Distribird: Literature-Informed Prior Distribution Design for Bayesian Model Calibration
Distribird is an agentic web application that automates the creation of literature-informed prior distributions for Bayesian model calibration by deploying a multi-agent pipeline to extract, weight, and fit scientific data, thereby offering a traceable and validity-checked alternative to the commonly used uniform priors or ungrounded single-prompt LLM baselines.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Science often relies on building models to understand how the world works, from the way crops grow in a field to how diseases spread through a population. These models are like complex machines made of equations, and they contain many adjustable parts called parameters. Some of these parts represent things we can measure directly, like the width of a river, but many represent hidden quantities, such as the rate at which a virus infects a cell or how deep plant roots reach into the soil. To make these models useful, scientists must tune these hidden parts until the model's predictions match real-world observations. This tuning process is known as calibration.
For decades, the most reliable way to tune these models has been through a method called Bayesian calibration. This approach asks scientists to state what they already believe about a hidden part before they look at new data. This initial belief is called a prior distribution. If a scientist knows that a certain temperature cannot exceed a specific limit, they can build that knowledge into the model. However, finding the right information to build these beliefs is incredibly difficult. The knowledge is scattered across thousands of scientific papers, and gathering it manually can take days of reading for a single model. Because this process is so slow, many researchers simply skip it and assume that every possible value for a parameter is equally likely. This shortcut is easy, but it often leads to models that are less accurate and slower to learn from new data.
A team of researchers has developed a new tool called Distribird to solve this problem. It is an automated system designed to read the scientific literature and build these initial beliefs for scientists, doing in minutes what used to take days. The system works by taking a description of a parameter, such as "the maximum temperature for photosynthesis in maize," and then sending out a team of digital agents to search scientific databases. These agents find relevant research papers, read them, and extract specific numbers reported by other scientists. The system then weighs these numbers based on how closely the original studies match the current situation. If a study was done on a different type of plant or in a different climate, the system treats that number as less important. Finally, it combines all the collected numbers into a single, well-defined probability distribution that represents the best current knowledge.
The researchers tested this tool on twenty-four different parameters across ten scientific fields, including climate science, ecology, and engineering. They compared the results of their automated system against a simpler method where a single artificial intelligence is asked to guess the distribution without looking at any papers. The results showed that both methods produced priors of similar accuracy. The automated system did not find a significantly better "center" for the distribution than the simple guess, but it provided something the simple guess could not: a complete trail of evidence. Every number used by Distribird is linked back to a specific sentence in a specific paper. A scientist can look at the result and see exactly which studies supported it, allowing them to verify the work or decide to ignore a source they do not trust.
Perhaps the most important finding was how the system handled questions it could not answer. When researchers asked the system about parameters that do not exist in nature, such as made-up names or internal software settings that have no physical meaning, the automated system correctly refused to give an answer. It flagged these requests as invalid and stopped. In contrast, the simpler method that did not search the literature confidently invented answers for these fake questions, providing specific numbers for things that do not exist. This ability to say "I do not know" rather than making things up is crucial for scientific safety. The researchers also ensured that the entire system runs on local computers using open-source software, meaning no private data or unpublished research details are sent to outside companies.
The cost of this extra safety and transparency is time and computing power. While a simple guess takes seconds, the full automated process takes about twenty to forty minutes per parameter and requires reading the full text of dozens of papers. The researchers found that for informal work, the quick guess might be sufficient. However, for serious scientific work where every number must be defensible and auditable, the automated system offers a level of trust that the quick guess cannot provide. It turns the slow, manual task of reading literature into a fast, repeatable process that leaves a clear paper trail, ensuring that the models scientists use are built on real evidence rather than guesses.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.