# Predicting a variably-placed value in a vector

I have $$m$$ vectors in $$\mathbb{R}^n$$, where $$m >> n$$, and I want to train a model to impute a value $$x_i$$ in $$\mathbf{x}$$, where $$1 \leq i \leq n$$ (and can vary by vector).

For instance, I may have the vector $$\mathbf{x} = [12.34, 12341, 234, 21.5643, \cdot, 42.2]$$, where I want to predict the second-to-last value.

I am guessing in theory this is similar to the masked word problem and approaches similar to BERT may work (though modified for a continuous domain), but I am wondering if their is an existing model architecture that does this?