Researchers have found a way to transfer know-how between AI models without touching a single weight.
A team describes Universal Textual Teaching (UTT), a framework that captures the performance gap between a stronger Teacher model and a weaker Student model, then writes it down as a plain-language document called a Primer. The system builds the Primer through a back-and-forth: the Student attempts a task, a Prompter module turns the resulting mistakes into teaching instructions, the Teacher demonstrates the correct approach, and a Synthesizer folds the validated lesson into the Primer. On the Omni-MATH-2 math benchmark, the technique lifted a Student model's accuracy from 27.6% to 51.7%. On KernelBench, a code-generation benchmark, accuracy rose from 9.4% to 48.6%, and its Fast1 accuracy (a measure of correct-and-efficient solutions on the first try) jumped from 9% to 35%.
That's a meaningful result because standard knowledge distillation bakes the lesson into a model's parameters, locking it to one architecture and training run. A Primer is just text, so it can be read, edited, and reused, including, the researchers found, by other Student models that never took part in making it. That matters most for API-only models, like many commercial LLMs, where you can't get at the weights to distill anything in the first place.
Call it a cheat sheet that works on any test, useful, provided nobody mistakes strong scores on two benchmarks for solved generalization.