arXivSimran Kaur, Narutatsu Ri, Yinghui He, Liam Fowl, Sanjeev AroraMon, Jul 6, 2026, 8:01 AM PDT
score 17.0
Self-distillation harms AI models that think step-by-step
Original: Rethinking On-Policy Self-Distillation for Thinking Models
Source: arxiv.org ↗
Writing ELI5 summary…