The Model-Assisted Reasoning in Science (MARS) project seeks to promote model-centered instruction as a means of improving middle-school science education. As part of the evaluation of the sixth-grade curriculum, performance of MARS and non-MARS students was compared on a curriculum-neutral task. Fourteen students participated in structured interviews in which they experimented with a balance apparatus that provided three manipulable variables (two affected balance, one was a non-causal distractor variable). Although both groups were equally able to identify and test variables, all MARS students discovered a quantitative rule to describe the operation of the balance, whereas only one non-MARS student did so. MARS students discovered this numerical relationship through experimentation, regardless of their scientific reasoning profile (i.e. theory-generating, theory-modifying, or theory-preserving). The critical components of MARS instruction that may foster the ability to flexibly coordinate theory and evidence include multiple opportunities to draw conclusions from data and an emphasis on the successive refinement of models.