LLM+P
Large Language Models + Classical Planners
An LLM translates a natural-language planning problem into a formal PDDL problem file, and the Fast Downward classical planner computes a sound or minimum-cost plan.
The loop
- Proposes
A PDDL problem specification (objects, initial state and goal predicates) translated from natural language by GPT-4
- Judged by
The Fast Downward classical symbolic planner (classical planners LAMA and FDSS) running against the domain action schema
- Steers
Fast Downward searches the symbolic state space to return a verified plan, which is translated back into natural language
Headline result
Across seven standard International Planning Competition (IPC) domains including Blocksworld, Barman, Floortile, Grippers, Storage, Termes and Tyreworld (20 tasks per domain), LLM+P with GPT-4 and in-context examples solved the majority of tasks with valid, cost-minimizing plans, whereas direct GPT-4 plan generation failed on almost all tasks in Barman, Floortile, Storage and Termes.
What this does not prove
LLM+P requires a human expert to supply the PDDL domain file (action preconditions and effects) in advance, and its one-shot translator has no validator-repair loop when the LLM emits syntax errors, undeclared objects or inverted goal predicates.
How it generalizes the harness
- The algorithm is the solver
- A domain authority judges
- The output outlives the model
- Continuous-discrete duality
The LLM does not search for action sequences at all; it compiles natural language into PDDL so Fast Downward can perform sound symbolic search.
Classification
| What it proposes | Formal specification |
|---|---|
| Generative model | GPT-4 (frozen) with in-context PDDL problem examples |
| Algorithmic authority | Fast Downward classical symbolic planner |
| Feedback | Test results |
| What gets updated | Prompt context |
| When it runs | At inference time |
| Validity | Formal (machine-checked) |
| Evidence | Experiments reported |
| Lineage | Original system |