Symbolic Intermediaries Bridge LLMs and Physics Simulators

A new research paper proposes 'symbolic intermediaries' as a linguistic-numerical interface to enable LLMs to reason over continuous numerical outputs from physics simulators. The approach uses symbolic regression to translate simulator traces (distances, curvatures, trajectories) into compact analytical expressions that language models can interpret while preserving geometric semantics. An agentic coordination-and-refinement loop is built around this interface, where a design agent maps numerical data to symbolic forms for tasks like mechanism design and motion planning. The work addresses a fundamental gap in LLMs' ability to handle spatially grounded engineering reasoning.