New tool automates literature-informed prior distribution design for Bayesian calibration

Bayesian calibration of process-based models typically relies on uniform priors due to the difficulty of constructing informative priors from scientific literature, which requires both domain and statistical expertise. Researchers from Patrik P. Süli, György Eigner, and Roland Hollós introduce Distribird, an agentic web application designed to automate this process.

Distribird’s multi-agent pipeline takes a parameter name, physical description, and domain context as input, then searches the literature, extracts and weights reported values by domain relevance, and fits a probability distribution via AIC model selection. When no literature is available, the system falls back to sensible uninformative alternatives and reports the evidence and confidence level for each prior.

The tool was evaluated on 24 parameters across 10 scientific domains, comparing three open-weight models—Qwen3.6 27B, Gemma 4 31B, and Mistral Small 4 119B—against a single-prompt LLM baseline. On prior quality, the full pipeline matched the baseline, but Distribird provided additional benefits: every prior is explicitly traced to the specific papers and values used, and a built-in validity layer declines to produce priors for out-of-scope requests.

In contrast, the single-prompt baseline returned confident but unfounded priors for out-of-scope requests in 11 of 30 model–parameter cases. All language-model calls in Distribird run locally, ensuring that no parameter description or unpublished modeling detail is transmitted to a third-party LLM provider; only generated search terms reach public literature databases.

The authors argue that for scientific use, these properties—traceability, refusal of out-of-scope requests, and local execution—are more important than marginal gains in point-estimate accuracy.