What's this about?
Another on-policy-distillation (OPD) paper, this one hardened against teacher models that pass on spurious correlations — where the student learns the teacher's shortcuts and biases instead of the real signal (9↑).
What exactly is a "spurious signal"?
A correlation that happens to hold in the training data but isn't the real cause — e.g. a model that "detects cows" by looking for grass. In distillation, if the teacher relies on such shortcuts, the student faithfully copies them. This paper makes the student aware of and resistant to that.
Why does it matter?
OPD variants have been a daily fixture this week (SAF-OPD, DAPD, Weak-to-Strong OPD). The volume signals that post-training distillation recipes — not architectures — are where the fast competitive iteration is happening right now, and "don't inherit the teacher's bad habits" is a natural next problem.