LLM-Elicited Priors for Model Predictive Control
Turning domain knowledge in language models into structured priors for decision-making under data scarcity
Bayesian methods need a prior, but in many emerging scientific settings there is not enough target-regime data to estimate one reliably. This project asks whether large language models can help elicit structured prior knowledge before new observations arrive.
What the LLM provides
Rather than a single point estimate, the LLM supplies a prior ensemble of plausible causal structures, parameter ranges, and scenario-conditioned samples that reflect uncertainty across possible regimes.
Application
I study this idea in epidemic control, a setting where historical data from a new pathogen is scarce by definition. LLM-elicited scenarios serve as a prior ensemble for model-predictive control of non-pharmaceutical interventions, allowing the controller to act before surveillance data accumulates.
Main message
The goal is not to replace data with an LLM, but to turn available domain knowledge into an explicit prior that can later be updated with data.