Reproduce “Amortizing Sequential Experimental Design in Reasoning LLMs” 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
Anita Kriz, Antonin Berthon, Nicolás Astorga, Mihaela van der Schaar
AR-BED trains a reasoning policy to choose questions in Twenty Questions instead of calculating Bayesian experimental design at inference time. It rewards expected information gain over targets and answers, including information from later turns, and updates the policy with GRPO. AR-BED has the highest solve rate in four tested categories; a separately trained policy improves on simulated MediQ cases.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.