Meta-Ctrl: Guaranteed Plan Generation by Decoupling Syntactic and Semantic Constraints

LLMs generate fluent plans for robots but routinely violate the syntactic and semantic constraints they must satisfy to execute, and existing remedies trade formal guarantees against plan quality: soft methods (affordance scoring, grounded decoding) give no guarantee, while symbolic planners (LLM+P) discard the LM's commonsense. We propose Meta-Ctrl, a constrained-decoding framework that guarantees the encoded constraints while preserving the base LM's plan quality. Meta-Ctrl introduces meta-tokens—a compact vocabulary of grounded actions—enforcing syntax at the token level and semantics (preconditions, goals, ordering) at the action level, an exact factorization that keeps the semantic search at action granularity instead of compiling it into the token stream, which for a typical task reduces the lookahead table from a theoretical 107TB to 1.6GB. With it, a small open-weight LM becomes competitive where it otherwise sits at the bottom of the leaderboard: Llama 3 8B leads VirtualHome action sequencing on the Embodied Agent Interface, ahead of recent frontier models including Claude Opus 4.8 and GPT-5.5; on WAH-NL under the LoTa-Bench protocol it reaches the highest reported subgoal success rate. We further demonstrate it on a real tabletop robot, where every generated plan satisfies its preconditions and goals by construction.