explaining a basic ml concept

Prompt

How and why does contrastive decoding work? When does it improve results compared to a neutral or negative impact? Explain conceptually / architecture wise which I understand much better than formulas which are difficult to make sense of.