Search papers, labs, and topics across Lattice.
This paper introduces Meta-Ctrl, a constrained-decoding framework that effectively separates syntactic and semantic constraints in plan generation for robots, addressing the limitations of existing methods that either compromise plan quality or lack formal guarantees. By utilizing meta-tokens to enforce constraints at both the token and action levels, Meta-Ctrl significantly reduces the memory requirements for constrained decoding from over 107TB to under 2GB. The framework achieves superior performance on the WAH-NL benchmark, surpassing GPT-4 in subgoal success rates, and demonstrates practical applicability by generating executable plans for a tabletop robot that meet all necessary conditions.
Meta-Ctrl cuts constrained decoding memory from over 107TB to under 2GB while achieving higher subgoal success rates than GPT-4.
LLMs generate fluent plans for robots but routinely violate the syntactic and se8mantic 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 \textbf{Meta-Ctrl}, a constrained-decoding framework that guarantees the encoded constraints while preserving the base LM's plan quality. Meta-Ctrl introduces \emph{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 cuts the memory of constrained decoding from over 107TB to under 2GB. With it, a small open-weight LM becomes competitive where it otherwise sits at the bottom of the leaderboard: on WAH-NL under the LoTa-Bench protocol it reaches the highest reported subgoal success rate, exceeding GPT-4's, with consistent gains across the Embodied Agent Interface. We further demonstrate it on a real tabletop robot, where every generated plan satisfies its preconditions and goals by construction. Project website: https://meta-ctrlg.github.io/.
