Microsoft has introduced a new set of small language models called Phi-4-reasoning ... by using reinforcement learning and more training data, especially on difficult tests.