Back in the late 1990s when I first was learning HB, I also assumed that the average of the individual-level partworths should be equal to the alpha estimates. They are not. The alpha estimates are updated in each iteration based on the individual-level betas and the upper-level covariance matrix. You will find that the average of the individual-level betas are very highly correlated with the alpha (population-level) part-worth estimates. But, not exactly equal.