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

• URL: http://arxiv.org/abs/2605.08137v1
• Date: Sat, 02 May 2026 05:27:40 GMT
• Title: Weight Pruning Amplifies Bias: A Multi-Method Study of Compressed LLMs for Edge AI
• Abstract summary: We show that activation-aware pruning (Wanda) preserves perplexity nearly perfectly (just 3.5% increase at 50% sparsity for Mistral-7B) yet produces the highest bias amplification.<n>Random pruning destroys language capability entirely (perplexity exceeding $104$ and reaching $108$) but produces only random-chance bias.<n>We further show that unstructured pruning provides zero storage savings and zero inference latency reduction on real edge hardware.
• Score: 0.0
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• 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.

Related papers

• Evaluating Fine-Tuning and Metrics for Neural Decompilation of Dart AOT Binaries [0.0]
  We present a systematic empirical study of fine-tuning effectiveness and validity for neural decompilation.<n>We evaluate six finetuned model variants across three base architectures using three metrics: CodeBLEU, compile@k, and pass@k.<n>Crosslingual interference from Swift training is highly significant at 4B (-2.66 pp, p but statistically indistinguishable from zero at 8B, consistent with the scaling hypothesis)
  arXiv  Detail & Related papers  (2026-07-07T10:36:12Z)
• Knowledge Index of Noah's Ark [63.143852586221534]
  We introduce KINA, an 899-item benchmark across 261 fine-grained disciplines.<n>We prove a bonus-on-bar tournament weakly FOSD-dominates flat payment in released-review quality.<n>The top model, Gemini-3.1-Pro-Preview, reaches 53.17%, followed by Claude-Opus-4.6 at 49.92% and GPT-5.4 at 48.55%.
  arXiv  Detail & Related papers  (2026-06-03T17:06:49Z)
• From Correlation to Cause: A Five-Stage Methodology for Feature Analysis in Transformer Language Models [0.0]
  We propose a five-stage methodology for causal feature analysis in transformer language models.<n>We demonstrate it end-to-end on GPT-2 small performing the Indirect Object Identification (IOI) task.
  arXiv  Detail & Related papers  (2026-05-21T13:25:16Z)
• Quantization Undoes Alignment: Bias Emergence in Compressed LLMs Across Models and Precision Levels [0.0]
  Large Language Models are routinely compressed via post-training quantization to reduce inference costs and memory footprint for cloud and edge deployment.<n>Existing studies typically compare only two conditions, rely on aggregate bias metrics, and evaluate a single model family.<n>We conduct a controlled empirical study of three instruction-tuned models at five precision levels (BF16 through 3-bit) on 12,148 BBQ bias benchmark items.
  arXiv  Detail & Related papers  (2026-05-02T05:41:47Z)
• Cut Your Losses! Learning to Prune Paths Early for Efficient Parallel Reasoning [30.70183055909553]
  We propose the first systematic taxonomy of path pruning, categorizing methods by their signal source (internal vs. external) and learnability (learnable vs. non-learnable)<n>This classification reveals the unexplored potential of learnable internal methods, motivating our proposal of STOP (Super TOken for Pruning)
  arXiv  Detail & Related papers  (2026-04-17T13:00:22Z)
• One Token Away from Collapse: The Fragility of Instruction-Tuned Helpfulness [12.183451602438753]
  We show that simple lexical constraints (banning a single punctuation character or common word) cause instruction-tuned LLMs to collapse their responses.<n>We show that base models show no systematic collapse under identical constraints, with effects that are small, noisy, and bidirectional.
  arXiv  Detail & Related papers  (2026-04-14T17:40:01Z)
• Evaluating Small Language Models for Front-Door Routing: A Harmonized Benchmark and Synthetic-Traffic Experiment [0.05586191108738562]
  Small language models (SLMs) have sufficient reasoning capability for sub-second, zero-marginal-cost, self-hosted task classification.<n>Study 1 is a harmonized offline benchmark of Phi-3.5-mini, Qwen2.5-1.5B, and Qwen-2.5-3B on identical Azure T4 hardware, serving stack, quantization, and a fixed 60-case corpus.<n>Study 2 is a pre-registered four-arm randomized experiment under synthetic traffic with an effective sample size of 60 unique cases per arm.
