AdaVDR is an adaptive agent for video that decides which tools to invoke based on the task and model capabilities, and only backtracks when intermediate results are unreliable. The authors built a data pipeline that generates training trajectories with task-specific tool-use strategies, plus a model-conditioned filter that removes tools the model can bypass. Supervised fine-tuning and with a redundancy-aware reward produce strong results on a new benchmark, VDR-EE, and improve over base models on VideoDR.
Adaptive tool use improves video research agents’ accuracy and efficiency
AdaVDR learns when to ground video content, search the Web, or backtrack rather than following a fixed tool-use workflow.
Chinese Tech
Xintong Zhang · Xiaomeng Fan · Shilin Yan · Ekko He · Zicheng Liu · Zijian Zou · +4 more
Accio Team, Alibaba Group · Beijing Key Laboratory of Intelligent Information Technology · Beijing Institute of Technology
Research Digest··2 min read
Thread:Skill Selection for Agents
In one line
AdaVDR adapts tool use and reflection in video deep research to cut costs and errors.
Before you read this
Three ideas this paper builds ondeep researchA multi-step AI system that autonomously gathers, evaluates, and synthesizes evidence from diverse sources to produce comprehensive reports.21 papers →vision-language modelA model that handles tasks combining visual and textual information, such as editing images from conversation or reasoning across images and text.216 papers →reinforcement learningA machine learning method where an agent learns to make sequences of decisions by receiving rewards or penalties for its actions.239 papers →
Read it as
Zhang et al. developed a video question-answering agent that combines video understanding with external retrieval while adapting its tools to each task.
The authors built a data pipeline that identifies retrieval-relevant events and entities in diverse videos, gathers supporting information through video grounding and Web retrieval, and turns that evidence into question-answer pairs with task-specific tool-use trajectories.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.