Researchers have proposed a new metric called the Cross-confounder Robustness Margin (CRoMa) to evaluate the robustness of pathology foundation models. These models can learn shortcuts that undermine generalization across institutions due to non-biological variation introduced during tissue preparation, staining, and scanning. CRoMa measures whether samples sharing the same biology but different confounders are closer together than those with the same confounder but different biology. The authors evaluated CRoMa on 24 encoders across three benchmarks and found that it can identify models more susceptible to shortcuts. --- Why it matters: This matters because pathology foundation models are widely used in medical research, and their robustness is crucial for accurate diagnosis and treatment. The proposed metric, CRoMa, provides a way to evaluate and compare the robustness of these models, which could lead to improved performance and reduced errors in real-world applications. Source: https://arxiv.org/abs/2607.25497