In Gradient Boosting, instead of tweaking the instance weights at every iteration like AdaBoost does, it tries to fit the new predictor to the residual errors made by the previous predictor.