  arXiv  Detail & Related papers  (2026-03-26T15:57:46Z)
• Brevity Constraints Reverse Performance Hierarchies in Language Models [0.0]
  On 7.7% of benchmark problems, larger language models underperform smaller ones by 28.4 percentage points despite 10-100x more parameters.<n>We identify the mechanism as spontaneous scale-dependent verbosity that introduces errors through overelaboration.<n>Our results establish that maximizing large model performance requires scale-aware prompt engineering.
  arXiv  Detail & Related papers  (2026-03-11T06:47:41Z)
• ODAR: Principled Adaptive Routing for LLM Reasoning via Active Inference [60.958331943869126]
  ODAR-Expert is an adaptive routing framework that optimize the accuracy-efficiency trade-off via principled resource allocation.<n>We show strong and consistent gains, including 98.2% accuracy on MATH and 54.8% on Humanity's Last Exam.
  arXiv  Detail & Related papers  (2026-02-27T05:22:01Z)
• Information Fidelity in Tool-Using LLM Agents: A Martingale Analysis of the Model Context Protocol [69.11739400975445]
  We introduce the first theoretical framework for analyzing error accumulation in Model Context Protocol (MCP) agents.<n>We show that cumulative distortion exhibits linear growth and high-probability deviations bounded by $O(sqrtT)$.<n>Key findings include: semantic weighting reduces distortion by 80%, and periodic re-grounding approximately every 9 steps suffices for error control.
  arXiv  Detail & Related papers  (2026-02-10T21:08:53Z)
• Think Before You Prune: Self-Reflective Structured Pruning for Reasoning Language Models [31.422773877490613]
  Reasoning LLMs (RLMs) deliver strong multi-step reasoning through chain-of-thought generation.<n>RLMs' large model sizes and lengthy decode-time outputs make them costly to deploy and unsuitable for resource-constrained settings.<n>We introduce RESP, a structured pruning framework that aligns pruning decisions with the model's reasoning dynamics.
  arXiv  Detail & Related papers  (2025-12-01T20:27:05Z)
• RefineBench: Evaluating Refinement Capability of Language Models via Checklists [71.02281792867531]
  We evaluate two refinement modes: guided refinement and self-refinement.<n>In guided refinement, both proprietary LMs and large open-weight LMs can leverage targeted feedback to refine responses to near-perfect levels within five turns.<n>These findings suggest that frontier LMs require breakthroughs to self-refine their incorrect responses.
  arXiv  Detail & Related papers  (2025-11-27T07:20:52Z)
• Benchmarking Reasoning Robustness in Large Language Models [76.79744000300363]
  We find significant performance degradation on novel or incomplete data.<n>These findings highlight the reliance on recall over rigorous logical inference.<n>This paper introduces a novel benchmark, termed as Math-RoB, that exploits hallucinations triggered by missing information to expose reasoning gaps.
  arXiv  Detail & Related papers  (2025-03-06T15:36:06Z)
• Advancing LLM Reasoning Generalists with Preference Trees [119.57169648859707]
  We introduce Eurus, a suite of large language models (LLMs) optimized for reasoning. Eurus models achieve state-of-the-art results among open-source models on a diverse set of benchmarks.
  arXiv  Detail & Related papers  (2024-04-02T16:25:30Z)
• How Easy is It to Fool Your Multimodal LLMs? An Empirical Analysis on Deceptive Prompts [54.07541591018305]
  We present MAD-Bench, a benchmark that contains 1000 test samples divided into 5 categories, such as non-existent objects, count of objects, and spatial relationship. We provide a comprehensive analysis of popular MLLMs, ranging from GPT-4v, Reka, Gemini-Pro, to open-sourced models, such as LLaVA-NeXT and MiniCPM-Llama3. While GPT-4o achieves 82.82% accuracy on MAD-Bench, the accuracy of any other model in our experiments ranges from 9% to 50%.
  arXiv  Detail & Related papers  (2024-02-20T18:31:27Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.