Non-negative coefficients in mathematical models ensure that the relationships between variables are either positive or neutral, preventing any decrease in the dependent variable as the independent variable increases. This constraint is particularly useful in fields like economics and machine learning, where it aligns with realistic assumptions about the nature of the relationships being modeled.