When a language model tries to teach itself without human-provided answers, it often fails on hard problems because its own attempts disagree. Relying on the most common answer among attempts is wrong about a third of the time on challenging questions.
The proposed method, Stone, instead has the model write simpler related problems that it can solve consistently. The model solves each of these easier problems multiple times and keeps only those it gets right every time. The teacher then sees these worked solutions, while the student only sees the original problem.
On five math benchmarks, Stone improved accuracy by 13.2 and 22.7 points on two models without thinking, about two and a half times the gain from methods that use ground-truth answers. This means models can improve without labels by learning from simpler problems they already handle well.
