Shortcut Learning
A widely observed phenomenon in deep learning where a model exploits surface-level statistical correlations in training data that are associated with labels but semantically irrelevant. Models exhibiting shortcut learning perform well on in-distribution data but fail to generalize to out-of-distribution scenarios.
Timeline (last 90 days)
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.
In traditional evaluations, models can rely on parametric knowledge or semantic cues in the compressed text to guess the answer even when the key evidence has been removed, a phenomenon called "shortcut learning."
In synthetic data, replacing identifiers, names, section order, and surface wording can create samples that appear different but retain the same underlying phrase templates, which provides shortcuts for the model; exact duplicate detection cannot find this overlap, and finer-grained methods such as n-gram overlap, embedding similarity, or template family tracking are needed.
All Facts (3)
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.
50%UnverifiedIn traditional evaluations, models can rely on parametric knowledge or semantic cues in the compressed text to guess the answer even when the key evidence has been removed, a phenomenon called "shortcut learning."
50%UnverifiedIn synthetic data, replacing identifiers, names, section order, and surface wording can create samples that appear different but retain the same underlying phrase templates, which provides shortcuts for the model; exact duplicate detection cannot find this overlap, and finer-grained methods such as n-gram overlap, embedding similarity, or template family tracking are needed.
50%