The authors frame on-policy distillation as a credit assignment problem.
Sequence-level credit assignment improves on-policy distillation of reasoning models
GR-OPD aggregates token-level teacher feedback into group-calibrated trajectory credit, beating vanilla on-policy distillation across five Qwen3 settings and six math benchmarks.
Chinese Tech
Yushu Wang · Renkai Yang · Jingcao Xu · Zili Geng · Haihong Tang · Yaohui Jin · +1 more
Shanghai Jiao Tong University · Alibaba Group
Research Digest··2 min read
The authors propose GR-OPD, a method that aggregates token-level feedback from a teacher into sequence-level credit for on-policy distillation of large language models.
§