How Quantization Shapes Bias in Large Language Models

TL;DR

This paper investigates how quantization techniques influence bias in large language models, revealing nuanced effects on toxicity, stereotypes, and fairness across various models and benchmarks, emphasizing the need for balanced optimization.

Contribution

It provides a comprehensive evaluation of quantization's impact on bias in large language models across multiple bias types and benchmarks, highlighting the importance of ethical considerations in model compression.

Findings

Quantization can reduce toxicity but may increase stereotypes and unfairness.

Effects of quantization are consistent across different model architectures and demographic groups.

Aggressive quantization tends to slightly increase bias in generative tasks.

Abstract

This work presents a comprehensive evaluation of how quantization affects model bias, with particular attention to its impact on individual demographic subgroups. We focus on weight and activation quantization strategies and examine their effects across a broad range of bias types, including stereotypes, fairness, toxicity, and sentiment. We employ both probability- and generated text-based metrics across 13 benchmarks and evaluate models that differ in architecture family and reasoning ability. Our findings show that quantization has a nuanced impact on bias: while it can reduce model toxicity and does not significantly impact sentiment, it tends to slightly increase stereotypes and unfairness in generative tasks, especially under aggressive compression. These trends are generally consistent across demographic categories and subgroups, and model types, although their magnitude depends on the specific setting. Overall, our results highlight the importance of carefully balancing efficiency and ethical considerations when applying quantization in practice.