Simulators for the Unmeasured World

Modern AI systems excel at learning patterns from existing data, but high-stakes scientific domains face a deeper limitation where the most critical data does not yet exist. Rare biological events, tail-risk financial scenarios, and cascading supply chain failures cannot be recovered through retrieval or scaled by scraping. We need models that learn not only from observation but from explicit causal structure and structured simulation.

In this talk, I will introduce the concept of Agentic Simulators. These are systems where agents operate inside Knowledge Graphs (KGs) that encode the constraints of a domain. Agents propose interventions and experiments, then call mechanistic and Large Quantitative Models (LQMs) as causal simulation engines to generate counterfactual trajectories. The resulting synthetic outcomes, along with their causal assumptions and provenance, are written back into the KG. This transforms the KG from a static memory into a computable causal world model.

I will outline design patterns for building such systems, drawing on examples from patient modeling and agentic chemistry workflows. Finally, I will argue that Simulation Augmented and Causally Grounded Reasoning over KGs offers a principled path beyond today’s Retrieval Augmented Generation. This shift enables AI systems to reason about rare events, latent mechanisms, and the unmeasured world.