Training the Orchestrator: A Supervised Approach to End-to-End PDDL Planning with LLM Agents

The article presents HALO (Hybrid Agent-Learned Orchestrator), a new supervised approach for end-to-end PDDL planning that utilizes refinement trajectories certified by an external verifier to train the orchestrator. HALO employs a QLoRA-tuned policy combined with hardcoded rules, operating over a 21-agent action space, and demonstrates superior performance on PlanBench and classical planning benchmarks, achieving success rates comparable to or exceeding those of GPT-5-mini and Gemini-3-Flash while significantly reducing orchestration costs and LLM call frequency. This development is crucial for AI practitioners as it offers a more efficient and cost-effective method for translating natural language planning intents into verified plans, enhancing the usability of LLMs in practical applications.