BroRL: Scaling Reinforcement Learning via Broadened Exploration

The rollout-width axis of RL scaling, from the ProRL team (NVIDIA, with Stanford and UW co-authors): where ProRL's step scaling plateaus after thousands of updates with clearly diminishing returns, BroRL scales the number of rollouts per example into the hundreds to exhaustively broaden exploration — and keeps improving beyond the saturation point of purely sequential scaling.

A key reference for the parallel-vs-sequential trade-off later systematized by IsoCompute's compute-optimal allocation analysis, which cites BroRL as the rollout-width anchor of the literature.