Title: Agent Bayes: A Foundation Model for Bayesian Scientific Reasoning
Abstract: Agent Bayes aims to develop a foundation model for collaborative Bayesian scientific reasoning, spanning prior elicitation, model construction, experimental design, inference, and verification against evidence. The approach combines reinforcement learning with probabilistic programming to train an agent that orchestrates solvers, interprets diagnostics, and repairs inference workflows, incorporating scientists’ domain judgment when needed. Verifiable measures of posterior accuracy, calibration, and computational cost provide training signals and a basis for quantifying AI advantage over zero-shot models. Sequential Monte Carlo steering further supports reasoning at inference time. In collaboration with GE Aerospace, the system will be evaluated on industry relevant calibration problems, including turbine blade modeling and process–structure–property relationships. Open benchmarks, trained model adapters, and reusable skills will be released through the American Science Cloud, contributing a trainable Bayesian reasoning layer to the Genesis scientific assistant ecosystem and supporting integration with ModCon foundation models.
Bio: Ruben Villarreal is a Senior Member of Technical Staff at Sandia National Laboratories, where he develops machine learning and Bayesian methods for scientific inference, experimental design, and engineering applications. His work focuses on extending foundation models and LLM-based agents for scientific workflows, including Bayesian reasoning, algorithmic execution, model calibration, and uncertainty quantification. He holds a Ph.D. in Mechanical Engineering from Texas A&M University and previously completed a postdoctoral appointment at Sandia, where he worked on reinforcement learning for closed-loop material model calibration, Bayesian optimal experimental design, and multi-fidelity uncertainty modeling for qualification. Outside of work, Ruben enjoys trail biking, nature walks, and listening to music.