Large language models fed low-quality data skip steps in their reasoning process.