The DeepSeek-R1 technical report indicates that this training approach can cause the model to spontaneously develop complex reasoning behaviors such as self-correction and re-thinking, without requiring explicit human annotation guidance