“Minimax identified that keeping the language model head in FP32 precision during reinforcement learning training was critical to closing the gap between the theoretical algorithm and its practical implementation.”