Can you quantify the minimum diversity of backgrounds needed before adding more stops improving out-of-distribution sensitivity?

I appreciate the specificity of your question, but the preprints provided don't directly address it.

Your question asks for a quantitative threshold—a minimum number or diversity measure of background processes beyond which additional backgrounds stop improving out-of-distribution sensitivity. While [AG-2024.01-1031] demonstrates that training on *multiple* background types improves anomaly detection compared to a single dominant background, and [AG-2025.05-1524] studies sensitivity to hyperparameters in anomaly detection methods, neither paper establishes a saturation point or diminishing-returns curve that would let you answer "you need N background types" or "diversity index D is sufficient."

The closest work is [AG-2024.01-1031], which shows benefits of the multi-background approach but doesn't systematically vary the number or diversity of backgrounds to find where adding more stops helping—that would require ablation studies (removing backgrounds one by one) or scaling studies (progressively adding more diverse backgrounds) that these papers don't appear to report.