Prompt for the base model

Hi,

Thanks for the nice work.

I just have one question about the [prompt](https://github.com/turningpoint-ai/VisualThinker-R1-Zero/blob/cd71978473449833120f0daff3ff265309716ae1/src/open-r1-multimodal/src/open_r1/grpo.py#L159) for the base model. It seems that the base model is not told to output `<think>` and `<answer>`. How does it know to follow the format?