 4.5, GPT-5 and Gemini 2.5 Pro used whatever access and prompting setup the researchers chose for their own tests, details the abstract does not spell out.