How to improve the $R^2$ score on an autoencoder model when the loss(KLDivergence) and validation MAE is giving desired scores?

Question

I have been training an autoencoder for data with 25k feats and 1k data points. The $R^2$ score is coming negative on every epoch, and around -27 on both train and test sets, although the MAE is 0.7 on both the train and test sets.