Reinforcement Learning for Locally Checkable Labeling Problems

AAAI 2025 Workshop NeurMAD Submission17 Authors

Keywords: Algorithmic Reasoning, Graph Learning, Reinforcement Learning, Locally Checkable Problems

Abstract: We address the challenge of solving locally checkable labeling (LCL) problems on graphs using machine learning. Unlike prior supervised approaches that depend on ground-truth algorithms or enforce unique solutions, we propose a reinforcement learning framework that requires only verifiers to evaluate correctness. This formulation allows models to learn solution strategies independently, without bias toward specific algorithmic procedures, and inherently supports the discovery of non-unique solutions. We evaluate our method on four fundamental LCL problems, demonstrating its ability to generalize effectively, outperform supervised baselines, and provide a versatile foundation for learning algorithmic reasoning on graphs.

Submission Number: 17