Researchers compared mathematical training data for Large Language Models (LLMs) using two approaches: topic and reasoning approach. They found that using a reasoning approach, where sources share the target's method but change the topic, resulted in better transfer after fine-tuning than using a topic-based approach.