Research reveals that large reasoning models outperform base models mainly due to a small number of early decision tokens related to planning. A new intervention method can enhance base model performance by selectively utilizing reasoning model insights at critical points, potentially leading to significant improvements in reasoning tasks.