pretty_name: AD-MCQ
arxiv: 2608.01755
task_categories:
- visual-question-answering
- video-classification
language:
- en
AD-MCQ
AD-MCQ is a candidate-trajectory multiple-choice dataset for autonomous-driving visual reasoning.
Paper
Deferred Exposure of Future Trajectories for Verifiable Reasoning in Autonomous Driving VLMs
Splits
| Split | File | Examples |
|---|---|---|
| train | train/data.jsonl |
5,000 |
| test | test/data.jsonl |
500 |
Each example contains the prompt, candidate-grounded supervision, rule-based reward metadata, and three camera views.
Media availability
The repository includes locally packaged MP4 media for 3,497 training examples and 250 test examples. Their video references use repository-relative paths under media/.
The remaining examples retain internal asset descriptors (asset_id, cam_id, and timestamps) but do not include the corresponding H265 media. These examples require access to the original asset store before they can be decoded. They are retained so that the official split sizes and annotations remain intact.
Codebook
The codebook/ directory contains the K=8192 trajectory codebook, waypoint representation, and summary metrics used to construct candidate trajectories.
Citation
@article{huang2026deferred,
title={Deferred Exposure of Future Trajectories for Verifiable Reasoning in Autonomous Driving VLMs},
author={Huang, Zixuan and Zhou, Yang and Wang, Kaixuan and Zhang, Guli and Xie, Hongyan and Zhu, Yakun and Geng, Hao and Ban, Yikun and Wang, Deqing},
journal={arXiv preprint arXiv:2608.01755},
year={2026}
}