Daily AI Paper

Large language models often get better with more data, but this paper asks a surprising question: how little data can you use and still get most of the benefit? The authors study on-policy distillation, where a student model generates its own rollouts and a teacher model provides detailed token-level guidance. Instead of training on a full dataset, they try the extreme case of just one training example. Surprisingly, that single example can keep improving for hundreds of steps and recover much of the performance of full-data training across different tasks and model families. The key insight is state coverage: even one query can drive the student through a wide range of internal states, creating rich supervision. Adding a few semantically different examples expands that coverage further, until a small set can match full-data results. The bigger message is that this method is not starved for data so much as slowed by the learning process itself, which points to a new direction for making post-training far more efficient.