MetaBeta
In-context Bayesian inference for hierarchical data.
Grouped tabular data is everywhere; GLMMs allow statistical inference on it. Bayesian models incorporate prior knowledge and properly quantify uncertainty. But fitting them (repeatedly) with Hamiltonian Monte Carlo is expensive.
Our bespoke neural model combines Laplace approximation, Set Transformers, Neural Spline Flows, and Independence Metropolis-Hastings. It takes data, models and priors as in-context inputs, and samples from the full posterior.
Across controlled benchmarks and real datasets, metabeta closely matches NUTS in parameter recovery, calibration, posterior geometry, and out-of-sample prediction. This still holds in stress tests including misspecification, distribution shift, collinearity, and data-poor regimes.