Distribird: Literature-Informed Prior Distribution Design for Bayesian Model Calibration

August 13, 2026 | Source: arXiv AI

Tags: Bayesian calibration, multi-agent systems, prior distributions, scientific AI, Qwen3, Gemma 4

Distribird automates informative Bayesian prior design using a multi-agent pipeline that searches the literature, weights reported values by domain relevance, and fits probability distributions — with all LLM calls running locally so no unpublished parameter data reaches external providers.

Details

Bayesian calibration of process-based models requires domain-specific prior distributions for each parameter, but researchers routinely fall back on uninformative uniform priors because building literature-informed priors demands both statistical expertise and time-consuming manual search. Distribird (arXiv:2608.11210) automates this bottleneck with a multi-agent pipeline. Given a parameter name, physical description, and domain context, Distribird searches the literature, extracts reported values, weights them by domain relevance, and selects a probability distribution via AIC model selection. When no literature exists, it falls back to sensible uninformative alternatives and reports its confidence level. A built-in validity layer declines out-of-scope requests. The system was evaluated on 24 parameters across 10 scientific domains using three open-weight models: Qwen3.6 27B, Gemma 4 31B, and Mistral Small 4 119B. The full pipeline matches a single-prompt LLM baseline on prior quality — but every prior is traced to specific source papers, and the system declined unfounded requests whereas the baseline returned confident but unsupported answers in 11 of 30 test cases. Crucially, all LLM inference runs locally, so no unpublished model parameters reach external providers. Authors argue provenance, scope-awareness, and data privacy matter more than marginal point-estimate accuracy for scientific use.