| title | SWAM: Scaling World Action Models with Cross-Modal Alignment |
|---|---|
| authors | Yu Zhou, Xinyuan Xia, Xuankun Yang, Yinghao Yang, Wendi Chen, Bowen Liu, Jianhan Ma, Weicheng Duan, Jun Lv, Cewu Lu, Yanwei Li |
| date | 2026-11-09 |
| venue | arXiv (preprint, not peer reviewed) |
| id | arXiv:2611.12870 |
| affiliation | Shanghai Jiao Tong University |
| upvotes | 20 (Hugging Face Daily Papers, 2026-11-12) |
| links | abs pdf hf code |
| cite | |
| related | ARIS: Autonomous Research via Adversarial Multi-Agent Collaboration SWE-Explore: Benchmarking How Coding Agents Explore Repositories Work Agents: Systems, Capabilities, and Optimization Are We Ready For An Agent-Native Memory System? SoftwareGym: Scaling Computer-Use Agent Training with Diverse Synthesized Software Environments |
| abstract | World action models jointly predict video and actions conditioned on language. Existing studies focus primarily on architectures and training mechanisms, typically treating language-video alignment acquired through video pretraining as an established foundation. However, introducing action learning may impair this alignment. Principled guidance on data and supervision design for maintaining language-video alignment during action learning remains limited. To investigate this, we curate a pool of approximately 47,000 hours of general videos, egocentric videos, UMI demonstrations, and robot demonstrations. Building on this, we systematically study how language-video supervision, language-video-action supervision, and the joint use of these two forms affect WAM performance. In particular, we examine how language annotation quality affects pretrained language-video capabilities, and how action supervision from UMI and robot demonstrations extends these capabilities to language-video-action modeling. To better scale the model, we further conduct detailed studies on joint training recipe and model architecture. Under a fixed training budget, mixing egocentric videos and UMI demonstrations improves control generalization. Model architecture affects how effectively these experiences are used, while scaling pretraining under the selected configuration brings further gains. The resulting recipe, post-trained on RoboTwin Clean demonstrations only, achieves success rates of 88.10% and 61.56% under Easy (Clean-to-Clean) and Hard (Clean-to-Random) conditions, respectively. Our model achieves 40.00% SR across three real-world tasks, compared with 23.33% for the strongest baseline, indicating better generalization under unseen conditions. We release the curated data, training code, and model weights. |
bibtex
@misc{zhou2026swam,
title={SWAM: Scaling World Action Models with Cross-Modal Alignment},
author={Yu Zhou and Xinyuan Xia and Xuankun Yang and Yinghao Yang and Wendi Chen and Bowen Liu and Jianhan Ma and Weicheng Duan and Jun Lv and Cewu Lu and Yanwei Li},
year={2026},
eprint={2611.12870},
archivePrefix={arXiv},
url={https://arxiv.org/abs/2611.12870}
}