Distribird: Bayesian priors drawn from the literature

After decades of methodological work, almost everyone calibrating a process based Bayesian model still falls back on uniform priors. Not out of statistical conviction, but because building an informed prior from the literature is slow and demands both domain knowledge and statistical fluency. That unglamorous but accurate diagnosis opens Distribird, published on 13 August on arXiv.

The proposal is narrow and checkable: you hand it a parameter name, its physical description and the domain context, and a multi agent pipeline searches the literature, extracts reported values, weights them by domain relevance and fits a probability distribution through AIC model selection. The authors evaluate it on 24 parameters spread across 10 scientific domains.

What separates it from a chatbot with a search engine

Four design decisions, and none of them is the language model:

1. Weighting by domain relevance. Not every published value carries the same weight, and that is precisely the part a researcher does by hand and rarely documents.
2. Fitting through AIC model selection instead of assuming every uncertainty is normal or lognormal because that is convenient.
3. An explicit fallback to uninformative alternatives when the literature does not stretch far enough. The system admits it does not know instead of inventing a narrow prior.
4. Reporting the evidence behind each prior and its confidence level.

That last point is the one we find most important. A prior with no provenance is a methodological decision hidden inside a paper. If the tool returns the number together with the sources supporting it and a declared confidence, a reviewer can argue with it. If it returns only the number, the shortcut has been automated, not the work.

Open weight models, not APIs

The evaluation compares three open weight models: Qwen3.6 27B, Gemma 4 31B and a third whose name is cut off in the public arXiv abstract. The choice is not incidental. A scientific pipeline depending on a proprietary API ages badly: the model shifts underneath, the result stops being reproducible and the data leaves the institution. With open weights, a calibration from two years ago can be run again as is.

Who it is useful for

For process modelling with physically interpretable parameters: hydrology, agronomy, ecology, epidemiology, pharmacokinetics, materials engineering. The authors themselves bound the scope, and rightly so: this is no use for a neural network whose individual weights mean nothing, nor for parameters nobody has ever published.

The pattern, beyond the paper

Distribird is also a clean example of something built daily in the Claude ecosystem with subagents and MCP servers: one stage per agent, structured output at each step and a verifiable artifact at the end. Searching, extracting, weighting and fitting are four different jobs, and separating them lets you audit where things went wrong when the result does not add up. The practical difference is between "the system proposes 0.35" and "the system proposes 0.35 from these six sources, with this fit and this confidence".

What is still missing

The public abstract cuts off before the results, so we do not know how the three models compare or how far the generated priors sit from what an expert would build. And there is an underlying risk no tool solves: the published literature is already biased. A prior distilled from papers inherits the bias of those papers, and the confidence the system declares measures how much evidence there is, not how good that evidence is.

Distribird interests us not for automating but for leaving a trail: it says where each number comes from and how sure it is. In LLM assisted scientific tooling, that traceability is worth more than an extra point of precision.