When Bits Break Recourse: Counterfactual-Faithful Quantization

TL;DR

This paper introduces a new quantization method called Counterfactual-Faithful Quantization (CFQ) that preserves the ability of models to provide reliable recourse after quantization, addressing a critical gap in existing low-bit deployment techniques.

Contribution

The paper formalizes counterfactual sensitivity in quantization, proposes metrics for recourse failure detection, and develops CFQ to maintain recourse stability without sacrificing accuracy.

Findings

CFQ significantly improves recourse stability metrics (VD and CRG) across various datasets.

Baseline quantization methods often degrade recourse quality without affecting accuracy.

CFQ maintains model accuracy while substantially reducing recourse failure risks.

Abstract

Quantization can preserve predictive accuracy under low-bit deployment while silently breaking algorithmic recourse: an actionable change that flips a decision before quantization may fail after quantization, or become substantially more costly. We formalize counterfactual sensitivity under quantization through validity, cost, and direction stability, and introduce two metrics: Validity Drop (VD) and Counterfactual Recourse Gap (CRG) that reveal recourse failures invisible to accuracy. We propose Counterfactual-Faithful Quantization (CFQ), which trains quantizer parameters and mixed-precision bit allocation to preserve counterfactual behavior by enforcing the target outcome at teacher recourse points under a global bit budget. A margin-based analysis gives a sufficient condition for recourse transfer under bounded quantization perturbations. Experiments on Adult, German Credit, and COMPAS show that accuracy-matched baselines can significantly degrade recourse stability, while CFQ maintains accuracy and substantially improves VD and CRG across bit budgets.