Finding Semantically Guided Repairs in PDDL Domains Using LLMs

Repairing Planning Domain Definition Language (PDDL) models is difficult because solutions must ensure correctness while remaining interpretable to human modelers. Existing hitting set methods identify minimal repair sets from whitelist and blacklist traces, but they cannot prefer semantically meaningful fixes and the true repair may not be minimal. We propose combining large language models (LLMs) with the hitting set framework, using semantic cues in PDDL action and predicate names to guide repairs. This hybrid approach provides contrastive, counterfactual explanations of why traces fail and how domains could behave differently.