While there're interesting findings here, https://arxiv.org/pdf/2502.03373 (also with a lot of good findings) suggested some contradicting theory on the critical mass of training process/data for the sake of reasoning capability.