Reproduce “CausalWAM: Learning the Causes of Actions 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

Juntao Cheng, Jinzhou Tang, Fan Feng, Henrik I. Christensen, Yujia Zheng

CausalWAM keeps a whole-scene representation for prediction while adding a sparse router from visual and task tokens to the action head. Under stated non-degeneracy conditions, its theory identifies the action-relevant latent block up to invertible re-encoding, not individual factors. On 17 RoboCasa365 tasks, it reaches 0.43 mean success versus 0.24 for matched Cosmos 3, winning on 13 tasks.

Summary by alphaXiv.

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