AI
Distribird automates design of Bayesian priors by mining scientific literature
A new agentic tool uses language models to extract parameter values from published research and automatically construct statistically sound prior distributions for scientific model calibration.
Bayesian calibration of computational models relies on setting appropriate prior distributions for each parameter, yet researchers have long defaulted to generic uniform priors because building informed alternatives demands deep domain knowledge and statistical skill. A new system called Distribird aims to remove that barrier by automating the entire prior construction pipeline.
How it works#
Distribird is an agentic web application that takes a parameter name, physical description, and scientific domain as input, then orchestrates a multi-step workflow: it searches published literature for reported values of that parameter, extracts and ranks them by relevance to the domain, and fits a probability distribution using AIC-based model selection. When no literature is available, the tool defaults to sensible uninformative alternatives and always reports the evidence base and confidence level for every prior it generates.
Evaluation across domains#
The researchers tested Distribird on 24 parameters spanning 10 scientific disciplines, comparing three open-weight language models (Qwen 3.6 27B, Gemma 4 31B, Mistral Small 4 119B) against a single-prompt LLM baseline. On prior quality, the full pipeline matched the simpler baseline. However, the system demonstrated critical advantages: every prior can be traced back to the specific papers and values used to construct it; an internal validity layer declines to produce priors for out-of-scope requests (whereas the baseline returned confident but unsupported priors in 11 of 30 test cases); and all language-model computations run locally, ensuring parameter descriptions and unpublished modelling details never reach external LLM providers—only generated search queries reach public literature databases.
Why it matters#
For scientific applications, the researchers argue that reproducibility, traceability, and user control matter more than marginal gains in point-estimate accuracy. By democratizing the construction of literature-informed priors, Distribird opens Bayesian calibration methods to researchers who lack statistical expertise, while maintaining transparency and scientific rigor in domains where physically interpretable parameters and published domain knowledge already exist.
