Spuriosity Didn't Kill the Classifier: Using Invariant Predictions to Harness Spurious Features Cian Eastwood 1,2 Shashank Singh 1 Andrei L. Nicolicioiu 1 Marin Vlastelica

Feb-10-2026, 12:44:10 GMT–Neural Information Processing Systems 

Theoretically, we prove that SFB can learn an asymptotically-optimal predictor without test-domain labels. Empirically, we demonstrate the effectiveness of SFB on real and synthetic data.