reproducing the alphazero-like "model learns to reason on its own without supervised fine-tuning" phenomenon that deepseek-r1-zero exhibited