The authors explore the role of comparison in improving graph fluency. The ability to use graphs fluently is crucial for STEM achievement, but graphs are challenging to interpret and produce because they often involve integration of multiple variables, continuous change in variables over time, and omission of certain details in order to highlight central higher-order relations. Can comparison facilitate graph fluency by focusing learners on the relations between multiple variables? Furthermore, does the comparison of highly similar graphs facilitate performance to a greater degree than comparison less similar cases? This experimental research was conducted in the Language and Cognition Laboratory at Northwestern University. In Experiment 1, 22 Northwestern undergraduates participated to fulfill a course requirement. In Experiment 2, 64 Northwestern undergraduates participated to fulfill a course requirement. The results suggest that comparison of highly similar examples promotes understanding and fluency with graphical representations of stock-and-flow scenarios, compared to a situation in which the same examples are not compared (Experiment 1). This finding supports prior work that demonstrates the benefits of comparing examples, and extends it to a novel domain--graphical representations. The authors suggest that pedagogical methods that assume that learners will abstract principles from single examples or that they will spontaneously draw comparisons across examples are likely to fall well short expectations. They suggest that one aim for instruction should be not simply to provide cases but to encourage active comparison of examples. The results of Experiment 2 are somewhat less clear, but are still valuable for what they can tell them about the range of permitted variation between cases. Experiment 2 showed no difference in performance between those who compared high-similarity examples and low-similarity examples. One possible explanation for this finding is that the paired cases in the low-similarity condition were, in fact, quite similar to one another: while the trajectories of the flows and stocks may have differed, the perceptual features of the graphs were identical (e.g., the inflow line was always blue). Prior empirical work on comparison has shown that when two examples share surface features that are consistent with deeper relational commonalities, identifying these relational commonalities becomes much easier (consider a volleyball and a soccer ball vs. a volleyball and a football) (e.g., Gentner & Medina, 1998). In these studies, it is possible that the low-similarity examples were similar enough to facilitate high-quality comparisons, which in turn would lead to better learning. Once the authors finish coding participants' similarity and difference listings, they can assess whether both the high-similarity and low-similarity groups generated equally good comparisons. One limitation of their findings concerns the generalizability of the results. The current studies were conducted in a laboratory setting with undergraduates at an academically rigorous university. Whether the authors replicate these results with other populations in other settings is an open question. However, clarifying the conditions under which comparison can help or hurt learning of graphical representations in a highly controlled environment is an important first step in developing useful classroom interventions. (Contains 1 table and 1 figure.)