Never Give Up

Michael Noukhovitch, Hamish Ivison, and Nathan Lambert show that RL post-training improves LLMs unevenly: large gains on problems the model can already mostly solve, small gains on hard ones, a "Matthew effect" the authors argue modern RL recipes worsen by spending rollouts on easy prompts. Never Give Up is an adaptive sampling rule that keeps generating for a problem until it finds a solution, reallocating compute toward hard prompts, and lifts hard-problem performance without extra total compute. Filed under the RL-scaling family as a compute-allocation axis.