Teacher-to-reward guidance

Made to be adapted.

Training progress linearly interpolates the teacher and reward coefficients. The two weights need not sum to one.

Paper mechanism on synthetic inputs. No benchmark results are implied.