Quantization Can Break Algorithmic Recourse: New Metrics and Mitigation

A recent paper on arXiv (2605.17160) explores the implications of model quantization, a prevalent strategy for minimizing memory, latency, and deployment expenses, on algorithmic recourse in decision-making systems. While quantization is often assessed based on its effect on predictive accuracy, the authors contend that merely maintaining accuracy is inadequate in scenarios where users require actionable changes to alter decisions. They present a framework called counterfactual sensitivity under quantization to evaluate how compression affects recourse behavior. Two new metrics are introduced: Validity Drop (VD), which measures the proportion of full-precision recourse actions that fail post-quantization, and Counterfactual Recourse Gap (CRG), indicating the rise in minimal recourse costs in quantized models. To counteract these issues, the authors propose Counterfactual-Faithful Quantization, aimed at alleviating the adverse impacts of quantization on recourse. This study emphasizes a deployment mismatch, where quantization may render recourse actions invalid or necessitate larger interventions, potentially affecting fairness and user agency in automated decision processes.