Weight Pruning Amplifies Bias in Compressed LLMs for Edge AI

A new study published on arXiv reveals that weight pruning, a common technique for deploying Large Language Models (LLMs) on resource-constrained edge devices, significantly amplifies model bias. Researchers conducted a comprehensive empirical analysis of three instruction-tuned models using various pruning methods and sparsity levels. The findings highlight a 'Smart Pruning Paradox,' where activation-aware pruning methods like Wanda preserve language perplexity but cause substantial increases in stereotypical behaviors, with up to 59% of previously unbiased items developing new biases at high sparsity. In contrast, random pruning destroys language capability but results only in chance-level bias. Furthermore, the study demonstrates that unstructured pruning offers no actual storage or latency benefits on real edge hardware, challenging its primary utility for IoT deployments. The results indicate that pruning poses a greater risk to model alignment than quantization, with response transition rates between biased and unbiased states nearly three times higher. Consequently, the authors argue that perplexity-based evaluations provide false assurances of fairness and urge the implementation of bias-aware validation protocols before deploying pruned models in edge computing environments.