Researchers have revisited the concept of preprocessing invariance in spectral foundation models, specifically Raman foundation models. They found that these models' normalization step can map differently preprocessed spectra to the same vector, making it difficult to attribute invariance to learning. The study suggests that the encoder's performance should be measured against its own normalization, which has no learned parameters. The results show that the model does not measurably outperform its own normalization on six Raman evaluation datasets. --- Why it matters: This research matters because it challenges the conventional understanding of preprocessing invariance in spectral foundation models and highlights the importance of accurately measuring their performance. It also raises questions about the validity of claims made by other systems that claim to have achieved this invariance through learning. Source: https://arxiv.org/abs/2608.14227