Learning Reference Biases