Weight Pruning Amplifies Bias: A Multi-Method Study of Compressed LLMs for Edge AI

A comprehensive empirical study reveals that weight pruning—a technique for compressing large language models for edge devices—paradoxically amplifies bias while preserving performance metrics. The research shows activation-aware pruning methods maintain perplexity but increase stereotype reliance by up to 84%, suggesting current evaluation methods fail to detect fairness degradation in compressed models.

The research addresses a critical blind spot in AI deployment practices: the assumption that compression techniques maintaining performance metrics also preserve model behavior. As organizations accelerate edge AI adoption for IoT applications, this study demonstrates that such assumptions are dangerously flawed. The Smart Pruning Paradox—where the most sophisticated pruning method (Wanda) produces the worst fairness outcomes despite near-perfect perplexity preservation—exposes fundamental limitations in relying on language modeling metrics to validate behavioral equivalence.

The broader context involves the race to democratize LLM deployment across resource-constrained devices. Companies and researchers have championed weight pruning as a solution, yet this study's testing of three popular models across multiple pruning strategies and sparsity levels reveals that bias amplification scales with compression intensity. The finding that 47-59% of previously unbiased responses become stereotypical at 70% sparsity suggests systemic rather than random effects.

For AI developers and organizations deploying models at the edge, this research introduces a material compliance and reputational risk. Products marketed as lightweight alternatives may inadvertently amplify harmful biases. The comparison to quantization—showing pruning creates three times higher bias transition rates—reframes pruning's safety profile. Additionally, the revelation that unstructured pruning provides zero actual performance gains on real hardware undermines the technical justification for adoption.

The path forward requires implementing bias-aware validation benchmarks before edge deployment, treating fairness metrics as deployment prerequisites rather than post-hoc checks. Organizations must reconsider compression strategy trade-offs with explicit fairness testing protocols.

• →Weight pruning amplifies model bias up to 84% while maintaining perplexity, creating false assurance of safe compression through inadequate metrics.
• →Activation-aware pruning (Wanda) performs worst for fairness despite best technical performance, revealing a fundamental paradox in compression evaluation.
• →Pruning causes 47-59% of unbiased responses to develop stereotypical behaviors at high sparsity, three times higher than quantization-induced bias shifts.
• →Unstructured pruning provides zero real-world storage or latency benefits on edge hardware, undermining its primary deployment justification.
• →Current perplexity-based evaluation frameworks provide insufficient validation for behavioral equivalence and cannot guarantee fair model deployment.