Spurious Correlation
A non-causal association between input features and labels learned by a machine learning model, typically arising from observational bias in the training data, which causes degraded performance on out-of-distribution data.
Timeline (last 90 days)
Injecting "ties" (i.e., pairs of responses that are equally preferred) into preference data can weaken spurious correlations, because ties provide no discriminative signal and force the model to ignore irrelevant features.
Different tie pairs vary greatly in their effect on reducing spurious correlation; some ties can more effectively correct the model's erroneous dependencies, while others may yield little effect.
In preference optimization, spurious correlation specifically refers to the non-causal association between input features and preference labels learned by the model, whose root cause lies in the observational bias of preference data.
All Facts (3)
Injecting "ties" (i.e., pairs of responses that are equally preferred) into preference data can weaken spurious correlations, because ties provide no discriminative signal and force the model to ignore irrelevant features.
50%UnverifiedDifferent tie pairs vary greatly in their effect on reducing spurious correlation; some ties can more effectively correct the model's erroneous dependencies, while others may yield little effect.
50%UnverifiedIn preference optimization, spurious correlation specifically refers to the non-causal association between input features and preference labels learned by the model, whose root cause lies in the observational bias of preference data.
50%