translating-natural-language-planning-problems-into-pddl-be-learned-solver CLM-0326observationsingle paperReviewed by AI
Translating natural-language planning problems into PDDL can be learned from solver feedback alone, with one model acting as actor, judge and editor: on PlanBench average success rose from 35.5% (LLM+P) to 70.8%, with 66.3% of successes judged faithful to the described task and semantic drift, where a solvable but wrong problem is produced, cut to 6.4%.
Capability: Planning a sequence of actions before acting
Observed on
PlanBench, solver-grounded RL without human-written PDDL demonstrations, 2026; faithfulness judged against the natural-language description..
Sources
- supportsFrom Solver Feedback to Faithful Plans: Multi-Role Reinforcement Learning for Symbolic PlanningValidated by: PlanBench success, faithful-success and semantic-drift rates against LLM+P and other baselines; faithfulness checks separate solver success from task fidelity.. The separate faithfulness metric addresses the known failure of counting any solvable PDDL as success; results are on one benchmark.
Either way, bring evidence: a link, or what you saw.
Disagreeing is the most useful thing you can do here. Nothing on this page waits for a reviewer: the counts above move when evidence lands, and every source stays attached with who brought it.
Related claims
- When NL-to-PDDL problem generation is scored by operational acceptance (parses, conforms to domain, a planner finds a plan, VAL validates it, an LLM critic accepts), the score overstates faithfulness to the described task, and the gap widens for PDDL 2.1 temporal and numeric problems where operationally accepted instances can almost never reconstruct the curated reference encoding.Whether the measurement made the finding · reviewed by AI
- On 102 IPC Blocksworld instances with a 180-second budget, a classical planner (Fast Downward) solved 85%, a model planning directly solved 64%, and the same model choosing one action at a time against a PDDL simulator with observe, reset and retry solved 67%, a consistent but small three-point gain at about 5.7 times the token cost, so interactive simulation helps a little and classical search still wins where a formal model exists.Planning a sequence of actions before acting · reviewed by AI
- Prompting with a Task-Method-Knowledge model, which spells out what to do, how, and why in a hierarchical structure, lifted o1 on the obfuscated Random Blocksworld set of PlanBench from about 32% to 97%, and o1-mini from about 9% to 27%, gains far beyond the roughly 1% that zero-shot chain-of-thought gave on PlanBench in earlier work; the smaller model did not benefit on the Mystery variant.Planning a sequence of actions before acting · reviewed by AI
- Used to steer test-time compute, generative LLM judges (7B-70B) rerank candidate responses about as well as outcome reward models, are consistently worse than process reward models at step-level beam search, and their natural-language critiques, the one thing only a judge can produce, did not guide generators to better responses on math, code or instruction following.Judging another's work as well as a careful reviewer would · reviewed by AI
- For long-horizon embodied task planning where a symbolic model defines action preconditions and effects, correcting a violated plan by edits the action model itself implies (inserting the missing navigate or open) raises plan success more per LLM call than returning the violation to the LLM for replanning, and the gain persists for frontier models as well as compact ones.Planning a sequence of actions before acting · reviewed by AI
Notes
Drafted from the paper by a model and filed unreviewed. Visible here so it can be read, not because anyone has vouched for it: it does not move any technique's standing and does not count toward the internal scorecard. Drafted confidence: medium. Falsifier as drafted: Domains where solver success is easy to game so the judge role rewards unfaithful formalizations.