We initially train without format rewards. As accuracy rewards are conditional on using the (tool calling) format correctly, we believe models should learn the format without further intervention. This notion was further supported by our base models format pass rate being extremely high (>0.95) for our group size, ensuring that there is a training signal even in step zero.

Unfortunately, we observed persistent regressions in format adherence in later stages of training, leading us to introduce format rewards. In future work, we want to study using an SFT model with great format adherence instead.