Reproduce “Probing Physics in Motion Diffusion Models Against Their Own Estimate” 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
Saeed Ghorbani
The authors test whether layers add information beyond a tuned readout of a model’s intermediate clean-motion estimate. On jump-heavy prompts, adding a layer improves prediction of final fall rate for the 50-step MDM from R² 0.35 to 0.43 at 20% denoising. This gain is transient; balance and flight show little added information, and the result is limited to tested models and readouts.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.