Teaching LLMs to Plan (arxiv.org)

Researchers introduced PDDL-Instruct, an instruction-tuning framework that teaches large language models to do formal symbolic planning by training them to produce logical chain-of-thoughts. Rather than relying on free-form reasoning, the method prompts models to explicitly check preconditions, apply effects, and verify invariants step-by-step in Planning Domain Definition Language (PDDL) style problems. By shaping model outputs into precise logical inference steps and structured reflection, PDDL-Instruct lets LLMs self-correct planning errors and rigorously evaluate action applicability and state transitions. On multiple standard planning benchmarks the instruction-tuned models show large gains—reaching up to 94% planning accuracy, a 66 percentage-point absolute improvement over baselines—demonstrating that chain-of-thought supervision can close the gap between LLMs’ general reasoning and the exactness required for automated planning. Technically, the work highlights that targeted instruction tuning (with explicit reasoning traces) can imbue LLMs with verifiable planning behaviors compatible with formal representations like PDDL. This points toward practical routes for combining neural language models with symbolic planners in robotics, autonomous agents, and verification tasks where correctness and stepwise justification are essential.