PhysStation: Compiling the Language of Physics into Formal Epistemic Multi-Agent Systems
Abstract
Autonomous discovery systems such as STATION (Chung & Du, 2025) demonstrate that open-world multi-agent environments produce emergent scientific behaviour, yet most existing systems rely on empirical score comparison for falsification---insufficient for physics, where formal laws encode hard logical constraints on what is possible. We introduce Formal Epistemic Multi-Agent Systems (FEMAS), a framework that compiles a domain's mathematical language directly into agent infrastructure. PhysStation, our physics instantiation, embeds group representation theory, renormalization group (RG) analysis, and scaling theory as first-class architectural primitives, spanning the full scientific discovery pipeline via a Formal Physics Knowledge Graph (FPKG), a three-engine Formal Falsification Engine (FFE), an RG-Guided Explorer (RGGE), and a three-level multi-scale agent hierarchy. We provide formal guarantees on falsification completeness, monotonic knowledge accumulation, and exponential exploration efficiency near phase boundaries, and target three open problems in frustrated quantum magnetism.