Universal Textual Teaching for LLMs
The paper proposes distilling model knowledge into a reusable text “Primer” instead of updating student-model weights.
UTT builds that Primer from cases where a Teacher and Student diverge, then refines it through task attempts, feedback-based teaching instructions, targeted demonstrations, and consolidated lessons. On KernelBench, the reported Student accuracy rises from 9.4% to 48.6%, with Fast1 accuracy moving from 9% to 35%. On Omni-MATH-2, mathematical reasoning accuracy rises from 27.6% to 51.7%. The authors say a Primer made for one Teacher-Student pair can also help other Students that were not part of synthesis. HF Daily Papers' note
UTT builds that Primer from cases where a Teacher and Student diverge, then refines it through task attempts, feedback-based teaching instructions, targeted demonstrations, and consolidated lessons. On KernelBench, the reported Student accuracy rises from 9.4% to 48.6%, with Fast1 accuracy moving from 9% to 35%. On Omni-MATH-2, mathematical reasoning accuracy rises from 27.6% to 51.7%. The authors say a Primer made for one Teacher-Student pair can also help other Students that were not part of synthesis. HF Daily Papers' note
score 4