Reproduce “Sharp Characterization of Bias in Post-Bandit Inference” 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

Lisu Wang, Yilun Chen, Jiaqi Lu

The authors analyze how stable index bandits’ adaptive allocations bias post-experiment estimates. Their empirical fluid approximation captures leading dependence between rewards and pull counts; effective exploration rate measures allocation responsiveness. Under UCB1, tied-optimal arms have negative standardized bias decaying only as 1/√log T. A Gaussian example gives 6.8% Type I error at nominal 5%; this is a leading-order calculation, not a general rate.

Summary by alphaXiv.

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