Unleashing the Potential of Model Bias for Generalized Category Discovery