Reproduce “Learning What to Ask and What to Believe: Distilling a Variational Bayesian Agent into Language Models” 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
Baichuan Li, Yanling Chang, Alfredo Garcia
The authors train an 8B language model to ask informative questions and track its evolving belief, using dialogues generated by a Bayesian teacher that demonstrates both. On Movie Recommendation, adding a belief-state scratchpad raises Success@10 from 0.085 to 0.345, while it makes little difference on Circuit Decoding, where queried bits are the final answer; tests found little transfer across tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.