The "black-box" nature of representations, combined with the known susceptibility of networks to learn spurious correlations [

This gave birth to learning-augmented algorithms, which use these predictions to go beyond the standard long-standing worst-case limitations. The design of such algorithms requires establishing good tradeoffs between consistency and robustness, i.e. having improved performance when the predictions are accurate, and not behaving poorly

We select a representative subset consisting of 22 datasets and propose it as the MESS benchmark.

This characteristic enables simulation-free training, bypassing the necessity to simulate the underlying diffusion processes by directly sampling the diffused data.

In active learning (AL), we focus on reducing the data annotation cost from the model training perspective. However, "testing", which often refers to the model