Reproduce “PDDLCoder: Agentic PDDL Generation for LLM-Assisted Symbolic Planning” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Veit Laule, Jiangtao Shuai, Manfred Hauswirth, Sonja Schimmler

PDDLCoder introduces an agentic framework for generating Planning Domain Definition Language (PDDL) specifications from natural language descriptions, leveraging external tools for iterative refinement. Using DeepSeek v4 Flash, the system achieved applicable plans for 89.6% of problems in the new NL-pddlgym benchmark, which evaluates plan applicability against executable environments.

Key insights

  • An agentic framework with iterative refinement and diverse external feedback tools substantially enhances the quality of PDDL generation and the applicability of derived plans compared to rigid, single-pass approaches.
  • Semantic evaluation of generated plans against natural language task descriptions, provided by a dedicated PlanAgent, is crucial for identifying and correcting logical inconsistencies in PDDL models beyond mere syntactic correctness.
  • The effectiveness of LLM-assisted PDDL generation is highly dependent on the underlying LLM's capabilities, particularly its ability to provide usable plan feedback and generate syntactically valid PDDL.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.