The authors combine reinforcement learning with verifiable rewards, which scores completed answers for correctness, with on-policy self-distillation.
Filtering teacher guidance by gradient alignment improves reasoning model training
Li and colleagues report that selectively retaining token-level teaching signals improves mathematical and logical reasoning, with relative gains of up to 9.5% on difficult mathematics tasks.
Chinese Tech
Shijie Li · Yilin Gao · Tieyuan Chen · Yun Luo · Weiyao Lin · Hang Yu
Shanghai Jiao Tong University · Tencent
Research Digest··2 min read
Li and colleagues introduce GIFT, a training method that selects which teacher signals a reasoning model should follow while learning from correctness rewards.
§