Reproduce “Asking Smarter Questions: Targeted Active Preference Learning for Underspecified Rewards” 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

Helena Merker, Nick Walker, Andreea Bobu

The method uses a diagnosis of demonstration coverage to shape a robot’s prior and preference questions. It preserves information for well-supervised features, removes demonstration influence from underspecified features, and pulls irrelevant weights toward zero; explanations direct human attention. In simulations, it reports 13.0% higher normalized reward than Standard after five queries. A 12-person study supports both components; it used a ground-truth partition.

Summary by alphaXiv.

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