In one sentence: transforming the original scalar continuous features to vectors before mixing them in the main backbone (e.g. in MLP, Transformer, etc.) improves the downstream performance of tabular ...