Preserving Fairness and Safety in Quantized LLMs Through Critical Weight Protection

TL;DR

This paper investigates how quantization affects fairness and safety in large language models across multiple languages and introduces Critical Weight Protection to mitigate these issues effectively.

Contribution

It systematically studies the impact of static and dynamic quantization on fairness and safety, and proposes a novel method to preserve critical weights during quantization.

Findings

Quantization degrades fairness and safety, especially in non-English languages.

Dynamic quantization methods are more stable than static ones.

Critical Weight Protection mitigates bias and safety issues without retraining.

Abstract

Quantization is widely adopted to reduce the computational cost of large language models (LLMs); however, its implications for fairness and safety, particularly in dynamic quantization and multilingual contexts, remain underexplored. In this work, we conduct a systematic study of how static and dynamic quantization methods impact fairness and safety across benchmarks measuring intrinsic and extrinsic bias and safety alignment. For fairness, we evaluate English, French, Dutch, Spanish, and Turkish; for safety, we focus on English, Korean, and Arabic. Our findings reveal that quantization consistently degrades fairness and safety, with dynamic methods demonstrating greater stability than static ones. Moreover, fairness degradation varies across languages, while safety deterioration is especially pronounced in non-English settings. To address these risks, we introduce Critical Weight Protection, a novel technique that identifies and preserves fairness- and safety-critical weights during quantization. This approach effectively mitigates bias and safety deterioration without costly retraining or alignment, maintaining trustworthiness while retaining efficiency.