Moving Beyond Heuristic Generation

Traditional LLM-based planning often relies on heuristic generation, which frequently results in brittle plans that fail to generalize across different problem instances. This research shifts the paradigm by introducing a framework for 'provably complete' generalized planning. Instead of treating the LLM as a black-box generator of specific action sequences, the authors leverage the model to synthesize plans that satisfy formal correctness criteria. This ensures that the generated policy is not just a one-off solution, but a generalized strategy that holds across a class of problems.

Formal Verification and Completeness

The core contribution is the integration of formal verification techniques with LLM reasoning. By mapping LLM-generated plans to a formal representation, the framework can verify whether a plan is complete—meaning it covers all reachable states and guarantees a solution for any valid input instance. When the LLM fails to produce a complete plan, the system uses the feedback from the verification step to iteratively refine the plan. This closed-loop approach transforms the LLM from a probabilistic guesser into a component of a rigorous, verifiable planning pipeline. The result is a system that maintains the flexibility of natural language reasoning while adhering to the strict requirements of formal planning theory.