Without the chain of thought approach, language models would simply predict the most likely token without sufficient context.