Efficient Test-time Inference for Generative Planning Models

2026-05-30

AI Summary

Existing generative planning models are constrained at test time by the distribution of their training data, and improving solution quality typically incurs substantial computational overhead during inference. This work proposes an enhanced open-closed list (OCL) search framework that, for the first time, deeply integrates learned generative models with heuristic functions. The approach leverages the generative model to efficiently expand intermediate states while incorporating a novel exploration control mechanism to guide the reasoning trajectory. This integration substantially enhances both inference efficiency and solution quality at test time, consistently outperforming neural-symbolic search baselines and classical solvers across multiple combinatorial planning benchmarks.

Abstract

Generative models have emerged as a powerful paradigm for AI planning, yet their performance remains constrained by the training data distribution. One approach is to improve generated solutions during inference by scaling test-time compute. A more efficient alternative is to optimize the inference process itself. In this paper, we show that a modified version of a classical Open-Closed List (OCL) search provides just such an efficient inference procedure. Our algorithm synergizes two learned components: a generative model that performs fast rollouts from intermediate states and a heuristic model that prioritizes among candidate reasoning paths. Key contributions include novel exploration control mechanisms and integration of learned models within the OCL framework. Across multiple combinatorial planning domains, our approach outperforms both neurosymbolic search baselines and classical solvers in computational efficiency and solution quality.

Problem

• generative planning
• test-time inference
• computational efficiency
• combinatorial planning
• reasoning paths

Innovation

• generative planning
• test-time inference
• Open-Closed List search
• learned heuristics
• exploration control