The softmax value y is a linear combination of the vectors you're attending over: y = a1v1 + a2v2 + ... + an*vn where a_i >= 0 and sum(a_i) = 1.

Then y is a convex combination of the v_i, and sits in the convex hull of the v_i.