Tag
This paper proposes property-guided LLM program synthesis, using counterexample-guided inductive synthesis (CEGIS) to provide concrete feedback when a candidate program fails a formal property, reducing the number of generations and evaluation costs. Applied to PDDL planning domains for synthesizing direct heuristic functions, the method outperforms prior approaches, generating seven times fewer programs and solving more tasks without search.
This paper presents the first systematic zero-shot evaluation of frontier large language models as goal recognizers on classical PDDL planning benchmarks, finding that some models scale with evidence while others rely on world-knowledge priors regardless of observation accumulation.
This paper introduces a neural network architecture that learns lifted action schemas from fully observed state traces with unobserved action arguments, aiming to enable robust learning of planning domains for neuro-symbolic models.