What's this about?
This paper shows large language models can be post-trained by distilling from themselves — on-policy self-distillation — with no external supervision at all: no human labels, no reward models, no ground-truth answers.
What exactly is "self-distillation"?
Distillation normally means training a model to imitate a stronger "teacher." Self-distillation removes the separate teacher: the model learns from its *own* better outputs, reinforcing its stronger behaviors without outside signals. "On-policy" means it learns from responses it generates itself during training, keeping the training data aligned with how it actually behaves.
Why it matters
Cutting external supervision out of post-training would make model improvement dramatically cheaper and more scalable — part of a broader push toward AI that improves itself without armies of human labelers.