A new arXiv paper, Branching On-Policy Distillation, tackles a known weakness in on-policy distillation (OPD). In OPD, a teacher model supervises a student on trajectories the student generates itself. The authors point out that at each decision point, a single trajectory only shows one possible continuation, so under a fixed sampling budget the teacher's supervision covers only a narrow slice of the student's behavior.
The paper proposes "branching" as a remedy, presumably to explore multiple continuations at each decision and thus widen supervision coverage. However, the abstract provided in the source is truncated mid-sentence, so the exact mechanism, experimental setup, and findings are not available in this excerpt.
Because the source is a single arXiv announcement and the abstract is incomplete, this summary can only reflect the problem statement and the general direction of the proposed solution. No results or comparisons are reported in the available text.