Analyzing the residual predictor: To understand whether PCBM-h overrides PCBM predictions, in Appendix B, we look at the consistency between PCBM and PCBM-h predictions. We show that the residual component in PCBM-h intervenes only when the prediction is wrong, and fixes mistakes. When PCBM is confident, PCBM-h does not modify the prediction or significantly increase the confidence. In general, the residual component may dominate when the bottleneck is insufficient and future work can aim to explicitly limit the residual component, e.g. PIE (Wang et al., 2021) regularizer