Reproduce “Every Token Needs Its Own Teacher: Multi-teacher Adaptive On-policy Self-distillation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yifu Luo, Chuang Ma, Zeyu Chen, Enkelejda Kasneci, Qi Chen, Gjergji Kasneci
MTA-OPSD trains mathematical reasoning by letting a student model generate responses while reference-conditioned versions of the same model provide token-level supervision. It selects a teacher for each token using majority agreement about the sampled token, then KL distance. On three math benchmarks, Qwen3-1.7B scores a 57.0 average, above OPSD and Paint, within the evaluated Qwen3 models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.