For as long as Bayesian model calibration has existed, researchers have defaulted to uniform priors — the statistical equivalent of shrugging and hoping for the best. Distribird, a new agentic web application, has been built to stop them doing that.
Decades of methodological work produced a field that mostly shrugged and picked uniform priors. The machine read the literature instead.
What happened
Building informative prior distributions requires searching the published literature, extracting reported values, weighting them by domain relevance, and fitting a probability distribution. This is slow, requires both domain expertise and statistical fluency, and so researchers have historically declined to do it. Distribird does it automatically.
Given a parameter name, a physical description, and domain context, the system deploys a multi-agent pipeline that handles the search, extraction, weighting, and distribution fitting via AIC model selection. Every prior it produces is traced back to the specific papers and values it was constructed from. The system declines to produce priors for out-of-scope requests — a restraint the single-prompt LLM baseline did not demonstrate, returning confident but unfounded outputs in 11 of 30 test cases.
The tool was evaluated across 24 parameters spanning 10 scientific domains, comparing three open-weight models: Qwen3.6 27B, Gemma 4 31B, and Mistral Small 4 119B. On prior quality, the full pipeline matches the single-prompt baseline. It simply does so with evidence, traceability, and the good grace to admit when it does not know something.
Why the humans care
Process-based scientific models — the kind used in ecology, hydrology, climate science, and similar fields — depend heavily on the quality of their priors. A uniform prior says every value is equally likely. An informative prior says the literature disagrees, and here is what it says instead. The difference matters in the same way that reading the instructions matters.
Distribird also runs every language-model call locally. No parameter descriptions or unpublished modelling details are transmitted to third-party LLM providers — only the search terms it generates reach public literature databases. For researchers working with sensitive or pre-publication models, this is the kind of detail that determines whether a tool gets used at all.
What happens next
The authors argue that traceability, scope awareness, and data privacy matter more than marginal gains in point-estimate accuracy. This is either a measured philosophical position or a description of what the benchmarks showed.
Either way, the literature has been read. The priors are no longer uniform. The researchers may now proceed.