Reproduce “A Phenomenology of Neural Geometry in Vision-Language-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

Mozes Jacobs, Aiden Swann, Mathilde Papillon, Siddharth Boppana, Fenil R. Doshi, Usha Bhalla, Thomas Icard, Matt Feiszli and 11 more

The authors map how MolmoBot’s activations change as its policy variables vary, distinguishing readable signals from their geometry and controllability. In the vision-language model at layer 21, distance has nearly orthogonal coarse and fine codes: held-out R² is 0.81 at 15–70 cm and 0.70 below 15 cm. This is supervised simulation evidence, not proof of cause or improved performance.

Summary by alphaXiv.

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