The main objective of applying dimensionality reduction techniques in Spark ML workflows, and in machine learning broadly, is to decrease the number of input features. These techniques strive to preserve the most significant data information while eliminating less important or redundant features. Benefits include faster model training times, mitigation of the curse of dimensionality, and better generalization by concentrating on the most pertinent information. Contrary to increasing model complexity or adding irrelevant features, the focus is on simplifying the data's representation without losing its key attributes.