Reproduce “Conditional Energy-Based Models for Implicit Policies: The Gap between Theory and Practice” 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

Duy-Nguyen Ta, Eric Cousineau, Huihua Zhao, Siyuan Feng

The authors analyze conditional energy-based policies for behavior cloning with multimodal demonstrations. They score observation–action pairs and search for actions; they interpret InfoNCE as a mutual-information lower-bound objective, not an approximation to negative log likelihood. Conventional Langevin better matches a Gaussian, while MAS appears on par with IBC in a particle-environment comparison. Mutual information is necessary, not sufficient, for generalization.

Summary by alphaXiv.

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