Study: On-Policy Distillation Transfers Reasoning Behavior, Lineage Matters More Than Teacher Scores

New research on On-Policy Distillation finds that student LLMs copy the teacher's reasoning behavior rather than answers or knowledge. A weak teacher from the same model lineage can outperform a stronger teacher from a different lineage, suggesting lineage matters more than benchmark scores.

2026-08-30 ~ 2026-08-30 · 3 related posts