#xor

The training set has 10 copies of each of $(x_1, x_2, y) = (1, 1, 1)$, $(0, 1, 0)$, $(1, 0, 0)$ and $(0, 0, 1)$, so $y = 1$ exactly when $x_1 = x_2$. A logistic regression with a bias term is built by greedy forward selection: starting from the bias alone, a candidate feature is added only if it strictly increases the training log-likelihood. The candidates are $x_1$, $x_2$ and $x_1 x_2$. Which of them can be added first?