Reproduce “Causal Analysis of Robot Foundation Models Enables Efficiency Gains” 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
Aiden Swann, Mathilde Papillon, Mozes Jacobs, Siddharth Boppana, Tal Haklay, Eric Bigelow, Andrew Lee, Conor Watts and 8 more
The authors trace causal influences in three vision-language-action models using activation patching, identifying a narrow interface of one to three middle-to-late layers. Early representations affect object identity, while later ones affect target location. In simulation, MolmoBot using layers 23–25 scores 46.7% aggregate versus 46.9% for all 36 layers across 3,000 episodes; this architecture-specific result does not establish exclusive information storage.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.