The theory of negative prompts

[feiyangsuo]

To my deficient knowledge, positive prompts are embedded by CLIP encoder and sent into the denoiser(Unet) model by cross-attention. During training, guided with the ground truth image, the model can learn the association between positive prompts and the image's content.
My question is: How does stable-diffusion add negative prompts guidance during inference, while there's no negative-prompt-training procedure? And how to put the negative embbedings into the model as condition?
I wish I could just read the code to check what happened with negative prompts, but the code is too hard for me to go through.