Reproduce “DexDICE: Reconciling Safety and Performance via Dual-Expert Offline Imitation 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
Seokin Seo, Sung Kuk Shyn, Minji Seo, Yisoo Lee, Jongmin Lee, Hongseok Yang, Kee-Eung Kim
DexDICE learns from safe but possibly suboptimal demonstrations, high-performing but possibly unsafe demonstrations, and unlabeled offline data, without reward or cost labels. It uses safe examples to estimate permitted state-action support and high-performing examples to guide behavior within it. Across 38 benchmark tasks, it has the highest safe return on 19, versus 9 for DemoDICE-S, while meeting cost limits on 31 versus 32 tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.