Aggregation Hides Out-of-Distribution Generalization Failures from Spurious Correlations

Benchmarks for out-of-distribution (OOD) generalization frequently show a strong positive correlation between in-distribution (ID) and OOD accuracy across models, termed "accuracy-on-the-line." This pattern is often taken to imply that spurious correlations-correlations that improve ID but reduce OOD performance-are rare in practice. We find that this positive correlation is often an artifact of aggregating heterogeneous OOD examples. Using a simple gradient-based method, OODSelect, we identify semantically coherent OOD subsets where accuracy on the line does not hold. Across widely used distribution shift benchmarks, the OODSelect uncovers subsets, sometimes up to over half of the standard OOD set, where higher ID accuracy predicts lower OOD accuracy. Our findings indicate that aggregate metrics can obscure important failure modes of OOD robustness. We release code and the identified subsets to facilitate further research. 2656 0.95 0.95 0.95 0.61 0.61 0.61 PACS Sketch 10 -0.48 (0.14) 0.37 (0.18) 0.17 (0.20) -0.05 (0.19) 0.39 (0.18) -0.06 (0.19) PACS Sketch 20 -0.33 (0.16) 0.71 (0.11) 0.17 (0.20) -0.41 (0.15) 0.34 (0.19) 0.00 (0.20) PACS Sketch 50 -0.47 (0.14) 0.84 (0.07) 0.00 (0.20) 0.23 (0.19) 0.47 (0.17) -0.12 (0.19) PACS Sketch 100 -0.33 (0.16) 0.73 (0.11) 0.11 (0.20) 0.29 (0.19) 0.70 (0.12) -0.08 (0.19) PACS Sketch 250 -0.30 (0.17) 0.79 (0.09) 0.19 (0.20) 0.35 (0.19) 0.70 (0.12) -0.04 (0.19) PACS Sketch 500 -0.22 (0.18) 0.83 (0.07) 0.28 (0.19) 0.41 (0.18) 0.69 (0.12) 0.07 (0.20) PACS Sketch 750 0.01 (0.20) 0.82 (0.08) 0.33 (0.19) 0.43 (0.17) 0.65 (0.13) 0.18 (0.20) PACS Sketch 800 0.05 (0.20) 0.81 (0.08) 0.33 (0.19) Dataset OOD N Pearson R Spearman ρ Ours Random Hard Ours Random Hard PACS Sketch 1500 0.21 (0.20) 0.82 (0.08) 0.41 (0.18) 0.42 (0.18) 0.68 (0.12) 0.39 (0.18) PACS Sketch 1750 0.24 (0.19) 0.82 (0.08) 0.46 (0.17) 0.42 (0.18) 0.67 (0.12) 0.43 (0.17) PACS Sketch 2000 0.29 (0.19) 0.82 (0.08) 0.49 (0.16) 0.48 (0.17) 0.66 (0.13) 0.48 (0.17) PACS Sketch 2250 0.48 (0.17) 0.81 (0.08) 0.54 (0.16) 0.52 (0.16) 0.67 (0.12) 0.51 (0.16) PACS Sketch 2500 0.56 (0.15) 0.81 (0.08) 0.58 (0.15) 0.54 (0.16) 0.67 (0.12) 0.56 (0.15) PACS Sketch 2750 0.62 (0.14) 0.81 (0.08) 0.63 (0.13) 0.56 (0.15) 0.67 (0.13) 0.61 (0.14) PACS Sketch 3000 0.67 (0.12) 0.81 (0.08) 0.68 (0.12) 0.58 (0.15) 0.67 (0.12) 0.64 (0.13) PACS Sketch 3250 0.71 (0.11) 0.81 (0.08) 0.72 (0.11) 0.61 (0.14) 0.67 (0.13) 0.66 (0.13) PACS Sketch 3500 0.75 (0.10) 0.81 (0.08) 0.76 (0.10) 0.63 (0.14) 0.66 (0.13) 0.66 (0.13) PACS Sketch 3750 0.81 (0.08) 0.81 (0.08) 0.79 (0.09) 0.67 (0.13) 0.66 (0.13) 0.67 (0.13) PACS Sketch 3929 0.81 0.81 0.81 0.67 0.67 0.67