The research paper presents a new framework called ‘Formal-LLM’ that integrates the expressiveness of natural language with the precision of formal language to control the plan generation of Large Language Model (LLM)-based AI agents. The framework allows human users to express their requirements or constraints for the planning process as an automaton. The LLM then generates a plan under the supervision of the automaton, ensuring the plan satisfies the constraints. This process makes the planning process controllable and prevents the generation of invalid or unsuccessful plans. Experiments showed this method led to a 50% overall performance increase.
Publication date: 2 Feb 2024
Project Page: https://github.com/agiresearch/Formal-LLM
Paper: https://arxiv.org/pdf/2402.00798
