On-policy distillation (OPD) supervises a student language model on trajectories sampled from its current policy, but assigns equal credit to response tokens with unequal supervision value. Selective OPD addresses this limitation by allocating supervision non-uniformly across response tokens according to their estimated training value.
arXiv:2608. 03632v1 Announce Type: new Abstract: On-Policy distillation (OPD) transfers teacher capabilities by supervising student-sampled trajectories with dense token-level teacher signals.
By Yinuo Jiang, Yongjie Ye, Zhou Tao, Xiang Zhuang, Qiang Zhang, Huajun Chen, Tiankai Li
On-Policy distillation (OPD) transfers teacher capabilities by supervising student-sampled trajectories with dense token-level teacher signals. Recent selective OPD methods improve this process by prioritizing signals that are confident, informative, or learnable.
COFFEQA introduces a counterfactual-guided few‑shot distillation method for open‑book question answering, leveraging minimally modified contexts that prompt the teacher model to produce alternative answers. By pairing factual samples with validated counterfactuals—filtered using a new Context Following Difficulty metric—COFFEQA constrains the student’s decision boundary to better match the teacher’s. Experiments on multiple QA benchmarks show that COFFEQA consistently outperforms factual‑only distillation, achieving significant gains in EM and F1 scores with a limited few‑shot budget.
By Kavinda Kehelella, Sanghamitra Dutta
arXiv:2608. 07935v1 Announce Type: new Abstract: On-policy self-distillation (OPSD) adapts a language model by distilling guidance from a frozen teacher on trajectories sampled from the student.
By Meilin Yang (Renmin University of China, Beijing, China), Zixuan Ding (Renmin University of China, Beijing, China), Jianhao Nie (Renmin University of China, Beijing, China), Weite Zhang (Renmin University of China, Beijing, China), Yuxin Zhang (Renmin University of China, Beijing, China), Zhiming Shao (Renmin University of China, Beijing, China), Li Yu (Renmin University of China, Beijing, China), Zhe Fu (Renmin University of China, Beijing, China)
arXiv:2604. 10688v2 Announce Type: replace-cross Abstract: On-policy reinforcement learning has become the dominant paradigm for reasoning alignment in large language models, yet its sparse, outcome-level rewards make token-level credit assignment notoriously difficult.
By Binbin Zheng, Xing Ma, Yiheng Liang, Jingqing Ruan, Xiaoliang Fu, Kepeng Lin, Benchang Zhu, Ke Zeng, Xunliang Cai