We propose Self-Denoising Monte Carlo Annotation (SCAN), an efficient Process Reward Model (PRM) data synthesis and noise-tolerant learning framework.