This paper asserts that evaluating generative AI requires more than benchmark scores—it demands a science of measurement that aligns metrics with real-world uses and evolves over time. It articulates how evaluation tools must shift from static, one-off tests to dynamic instruments embedded in institutional ecosystems. In doing so, it charts a path for generative AI evaluation to support trustworthy claims about capability, safety, and deployment in society.

This paper reframes AI evaluation as a science of measurement. It proposes a validity framework that determines when evaluation results, such as benchmark scores, truly support claims about underlying capabilities like reasoning or intelligence. The same framework extends to assessing AI risks, offering a principled basis for reliable, meaningful evaluation.

This paper introduces ImageNot, a dataset intentionally designed to differ from ImageNet while matching its scale, to test whether the deep learning revolution sparked by ImageNet was benchmark-specific. Re-training the full trajectory of architectures from that era from scratch on ImageNot reproduces the same pattern of progress, showing that the common practice of iterative development on public benchmarks can yield externally valid methodological advances—even though similar practices often lead to false discoveries in other scientific fields.

This paper revisits the design of domain generalization (DG) benchmarks and shows that the widely observed “Accuracy-on-the-Line” pattern—where in-domain and out-of-domain accuracy are strongly and positively correlated—reveals a deeper misspecification. It argues that current DG datasets capture simple or narrow distribution shifts that miss the complexity of real-world variation, such as geographical differences in COVID-19 diagnosis. The paper provides diagnostic analyses and guidance for constructing DG evaluations that genuinely test generalization beyond training domains.

This paper shows that common aggregation practices in evaluating out-of-distribution (OOD) generalization can mask systematic failures caused by spurious correlations. When performance is averaged across heterogeneous subsets, models can appear robust even while failing sharply on specific environments or attributes. The paper introduces diagnostic methods to disentangle these hidden failures, showing that finer-grained evaluation is essential for detecting and mitigating OOD brittleness.

This paper develops a causal regularization method that is effective across a broad class of domain generalization settings previously treated individually, defined by Reichenbach’s Common Cause Principle. The method yields representations that generalize across domains even when spurious factors are unobserved, improving robustness to distribution shifts across well-specified state-of-the-art benchmarks.

This paper studies domain adaptation under latent subgroup shift, where unobserved subpopulations drive changes in both features and labels. It develops a causal framework showing that the Bayes-optimal target predictor is identifiable using observable concepts and proxy variables, even without target labels. The resulting method adapts across hidden subgroups that confound model performance, improving reliability in critical real-world applications.

This paper studies domain adaptation under latent shift, where unobserved confounders jointly affect inputs and labels, invalidating standard covariate- or label-shift assumptions. It develops a proxy-based adaptation method grounded in proximal causal inference, using observable proxies or access to multiple environments to mitigate the effects of latent distribution shift and align domains without direct supervision.

When environmental cues are insufficient, labeling data becomes the only way to adapt. We study this single-round adaptation setting—where only one batch can be labeled—and show that using prediction variability to guide selection improves long-term adaptation efficiency.