Distribird automates Bayesian priors from literature with traceable AI

Bayesian calibration of process-based models has long required researchers to manually construct prior distributions, so most default to uniform priors out of convenience. Distribird, a new tool from Patrik P. Süli, György Eigner, and Roland Hollós (arXiv:2608.11210), aims to eliminate that bottleneck by automating the entire prior-elicitation process. Given a parameter name, description, and domain context, its multi-agent pipeline searches published literature, extracts reported parameter values, weights them by relevance, and fits probability distributions using AIC model selection. It supports physically interpretable parameters in fields where domain knowledge exists in published papers.

In tests across 24 parameters and 10 scientific domains, Distribird's full pipeline matched a single-prompt LLM baseline in prior quality, but with two crucial distinctions. First, every generated prior is traced to the specific papers and values it was built from. Second, a built-in validity layer declines requests that fall outside its scope, whereas the baseline confidently produced unfounded priors in 11 of 30 model-parameter cases. The tool runs all language-model calls locally—so unpublished parameter details never reach third-party providers—and only search terms are sent to public literature databases. The authors deliberately prioritize transparency and safety over marginal point-estimate accuracy, making Distribird a practical step toward rigorous, reproducible Bayesian modeling.