Reproduce “Learning the Task or Learning the Verifier? Rethinking Criteria Exposure in Embodied Self-Improvement” 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

Wenbo Huang, Jianxin Bi, Shunyu Yang, Ruiyang Si, Rui Ni, Qiang Wang, Shulong Jiang, Yintianrun Zhang and 5 more

The authors test whether access to a benchmark verifier improves robot task performance or mainly its score under that verifier. On 16 RoboDojo tasks with mismatched criteria, a verifier-exposed library reaches 51.6% success under original criteria, versus 27.8% without exposure. Under instruction-faithful rewritten criteria, these results reverse: 45.0% versus 47.2%, respectively, in this simulated setting.

Summary by alphaXiv.

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