BlocksWorld, Logistics, Mini-Grid, Trip Planning, Calendar Scheduling
收藏资源简介:
本研究构建了包括BlocksWorld、Logistics、Mini-Grid等在内的多个数据集,用于评估大型语言模型(LLMs)的规划能力。这些数据集涵盖了从简单的积木世界到复杂的物流规划等多种场景,通过PDDL和自然语言两种形式描述问题。数据集的创建过程遵循严格的步骤,包括初始状态和目标状态的设定、问题在PDDL中的表达以及使用经典规划器解决问题的过程。这些数据集不仅用于评估LLMs的规划性能,还用于研究模型在不同复杂度问题上的泛化能力,旨在提升LLMs在实际应用中如会议安排和旅行规划等任务的规划能力。
This study develops multiple datasets including BlocksWorld, Logistics, Mini-Grid, and others, to evaluate the planning capabilities of Large Language Models (LLMs). These datasets cover diverse scenarios, ranging from simple block-based environments such as BlocksWorld to complex logistics planning tasks, with problems described in both PDDL and natural language formats. The construction of these datasets follows a rigorous procedure, which encompasses the specification of initial and goal states, the formalization of problems in PDDL, and the solution process via classical planners. These datasets serve not only to assess the planning performance of LLMs, but also to explore the generalization ability of models across problems of different complexities, with the ultimate goal of enhancing the planning capabilities of LLMs in real-world applications like meeting scheduling and travel planning.

- 1Exploring and Benchmarking the Planning Capabilities of Large Language Models谷歌深度思维 · 2024年



