The authors identify a Matthew Effect in reinforcement learning for LLMs, where easy problems receive large accuracy gains during training while hard problems see almost none, and argue this stems from standard methods wasting rollout budget on already-easy problems rather than from raw compute limits. They propose Never Give Up (NGU), an adaptive sampling method that, under asynchronous RL, keeps generating samples for a given problem until a correct solution is found, naturally spending less compute on easy problems and more on hard ones. On the Deepscaler math benchmark and the Manufactoria coding benchmark, NGU improves performance per unit of compute, with the largest gains on the hardest problems, enabling progressive solving of complex coding tasks that standard methods fail on.
