GRACE is a new method for optimizing reasoning data curation by scoring individual steps in reasoning traces, leading to more efficient post-training performance with significantly less data. This approach has shown to enhance model performance while reducing data requirements, indicating a potential shift in data utilization strategies in AI development.