Autonomy Tea Talk

"Are Domain Generalization Benchmarks with Accuracy on the Line Misspecified?"

Spurious correlations are unstable statistical associations that hinder robust decision-making. Conventional wisdom suggests that models relying on such correlations will fail to generalize out-of-distribution (OOD), particularly under strong distribution shifts. However, a growing body of empirical evidence challenges this view, as naive empirical risk minimizers often achieve the best OOD accuracy across popular OOD generalization benchmarks. In light of these counterintuitive results, we propose a different perspective: {\em many widely used benchmarks for assessing the impact of spurious correlations on OOD generalization are misspecified.} Specifically, they fail to include shifts in spurious correlations that meaningfully degrade OOD generalization, making them unsuitable for evaluating the benefits of removing such correlations. We establish sufficient—and in some cases necessary—conditions under which a distribution shift can reliably assess a model's reliance on spurious correlations. Crucially, under these conditions, we provably should not observe a strong positive correlation between in-distribution and out-of-distribution accuracy—often referred to as accuracy on the line. Yet, when we examine state-of-the-art OOD generalization benchmarks, we find that most exhibit accuracy on the line, suggesting they do not effectively assess robustness to spurious correlations. Our findings expose a limitation in evaluating algorithms for domain generalization, i.e., learning predictors that do not rely on spurious correlations. Our results highlight the need to rethink how we assess robustness to spurious correlations.