In this paper, a molecular theory of self-diffusion coefficient is developed for polymeric liquids (melts) on the basis of the integral equation theory for site–site pair correlation functions, the generic van der Waals equation of state, and the modified free volume theory of diffusion. The integral equations supply the pair correlation functions necessary for the generic van der Waals equation of state, which in turn makes it possible to calculate the self-diffusion coefficient on the basis of the modified free volume theory of diffusion. A random distribution is assumed for minimum free volumes for monomers along the chain in the melt. More specifically, a stretched exponential is taken for the distribution function. If the exponents of the distribution function for minimum free volumes for monomers are chosen suitably for linear polymer melts of N monomers, the N dependence of the self-diffusion coefficient is N − 1 for the small values of N , an exponent predicted by the Rouse theory, whereas in the range of 2.3 ≲ ln N ≲ 4.5 the N dependence smoothly crosses over to N − 2 , which is reminiscent of the exponent by the reptation theory. However, for ln N ≳ 4.5 the N dependence of the self-diffusion coefficient differs from N − 2 , but gives an N dependence, N − 2 − δ ( 0 < δ < 1 ) , consistent with experiment on polymer melts in the range. For polyethylene δ ≈ 0.48 for the parameters chosen for the stretched exponential. Because the stretched exponential function contains undetermined parameters, the N dependence of diffusion becomes semiempirical, but once the parameters are chosen such that the N dependence of D can be successfully given for a polymer melt, the temperature dependence of the self-diffusion coefficient can be well predicted in comparison with experiment. The theory is satisfactorily tested against experimental and simulation data on the temperature dependence of D for polyethylene and polystyrene melts.