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

Researchers introduce HALO, a trained orchestrator system that reduces LLM API costs by 45x compared to GPT-4-mini while matching performance on PDDL planning tasks. By leveraging verifier-certified trajectories as direct supervision rather than prompting frontier models at every step, HALO achieves significant cost efficiency improvements across multiple planning benchmarks.

HALO addresses a fundamental inefficiency in current agentic AI systems: the reliance on expensive frontier LLM API calls during every refinement step in planning tasks. The research demonstrates that verifier outputs—already present in existing systems—constitute high-quality supervision signals that can train smaller, fine-tuned models to make orchestration decisions effectively. This shift from prompt-based to learned orchestration represents an important maturation of agentic frameworks.

The broader context involves the growing tension between LLM capability and operational cost. As organizations deploy AI agents for complex tasks like formal planning, per-step API costs accumulate rapidly. HALO's approach of training a QLoRA-tuned policy on gold-standard trajectories offers a practical solution that doesn't sacrifice performance—it matches GPT-4-mini and comes within three percentage points of Gemini-3-Flash while dramatically reducing expenses. The 40-50% reduction in total LLM calls per episode indicates efficiency gains beyond just orchestration costs.

For developers and organizations building production AI systems, this research validates the economics of learning custom orchestrators rather than relying on frontier models. The 45x cost reduction ($0.18 to $0.004 per task) transforms planning-based applications from potentially prohibitive to scalable. This directly impacts feasibility for resource-constrained teams and enterprise deployments where API costs represent significant operational expenses.

Future work should examine whether this approach generalizes to other agentic tasks beyond planning, and whether similar supervision-from-verification patterns exist in other domains where verifiers can provide trajectory guidance.

• →HALO achieves 45x cost reduction in orchestration compared to GPT-4-mini while maintaining competitive performance on planning tasks.
• →Training orchestrators on verifier-certified trajectories provides superior supervision compared to sparse end-of-episode rewards or pure prompting.
• →The approach reduces total LLM calls per episode by 40-50%, addressing a critical efficiency bottleneck in agentic systems.
• →A small QLoRA-tuned model paired with hardcoded rules matches or exceeds frontier LLM baselines, validating learned orchestration over prompt-based approaches.
• →Verifier outputs can be leveraged as direct supervision signals, enabling cost-effective customization of agent orchestration systems.