Westlake researchers distill LLM knowledge into text 'Primers'

A text file, generated by a four-way interaction between an LLM Student, Prompter, Teacher, and Synthesizer, is enough to lift a weaker model's benchmark scores by tens of points without updating any of its weights. That is the claim of Universal Textual Teaching, a paper posted to arXiv this month by Zhanyi Lu and Huan Wang of Westlake University.

The method writes the Teacher-Student knowledge gap into "a textual, interpretable, and reusable natural-language artifact called Primer." The loop that produces it is spelled out plainly in the abstract: "the Student attempts each task, the Prompter turns evaluation feedback into a teaching instruction, the Teacher provides a targeted demonstration, and the Synthesizer consolidates validated lessons."

The reported numbers are large. On KernelBench, a code generation benchmark, UTT "raises the Student's accuracy from 9.4% to 48.6% and Fast1 accuracy from 9% to 35%." On Omni-MATH-2, mathematical reasoning accuracy moves from 27.6% to 51.7%. The authors also claim UTT "performs better than representative prompt engineering and parameter-based KD methods."

The motivation is stated flatly: "most methods require training the Student parameters, thereby binding the distilled knowledge to a specific architecture and checkpoint." A Primer, by construction, is not.

This is a two-author preprint on two benchmark families. The abstract states that a Primer "can generalize to other Students that do not participate in the synthesis" but publishes no cross-model transfer matrix in the text available.