An adaptive AI agent for video research reported higher accuracy than its base model on two benchmark tasks, with the paper’s reported average gap reaching 29 percentage points on VideoDR and 15.2 points on VDR-EE. The result comes from an arXiv preprint, so it is a reported model comparison rather than evidence that the approach will work equally well in real-world research.
AdaVDR is designed for questions that require both understanding what appears in a video and finding related information on the open Web. It chooses tools according to the question type, the model’s video-understanding capability and its internal knowledge, then backtracks when an intermediate piece of evidence looks unreliable.
Building a selective research loop
To test that setup, the researchers used VDR-EE, a benchmark of 250 manually verified questions requiring both video and external evidence. Its sample contains 153 entity questions and 97 event questions.
The training pipeline generated task-specific interaction traces, retried unreliable steps, kept those correction traces as reflection data and filtered out tool calls or chains the model did not need. Supervised fine-tuning first used 2.9K verified trajectories. The later NASA-GRPO stage, a reinforcement-learning step, assigned rewards at the tool level by combining the contribution of evidence segments with whether a tool was necessary, while normalizing rewards by tool type.
The necessity test was model-conditioned, meaning it depended on what this particular model could already answer. A tool or consecutive tool chain was marked necessary when bypassing it changed the predicted result compared with the result obtained after execution; if the predictions matched, the step was treated as removable.
Where the score moved
In the supervised fine-tuning comparison, AdaVDR-9B-SFT was reported at 52% on VideoDR versus 37% for the Qwen3.5-9B base model, and at 48% versus 40% on VDR-EE. With the full agentic setup, the reported average-accuracy gains over the base model were 29 percentage points on VideoDR and 15.2 points on VDR-EE.
The necessity-weight ablation also reported fewer interaction turns. The weighted variant was reported at 66% on VideoDR and 55.2% on VDR-EE. It averaged 11.67 turns rather than 13.67 on VideoDR, and 13.26 rather than 15.43 on VDR-EE, compared with the variant without the weight.
The process measures pointed in the same direction. For VideoDR visual segments, NASA-GRPO produced 76.1% useful segments, compared with 58.4% for SFT. Among useful segments, 84.6% were classified as necessary under NASA-GRPO versus 68.9% under SFT, while false acceptance was 12.8% versus 27.6%.
Another analysis reported higher shares of necessary segments and calls when the necessity weight was used. On VideoDR, necessary segments accounted for 65.9% versus 54.9% without the weight, and necessary calls for 76.2% versus 54.6%. On VDR-EE, the corresponding figures were 65.9% versus 50.8% for segments and 74.5% versus 53.2% for calls.
Removing either retrieval option was associated with lower benchmark scores. With all tools available, accuracy was 66% on VideoDR and 55.2% on VDR-EE. Disabling image search gave 58% and 48.8%, respectively, while disabling web search gave 53% and 44%.
What the benchmark leaves open
The numbers are still benchmark comparisons, not a randomized test of the agent in use. The supplied analysis reports no human control group, no confidence intervals and no significance tests, so the size and reliability of the gaps cannot be assessed as they would be in a fuller statistical study.
Evaluation also depended on automated judgment: answer accuracy was judged by GPT-5.4, while segment analyses used a VLM judge. VDR-EE was constructed and manually verified within the study, and the supplied analysis did not report independent benchmark validation. The findings therefore do not establish real-world web-research reliability, deployment cost or user utility, or show that they generalize to unseen models, tools and domains.
The document is an arXiv preprint labeled arXiv:2608.25559v2 and dated 14 Oct 2026. No funding statement is reported in the supplied text; its front matter lists Accio Team, Alibaba Group as an affiliation.
Paper data and sources
Original title: AdaVDR: Adaptive Tool Use and Reflection for Video Deep Research
Authors: Xintong Zhang, Xiaomeng Fan, Shilin Yan et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text