On-policy distillation (OPD) is a training approach in which a teacher model gives token-level feedback to a student model. According to the arXiv abstract, however, common sampled-token updates assign credit only locally, omitting the influence of earlier actions on the feedback that follows. That means the student may not learn which prior decisions led to a good or bad outcome later in the sequence.
To address this, the paper introduces GR-OPD (Group-Relative Sequence Credit). The idea is to evaluate actions in the context of the sequence, rather than treating each token as an independent event. This is intended to give the student a more complete signal about how its choices affect subsequent teacher feedback.
The abstract is short and does not include experimental results or comparisons with other methods. This summary is based solely on the single arXiv source, so the significance of GR-OPD rests on the proposed mechanism rather than demonstrated performance.