F-PvE: Fairness-Aware Structured Neural Network Pruning via Evolution

ICLR 2026 Conference Submission16360 Authors

Keywords: Structured neural network pruning, fairness, evolutionary optimization

TL;DR: This paper introduces F-PvE, a general framework for fairness-aware structured neural network pruning via evolution, where multi-objective evolution and cooperative evolution techniques are further adopted to address the inherent challenges.

Abstract: Model compression plays a crucial role in deployment, as it can significantly reduce computational costs with minimal loss in accuracy. However, recent studies have shown that model compression may involve additional bias, posing fairness risks that can potentially lead to social impact. As a result, mitigating bias during model compression has emerged as an important topic. In this work, we focus on structured neural network pruning, a widely adopted model compression technique that remains rarely explored in the context of fairness. Specifically, we introduce evolutionary algorithms as a general yet powerful approach to achieve fairness-aware structured pruning. That is, we formulate structured pruning as a subset selection problem and use evolutionary search to identify an optimal set of structural components to retain, balancing both accuracy and fairness objectives. Given the multi-objective nature and the large combinatorial search space of structural components, we further incorporate multi-objective evolution and cooperative coevolution to effectively address them. To verify the effectiveness of our method, we conduct experiments that cover three typical fairness scenarios: class-wise and group-wise fairness in classification models, and toxicity in language models. Compared with classic structured pruning methods and state-of-the-art competitors on fairness-aware structured pruning, our method can preserve better fairness while keeping competitive accuracy, demonstrating the superiority of evolutionary optimization for fairness-aware structured pruning in practice.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Submission Number: 16360