文章背景与核心概要
大语言模型（LLM）擅长为机器人生成流畅的计划，但往往无法满足成功执行所需的严格语法和语义约束。现有的解决方案往往需要在两者之间进行妥协：“软方法”（如基础解码）缺乏形式化保证，而符号规划器（LLM+P）则常常牺牲底层模型的常识推理能力。
Meta-Ctrl 引入了一种新颖的受限解码框架，弥补了这一鸿沟。通过将约束解耦为两层——在词元（token）级别强制执行语法，在动作级别强制执行语义（前置条件、目标和顺序）——作者在不妥协基础模型性能的情况下，实现了有保证的计划生成。这种方法显著优化了内存使用，将受限解码的资源需求从 107TB 以上降低到 2GB 以下。实验结果表明，Meta-Ctrl 使得较小的、开放权重的模型在 WAH-NL 基准测试中超越了 GPT-4，并在桌面机器人任务上成功完成了现实世界的验证。
Meta-Ctrl: Guaranteed Plan Generation by Decoupling Syntactic and Semantic Constraints
Authors: Gwen Yidou-Weng, Edward Sun, Tianyi Ma, Metin Alp Dogan, Benjie Wang, Allen Peng, Guy Van den Broeck, Yuchen Cui
Date: August 23, 2026
Subject: Robotics (cs.RO); Artificial Intelligence (cs.AI)
arXiv ID: 2608.22149
摘要
Large Language Models (LLMs) are adept at generating fluent plans for robots but frequently fail to adhere to the strict syntactic and semantic constraints required for successful execution. Existing solutions often force a trade-off: "soft" methods (like grounded decoding) lack formal guarantees, while symbolic planners (LLM+P) often sacrifice the commonsense reasoning capabilities of the underlying model.
Meta-Ctrl introduces a novel constrained-decoding framework that bridges this gap. By decoupling constraints into two layers—enforcing syntax at the token level and semantics (preconditions, goals, and ordering) at the action level—the authors achieve guaranteed plan generation without compromising the base model's performance. This approach significantly optimizes memory usage, reducing the requirements for constrained decoding from over 107TB to under 2GB. Experimental results demonstrate that Meta-Ctrl allows smaller, open-weight models to outperform GPT-4 on the WAH-NL benchmark, with successful real-world validation on tabletop robotic tasks.
核心贡献
- Meta-Token Architecture: Introduces "meta-tokens"—a compact vocabulary of grounded actions—that allows for precise control over plan generation.
- Decoupled Constraint Enforcement: Separates syntactic and semantic constraints, enabling formal guarantees on plan validity by construction.
- Efficiency: Achieves a massive reduction in memory overhead for constrained decoding, making the approach scalable and practical.
- Superior Performance: Demonstrates state-of-the-art subgoal success rates on the LoTa-Bench protocol and provides robust, error-free execution on physical robotic hardware.
访问与资源
- Project Website: meta-ctrlg.github.io
- Full-text Links:
