Title: Simulate, Reason, Decide: Scientific Reasoning with LLMs for Simulation-Driven Decision Making

Abstract:

High-stakes decision-making processes are increasingly leveraging LLM-driven systems that incorporate scientific simulators. However, current frameworks typically treat these simulators as opaque black-box interfaces, employing LLMs merely to generate, calibrate, or execute them. This approach fails to treat simulators as structured mechanistic systems capable of being reasoned about, thereby preventing the identification, representation, and analysis of the assumptions and mechanisms that drive simulator behavior. Consequently, existing methods suffer from limited transparency, auditability, and the ability to justify decisions.

To address these limitations, we present MechSim, a neuro-symbolic reasoning framework grounded in mechanisms designed for executable scientific simulators. In contrast to previous neuro-symbolic models that focus on static symbolic structures, MechSim empowers LLM agents to reason directly about the mechanisms, underlying assumptions, and execution dynamics of scientific simulators. The framework utilizes a shared structured schema to represent simulators, capturing critical elements such as assumptions, variables, mechanism dependencies, and execution traces. Building upon this representation, LLM agents function as constrained reasoning engines, producing structured explanations that are grounded in evidence and explicitly link simulator outcomes to their foundational mechanisms. Our evaluation across several high-stakes domains demonstrates that this approach enhances the quality of mechanism-level explanations, improves simulator analysis, and increases the reliability of downstream decision-making.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC