Distribird Automates Bayesian Prior Design from Scientific Literature.

Bayesian model calibration often relies on uniform priors due to the complexity and time required to construct informative priors from scientific literature. This paper introduces Distribird, an agentic web application designed to automate this process. Given a parameter name, physical description, and domain context, Distribird deploys a multi-agent pipeline. This pipeline searches academic literature, extracts and weights reported values based on their relevance, and then fits an appropriate probability distribution using AIC model selection. If no literature is found, it defaults to sensible uninformative priors, clearly reporting the evidence and confidence level for each. Evaluated on 24 parameters across 10 scientific domains, Distribird matched a single-prompt LLM baseline in prior quality but significantly surpassed it in reliability. It traces every prior back to its source papers and values, and includes a validity layer to decline out-of-scope requests, preventing unfounded priors. Crucially, all language model calls run locally, ensuring that sensitive parameter descriptions or unpublished modeling details remain private.