In this talk I will discuss a series of controlled numerical experiments modelling the formation and evolution of high-redshift galaxies in an explicit cosmological framework. As these simulations reach a spatial resolution of 10pc, enabling us to partially capture the structure of the ISM, I investigate the validity of the most commonly adopted prescription for star formation which is based on a simple density threshold and a Kennicutt/Schmidt law, and discuss a more plausible alternative model on these scales. I will show that such an apparently innocuous modification can have a dramatic impact on the inferred properties of galaxies (stellar mass, rotation velocity etc) and help resolve longstanding issues with supernova feedback.