In simpler terms, the variable that gives the minimum AIC when dropped, is dropped for the next iteration, until there is no significant drop in AIC is noticed. The AIC of the models is also computed and the model that yields the lowest AIC is retained for the next iteration. In each iteration, multiple models are built by dropping each of the X variables at a time. It performs multiple iteractions by droping one X variable at a time. It iteratively searches the full scope of variables in backwards directions by default, if scope is not given. In stepwise regression, we pass the full model to step function.
