Slow And Steady

In a gradient boosting model (GBDT), each new tree is fit to the negative gradient of the loss (pseudo-residuals) from the current ensemble, and its output is scaled by a learning rate before being added. If you lower the learning rate from 0.3 to 0.01 while keeping n_estimators fixed, what happens?

The model overfits harder because smaller steps let it memorize noise in the residuals

The model underfits because each tree's correction to the residuals is now too small to be compensated for without adding more trees

Training becomes faster since each tree needs less precision to fit the residuals

The model is unaffected because learning rate only rescales feature importances, not predictions