This article reviews the research literature and theory concerned with accuracy of judgments in romantic relationships. We initially propose a model of cognition in (romantic) relationships that distinguishes between 2 forms of accuracy: mean-level bias and tracking accuracy. We then report the results of meta-analyses of research on heterosexual, romantic relationships, which used external benchmarks and reported levels of tracking accuracy (98 studies) and/or mean-level bias (48 studies). The results revealed robust overall effect sizes for both tracking accuracy (r = 0.47) and positive mean-level bias (r = 0.09). As expected, the effects were substantial and positive for tracking accuracy across 6 judgmental categories, whereas signed mean-level bias was negative for the interaction attributions (e.g., love, communication). The results showed, as expected, that these 2 forms of accuracy were independent--the 2 kinds of effect size derived from the same set of 38 studies were uncorrelated. As expected, gender, relationship length, and relationship evaluations moderated mean-level bias across studies but (unexpectedly) not for tracking accuracy. In the Discussion we evaluate the prior model in light of the findings, other research, moderating variables (such as self-esteem), the role of projection, the early stages of mate selection, metacognition, and the rationality and nature of motivated cognition. We conclude that our model, findings, and analyses help to resolve the apparent paradox that love is both riven with illusions and rooted in reality, and support both evolutionary and social psychological approaches to understanding cognition in romantic relationships. (Contains 17 footnotes, 3 tables, and 5 figures.)