Sampling a model N times and keeping the best answer. During Composer 2's RL, best-of-16 on a held-out set and best-of-4 on CursorBench rose along with the average, which Cursor's report takes to suggest RL also widened the set of correct solutions.