Enabling Robust Epidemic Control via LLM-Elicited Causal Discovery

Large language models (LLMs) can supply scientific priors, but it remains unclear how such priors might enhance high-stakes control under uncertainty. We study this in a zero-current-data epidemic-control setting on a deterministic SIQR benchmark with five non-pharmaceutical intervention levers. Our co-scientist pipeline elicits a consensus DAG, constructs an LLM-elicited sample pool, fits an LLM prior, and exposes the resulting artifacts to model-predictive control (MPC). Against a Uniform-prior MPC baseline and a literature prior, the LLM prior plays a bounded but useful role: its open-loop mode reduces cumulative infections by $25.8\\%$ relative to the content-free $\mathrm{Beta}(1,1)$ (Uniform) baseline, but the intervention strengths it recommends inside the controller are less well calibrated than those of the literature-based prior. Keeping the full pool of LLM samples inside the controller works better than collapsing it into a single point prior. Disagreement among the samples then becomes an explicit measure of regime uncertainty. These results suggest that co-scientist systems are useful for prior structure and scenario coverage, while robust control should ultimately update the LLM prior with real observations and propagate posterior uncertainty through MPC.