PDDLCoder: Agentic Framework Enhances LLM-Assisted Symbolic Planning
A new framework named PDDLCoder has been developed by researchers to enhance the dependability of large language models (LLMs) in tasks requiring long-term planning. This framework systematically generates, evaluates, and refines Planning Domain Definition Language (PDDL) specifications based on natural language inputs, tackling prevalent challenges like logical inconsistencies and irrelevant plans. To facilitate assessment, the researchers also introduced NL-pddlgym, a benchmark dataset featuring 711 planning scenarios across 23 domains, complete with executable gym environments for automated plan verification. Tests conducted on the NL-pddlgym test set, which includes 106 problems, showcase the framework's efficacy. This research is documented in a paper available on arXiv (arXiv:2608.16637), marking a significant step towards standardized benchmarks in LLM-assisted symbolic planning.
Key facts
- PDDLCoder is an agentic framework for PDDL generation from natural language.
- It iteratively generates, analyzes, and refines planning specifications.
- NL-pddlgym is a new benchmark dataset with 711 planning problems across 23 domains.
- The benchmark includes executable gym environments for automated verification.
- Experiments were conducted on the NL-pddlgym test set containing 106 problems.
- The framework aims to improve reliability of LLMs in long-horizon planning.
- The paper is available on arXiv with ID 2608.16637.
- The research addresses the lack of standardized benchmarks with automated verification.
Entities
Institutions
- arXiv