Reproduce “PASA-KV: Attacking Action-Facing Visual Routes in World Action Models” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Zhongbo Wu, Jiakai Huang, Zheng Zhang, Siyu Xu, Kaixiang Chen, Xiaohong Liu, Shan You, Chang Xu and 1 more

PASA-KV probes action-facing visual routes in world action models (WAMs) before training an image patch against selected routes. Its Tail-AV selector tracks changes in visual information read by action queries, rather than raw key/value displacement. On matched FastWAM tests, attacked Policy SR was 1.5% versus 47.5% for displacement selection; this result is limited to the evaluated checkpoint.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.