Large-eddy simulations (LES) of a turbulent interfacial gas-liquid flows are described in this paper. The variational multiscale approach (VMS) introduced by Hughes for single-phase flows is systematically assessed against direct numerical simulation (DNS) data obtained at a shear Reynolds number Re-star=171, and compared to LES results obtained with the Smagorinsky model, modified by a near-interface turbulence decay treatment. The models are incorporated in the same pseudospectral DNS solver built within the boundary fitting method used by Fulgosi for air-water flow. The LES are performed for physical conditions allowing low interface deformations that fall in the range of capillary waves of wave slope ak=0.01. The LES results show that both the modified Smagorinsky model and the VMS are capable to predict the boundary layer structure in the gas side, including the decay process, and to cope with the anisotropy of turbulence in the liquid blockage layer underneath the interface. Higher-order turbulence statistics, including the transfer of energy between the normal stresses is also well predicted by both approaches, but qualitatively the VMS results remain overall better than the modified Smagorinsky model. The study has demonstrated that the key to the prediction of the energy transfer mechanism is in the proper prediction of the fluctuating pressure field, which has been found out of reach of any of the LES methodologies. The superiority of the VMS is demonstrated through the analysis of the subgrid transport and exchange terms in the resolved kinetic energy, where it is indeed shown to be self-adaptive with regard to the eddy viscosity. Although VMS is shown to be sensitive to filter scale partition and model constant, the optimal setting can be easily translated in the interface tracking/finite-volume context, which makes it very useful for practical purposes. An important point is that the VMS approach yields very satisfactory results without the need for prescribing an ad-hoc damping function and the required distance to the interface. (c) 2006 American Institute of Physics.