Reproduce “Gibbs Variational Last Layers for Long-Horizon Dynamics Learning” 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

Paul Brunzema, Henrik Hose, Sebastian Trimpe

The authors train variational Bayesian last-layer dynamics models with multi-step losses, including CRPS and autoregressive rollouts, without changing the architecture. They compare predictive accuracy and distributional fit on synthetic and robot-motion tests. Autoregressive CRPS improves long-horizon rollout performance: median 100-step energy distance is 0.065, versus 0.087 for m-KDE. This does not guarantee safer decisions or better calibration.

Summary by alphaXiv.

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