Concept捷径学习 / spurious correlation
Spurious Correlation
A phenomenon in which a machine learning model relies on superficial features in training data rather than genuine causal relationships during learning; in large-model safety alignment it manifests as judging risk based on keywords rather than intent.
Timeline (last 90 days)
Oct 26
Safety samples in alignment training data often carry specific keywords and sentence patterns, and models readily learn "surface features" rather than genuine intent understanding, which is called spurious correlation or shortcut learning.
Unverified50%