In reported tests, AdaVDR-8B reached 38.40% overall accuracy in the Agentic VDR-EE evaluation, compared with 28.40% for the Qwen3-VL-8B-Instruct baseline. The reported difference was 10.00 percentage points.
The study examines whether AdaVDR can choose tools according to the task and the model's capabilities, and backtrack when intermediate evidence is unreliable.
VDR-EE contains 250 questions across seven domains, and each question requires both video evidence and external knowledge.
The benchmark gains varied by question type
The benchmark included 153 entity questions and 97 event questions. Reported gains were 9.15 percentage points for entity questions and 11.34 points for event questions.
In the long-event subgroup, AdaVDR's reported accuracy was 39.29% versus 17.86% for the baseline, a 21.43-point difference.
In a separate comparison, AdaVDR-9B recorded 47.60% overall VDR-EE accuracy, versus 40.00% for Qwen3.5-9B. The reported overall difference was 7.60 points, while the event-question gain was 15.47 points.
The authors report that AdaVDR had the best VDR-EE performance among the evaluated open-source models and substantially exceeded its corresponding base models on VideoDR.
Across baseline models on VDR-EE, the Agentic setting had higher reported average accuracy than Direct by 16.00 to 30.40 percentage points. For Qwen3.5-35B-A3B, accuracy was 16.00% in Direct and 46.40% in Agentic.
How the system was trained
AdaVDR's data-construction pipeline had two stages, QA generation and trajectory generation. During supervised fine-tuning, or SFT, approximately 2.9K verified trajectories were used to initialize adaptive tool use and reflection.
The reinforcement-learning stage started from the SFT checkpoint and used GRPO on approximately 1K training instances. Its adaptive reward gave no penalty to the most efficient correct trajectory, penalized each additional tool call, and was not applied to incorrect trajectories.
A second benchmark showed different results by difficulty
On VideoDR, the version based on Qwen3-VL-8B reached 51.00% average accuracy, while the version based on Qwen3.5-9B reached 56.00%.
By reported difficulty, the Qwen3-VL-8B-based version scored 65.62% on the low-difficulty subset, 44.44% on the medium subset and 43.75% on the high subset. The corresponding figures for the Qwen3.5-9B-based version were 75.00%, 50.00% and 43.75%.
Accuracy was not tied to every extra tool call
Tool-use traces show that Web Search averaged 2.50 calls per VideoDR question and 2.77 calls per VDR-EE question. Within VDR-EE, event questions averaged 3.13 calls, compared with 2.54 for entity questions.
The figures do not show a simple rule that more calls always produce a better answer. In an ablation, Base Data produced 45.00% accuracy with 6.18 average calls per question; Reflection plus Adaptive Data produced 48.00% with 7.26 calls; and the Accuracy plus Format plus Adaptive configuration reached 51.00% with 6.80 calls, compared with 7.84 calls for Accuracy plus Format.
What the scores can show
Accuracy was defined as the proportion of predictions that GPT-5.4 judged semantically consistent with the reference answer. The percentages therefore depend on that semantic-judging method.
The reported comparisons do not include confidence intervals, standard errors, p-values or significance tests.
These are benchmark results for the evaluated models and settings, rather than evidence of performance for human users or real-world deployment. Whether the approach generalizes beyond VDR-EE and VideoDR remains an open question.
The work is an arXiv version-1 preprint dated 26 Aug 2026. Its abstract provides a project-page URL and a GitHub repository URL for AdaVDR.
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