On-policy distillation (OPD) improves a student language model by training it on trajectories the student itself generates, with supervision from a stronger teacher. This approach is designed to keep the student's outputs close to its own distribution while pushing quality toward the teacher's level.
The new arXiv paper (2610.21195) flags a complication: when the teacher is conditioned on student-generated prefixes, it can favor certain outputs — the abstract's sentence trails off before specifying the exact failure. To counter this, the authors propose condition-contrastive extrapolation, a technique named in the title.
Since there is only one source, there are no conflicting accounts to compare. The abstract as provided does not include implementation details or experimental results.