Prior-Amortized In-Context Bayesian Inference for Generalized Linear Mixed-Effects Models
A pretrained network called metabeta is presented as a faster Bayesian inference route for GLMMs without fixing the prior at training time.
The paper says metabeta takes prior families and hyperparameters as test-time inputs, allowing zero-shot use across new datasets, models, and priors. Its default flow posterior is refined with Independence Metropolis-Hastings, with inference reported as two to three orders of magnitude faster than NUTS. The authors say it matches NUTS on parameter recovery, calibration, prediction, and real out-of-distribution datasets. The model is described as open-source and open-weights. ArXiv · AI/CL/LG's note
The paper says metabeta takes prior families and hyperparameters as test-time inputs, allowing zero-shot use across new datasets, models, and priors. Its default flow posterior is refined with Independence Metropolis-Hastings, with inference reported as two to three orders of magnitude faster than NUTS. The authors say it matches NUTS on parameter recovery, calibration, prediction, and real out-of-distribution datasets. The model is described as open-source and open-weights. ArXiv · AI/CL/LG's note
score 4