Weight Pruning Amplifies Bias in LLMs for Edge AI

A recent investigation published on arXiv indicates that weight pruning, a method for implementing Large Language Models (LLMs) on limited-resource IoT and edge devices, can greatly enhance model bias. The study involved a systematic empirical analysis of three instruction-tuned models (Gemma-2-9b-it, Mistral-7B-Instruct-v0.3, Phi-3.5-mini-instruct) and examined three pruning techniques (Random, Magnitude, Wanda) across four sparsity levels (10-70%) using 12,148 BBQ bias benchmark items with five random seeds, resulting in 2,368,860 inference records. Findings reveal a "Smart Pruning Paradox": while activation-aware pruning (Wanda) maintains perplexity with only a 3.5% rise at 50% sparsity for Mistral-7B, it also leads to the greatest bias increase. At 70% sparsity, the Stereotype Reliance Score jumps by 83.7%, with 47-59% of previously unbiased items exhibiting new stereotypical behaviors. Random pruning entirely undermines language capability, with perplexity surpassing 10^4. This research underscores the unintended consequences of efficiency-focused pruning methods, which may compromise fairness in edge AI applications.