Reproduce “Stacking-Based Weighting for Large Language Bayes in M-Open Settings” 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
Anant Bhide, Akarsh Gupta, Edmond Cunningham, Justin Domke
Large Language Bayes replaces evidence-based Bayesian model averaging with KL-regularized stacking weights optimized using variational estimates of leave-one-out prediction. It retains language-model-generated candidate probabilistic models and inference within each. On three hazard-loss tasks, weights spread across candidates while held-out predictive scores remain comparable to BMA; hurricane scores are −5.983 and −5.999, respectively. The results do not establish uncertainty recovery.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.