Only because each individual algorithm is slightly overfit. Blending it all together keeps the good predictive part and reduces the overfit noise.