An engineer is using a pre-trained language model to solve multi-step mathematical word problems. The model consistently provides the correct final numerical answer but fails to output the intermediate calculations and logical steps it used to arrive at the solution. This makes it impossible to verify the model's reasoning process. What is one specific adjustment the engineer can make to the text generation process itself (without retraining the model) to encourage the model to produce a more detailed, step-by-step output? Explain the rationale behind your suggested adjustment.