Comment by forrestbao

A smaller model has less capacity and thus is less prone to overfitting, which is a reason of hallucination. Overfitting means that a model cannot adjust its output based on the input that is unseen during training.