Counterfactual-Faithful Quantization aims to keep algorithmic recourse stable after compression

An arXiv machine learning paper studies how model quantization can break algorithmic recourse even when predictive accuracy is preserved. Small actionable changes that flip a full-precision model may fail after quantization or require a larger intervention. The authors define counterfactual sensitivity under quantization with two metrics: Validity Drop, the fraction of full-precision recourse actions that no longer achieve the target after quantization, and Counterfactual Recourse Gap, the increase in minimal recourse cost under the quantized model. They propose Counterfactual-Faithful Quantization, a quantization-aware training method that jointly learns quantizer parameters and mixed-precision bit allocation while preserving the target prediction at teacher-generated recourse points. CFQ works with standard LSQ/PACT-style quantizers and mixed-precision policies and can also run as training-free calibration for post-training quantization. On Adult, German Credit, and COMPAS, standard quantization-aware training and mixed-precision baselines can keep accuracy while degrading recourse stability. At matched accuracy and bit budget, CFQ reduces Validity Drop and Counterfactual Recourse Gap; on Adult, the paper reports VD/CRG falling from 0.121/0.162 for an accuracy-centric mixed-precision baseline to 0.061/0.071.