Inquiry Regarding Paper Content

[Tangyueming311]

I have a question regarding the content of your paper. I am curious as to why the decrease in model performance does not correlate proportionally with the decrease in similarity between the OOD dataset and the training dataset. For instance, the similarity between the Amazon and SST datasets is only 33, which is the second lowest, yet the OOD performance reaches 75.16. Have you explored this issue in your research?

Thank you for your time, and I look forward to your insights.