Quantifying the Gain in Weak-to-Strong Generalization