MTA-OPSD gives each token its own teacher, +5.9 on Qwen3 math

"Every Token Needs Its Own Teacher," a new arXiv preprint from Yifu Luo and collaborators, argues that on-policy self-distillation, a technique used in post-training large language models, has been using too coarse a teaching signal. Existing OPSD methods either keep one teacher for the whole run or vary teacher exposure across responses. The paper's claim is that the right teacher can change token by token inside a single response.

The authors propose MTA-OPSD, which the abstract calls "the first token-level adaptive OPSD method," building a pool of candidate teachers and letting each token pick one via majority voting and KL-based selection. On Qwen3 models tested on mathematical reasoning, the paper reports "an average gain of up to +5.9 over the base model," and says the method "consistently outperforms existing OPSD approaches."

The abstract reports a single aggregate number and does not name the specific math benchmarks, the Qwen3 parameter sizes, the size of the teacher pool, or the training-time overhead. Code is released at the group's GitHub repository.