PlanForge: Synthesizing Planners from Natural Language with LLM-driven Architect–Builder–Runner Pipeline

Large language models can generate plausible step-by-step plans, but such plans are not necessarily executable or correct. Reliable planning requires translating a natural-language task into a computational procedure that tracks state, enforces constraints, selects valid actions, and searches for a solution. Recent programmatic planning approaches improve reliability by operating over such procedures, but typically assume the underlying blueprint (state representation, action structure, and solving strategy) is already defined, often by humans. This is limiting, since natural-language planning requires choosing the representation, the decomposition, and the strategy, not just executing them. PlanForge addresses this gap by synthesizing the blueprint that prior methods assume is given. From natural-language task families, PlanForge induces a blueprint comprising the state representation, a decomposition into function signatures, an algorithmic solving strategy, and the action structure. An architect designs this blueprint, a builder compiles it into an executable program using development-set feedback, and a runner runs the program on held-out instances under external validation. We evaluate generalization across three benchmarks spanning complementary planning and constraint-satisfaction problems, namely NaturalPlan (trip, meeting, and calendar planning), ACPBench (sequential planning across 15 domains), and StructureSAT (9 structured constraint satisfaction tasks). Across these benchmarks, PlanForge achieves a 100 plan-generation success rate on all held-out test sets, outperforming direct prompting by up to 72.9 and the programmatic baseline SCOPE by up to 57.8. Furthermore, by compiling the planner into a standalone program, PlanForge executes locally on new instances, eliminating test-time API queries and reducing cumulative token costs by up to 15x compared to programmatic baselines.