Original title: Rethinking On-Policy Distillation of Large Language Models II: One Training Example
Zixuan Fu, Bingxiang He, Yuxin Zuo, Haohuan Huang +9 · September 2026
Teaching a smaller AI model using a larger teacher model usually requires huge datasets. In this process, called on-policy distillationTraining a smaller model by having it generate its own practice examples and learn from a teacher model., the student generates its own text attempts, and the teacher corrects each word. Researchers tested what happens if you train the student on just one single prompt. Surprisingly, the student kept improving for hundreds of steps and captured most of the gains seen with full datasets across various tasks and model families. It works because the student's own varied attempts quickly visit many possible text paths, known as states. Just one prompt reaches 71.5% of the states seen in full training, mostly in the first 100 steps. Using 16 distinct prompts reaches 98.9% and matches full training. The student slows down when learning these paths, taking hundreds of steps even on a fixed set. The method actually gets plenty of data, but struggles to absorb it quickly.
Imagine learning to play a musical instrument by having a teacher guide you through pieces. You might think you need to practice many different songs to improve, but mastering just one complex piece can teach you most techniques. In this paper, training an AI with one example is like mastering one song; the example involves many decisions that cover various skills, and the AI absorbs these skills slowly through repeated practice.
This finding could make training AI models much cheaper and faster, since developers might not need massive datasets. It shifts focus from collecting data to improving learning algorithms.
The abstract does not specify the exact tasks, models, or performance metrics used. It is limited to on-policy distillation and may not apply to other training methods. The focus is on data minimization but does not address broader implications like computational cost.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Rethinking On-Policy Distillation of Large Language Models II: One Training Example — Zixuan Fu, Bingxiang He, Yuxin Zuo et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.