GuideMe: Multi-Domain Task Guidance and Intervention in Streaming Video (ECCV 2026)
GuideMe is a benchmark for training and evaluating multimodal large language models as real-time procedural coaches. It covers next-step instruction, completion feedback, error detection, and corrective guidance across cooking, daily-life tasks, object manipulation, and fitness.
- Project: https://fawnliu.github.io/project/guideme/
- Code: https://github.com/fawnliu/GuideMe
- Paper: https://arxiv.org/abs/2607.02991
@inproceedings{liu2026guideme,
title={GuideMe: Multi-domain Task Guidance and Intervention in Streaming Video},
author={Liu, Fang and Chen, Jinpeng and Xu, Ke and Liu, Yuhao and Guan, Huankang and Lu, Xudong and Bo, Yang and Hancke, Gerhard and Liu, Rui and Lau, Rynson WH},
booktitle={European Conference on Computer Vision},
pages={57--74},
year={2026},
organization={Springer}
}