Larger RL batches only help when throughput beats the sample penalty

Ziniu Li highlights a study finding that learning-rate retuning is the difference between a 29% GRPO speedup and a 42% slowdown. The underlying preprint was first submitted in August.