Reproduce “Step-DAD: Semi-Amortized Policy-Based Bayesian Experimental Design” 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

Marcel Hedman, Desi R. Ivanova, Cong Guan, Tom Rainforth

Step-DAD adapts Bayesian experimental design policies during experiments. Starting from a pretrained policy, it uses observed outcomes to update the posterior and fine-tune plans for remaining noisy measurements. In a simulated source-location task with 10 measurements and an update after six, its estimated lower-bound EIG was 7.759 ± 0.114, above DAD’s upper bound of 7.089 ± 0.013. Under tested prior shifts, it retained positive information gain, without establishing broad robustness.

Summary by alphaXiv.

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