Title: Streamlining Inference for Generative Planning Models

Abstract: While generative models have become a potent framework for artificial intelligence planning, their effectiveness is often limited by the distributions present in their training data. Although one strategy involves enhancing solution generation during inference by increasing computational resources, a more resource-efficient method is to refine the inference process itself. This study demonstrates that a modified iteration of the classical Open-Closed List (OCL) search algorithm serves as an effective inference procedure. Our proposed method combines two learned elements: a generative model capable of executing rapid rollouts from intermediate states and a heuristic model designed to rank candidate reasoning paths. The primary contributions of this work involve new mechanisms for controlling exploration and the seamless integration of these learned models into the OCL structure. Evaluations across various combinatorial planning domains reveal that our approach surpasses both neurosymbolic search baselines and traditional solvers in terms of solution quality and computational efficiency.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC