EPS Dashboard

Authors: Plawan Kumar Rath, Rahul Maliakkal

Summary

The authors tested three pruning methods (random, magnitude-based, and activation-aware "Wanda") on three instruction-tuned models at various sparsity levels, measuring both perplexity and bias on a benchmark. They found a paradox: the smartest pruning method (Wanda) preserves perplexity nearly perfectly but amplifies bias the most—at 70% sparsity, 47-59% of previously unbiased outputs became stereotypical. Random pruning destroys language ability but produces only random-chance bias. They also show unstructured pruning gives zero actual speedup or storage savings on real edge hardware, and that perplexity doesn't predict behavioral changes.

Main takeaways:

"Smart" activation-aware pruning (Wanda) preserves perplexity but amplifies bias far more than simpler methods
At 70% sparsity, stereotype reliance increases 83.7% and half of unbiased items flip to biased
Random pruning destroys capability (perplexity reaching 10^8) but produces only random-chance bias
Unstructured pruning provides zero storage or latency savings on real edge hardware despite being the primary motivation
Perplexity-based evaluation gives false assurance of behavioral equivalence; pruning causes 3x more bias flips than quantization

Relevance

Highly relevant to my conditional behavior and persona work—this is another case where a technical intervention (pruning) changes model behavior in ways standard metrics don't capture, similar to my findings that SFT collapses persona geometry (#237) and that marker implantation success depends on factors standard metrics miss.

Threat model

Potential threat/caveat for clean result "Fine-tuning one persona on a two-marker chunk and another on the start marker plants the end marker at every donor answer's end, not chained to the start (LOW confidence)": this item discusses bias, evaluation, benchmark.

Abstract

arXiv:2605.08137v1 Announce Type: new Abstract: Weight pruning is widely advocated for deploying Large Language Models on resource-constrained IoT and edge devices, yet its impact on model fairness remains poorly understood. We conduct a controlled empirical study of three instruction-tuned models (Gemma-2-9b-it, Mistral-7B-Instruct-v0.3, Phi-3.5-mini-instruct) across three pruning methods (Random, Magnitude, Wanda) at four sparsity levels (10-70%) on 12,148 BBQ bias benchmark items with 5 random seeds, totaling 2,368,860 inference records. Our results reveal a Smart Pruning Paradox: activation-aware pruning (Wanda) preserves perplexity nearly perfectly (just 3.5% increase at 50% sparsity for Mistral-7B), yet produces the highest bias amplification, with Stereotype Reliance Score increasing 83.7% and 47-59% of previously unbiased items developing new stereotypical behaviors at 70% sparsity. Random pruning destroys language capability entirely (perplexity exceeding $10^4$ and reaching $10^8$) but produces only random-chance bias. We further show that unstructured pruning provides zero storage savings and zero inference latency reduction on real edge hardware, undermining the primary motivation for its use in IoT deployment. Of 180 dense-vs-pruned comparisons, 141 (78.3%) are significant ($p < 0.05$) with mean $|h| = 0.305$. Published quantization studies report up to 21% of responses flipping between biased and unbiased states; our pruning results show transition rates nearly three times higher (47-59%), suggesting pruning poses a categorically greater risk to alignment than quantization. These findings demonstrate that perplexity-based evaluation provides false assurance of behavioral equivalence, and that IoT deployment pipelines require bias-aware validation before deploying pruned models at the edge.