UTT distills teacher LLM knowledge into text, lifting math accuracy 27.6% to 51.7%
Zhanyi Lu · hf · 2026-10-12
Researchers from Westlake University introduce Universal Textual Teaching (UTT), a parameter-update-free knowledge distillation framework that converts teacher-student knowledge gaps into an interpretable, reusable natural-language artifact called a Primer.
Method
- Identifies representative gap cases via paired evaluations
- Iteratively refines the Primer through multi-role LLM interactions: Student attempts tasks, Prompter turns feedback into teaching instructions, Teacher provides targeted demonstrations, Synthesizer consolidates validated lessons
Results
- Raises student accuracy on KernelBench from 9.4% to 48.6% (Fast1: 9% to 35%)
- Improves math reasoning accuracy from 27.6% to 51.7% on Omni-MATH-2
- Outperforms representative prompt engineering and parameter-based KD methods
- A Primer synthesized for one teacher-student pair generalizes to other students not involved in its synthesis
Especially useful for API-only or costly-to-train models.
More from Research
- RAIL principles paper argues neurosymbolic AI is key to trustworthy systems, soon in CACM — luislamb · 2026-10-12
- If mathematicians value understanding, why won't they read AI-generated proofs? — LucaAmb · 2026-10-12
- Nature: The century-old debate over whether adult human brains grow new neurons remains unresolved — AnnaCiaunica · 2026-10-12
- Half of a study's 1,168-row dataset are duplicated observations, replication watchdog finds — RexDouglass · 2026-10-12
- Replication watchdog admits failure after independent researcher flags data irregularities — RexDouglass · 2026-10-12
- Brazilian data looks randomly generated: report cites bizarre correlations, implausible effect sizes — RexDouglass · 2026-10-12