LLM-Symbolic PDDL Model Repair

PDDL models are central to AI planning, yet they are often just assumed to be correct. Model repair allows us to verify whether a model is correct based on a given test suite, and if not, to repair it automatically. Existing methods are purely symbolic and search for repairs with minimal edits as a proxy for reasonable repairs. Although they guarantee test satisfaction, they can yield implausible fixes that ignore the semantic cues in action and predicate names. We introduce hybrid LLM-guided repair methods that incorporate this missing semantic knowledge. By injecting LLM preferences into the underlying MaxSAT search as soft constraints, our approach preserves the symbolic guarantee of passing all tests while steering the search toward semantically meaningful repairs. Across error-injected IPC domains and new unpublished domains, our methods improve the average F1 score from 42% to 83% (a 98% improvement).