FACT: Failure-Aware Causal Training for World-Action Models
Action-conditioned world-action model that also learns from failed rollouts, predicting future video and task progress for the action actually taken; progress scores can rank sampled actions.以实际执行动作为条件的世界动作模型，把失败轨迹也作为监督，预测未来视频与任务进度；进度分可在推理时为候选动作排序。
BibTeX
@misc{peng2026fact,
title = {{FACT: Failure-Aware Causal Training for World-Action Models}},
author = {Quanquan Peng and Yutong Liang and Rui Yan and Nicklas Hansen and Xiaolong Wang},
year = {2026},
eprint = {2608.10232},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2608.10232},
}