{"ID":6138334,"CreatedAt":"2026-07-09T01:07:32.349475501Z","UpdatedAt":"2026-07-11T15:39:16.154462532Z","DeletedAt":null,"paper_url":"https://arxiv.org/abs/2607.07557","arxiv_id":"2607.07557","title":"PALS: Percentile-Aware Layerwise Sparsity for LLM Pruning","abstract":"One-shot pruning methods like Wanda and SparseGPT apply the same sparsity ratio to every layer of a transformer, ignoring known variation in layer importance. We propose PALS (Percentile-Aware Layerwise Sparsity), which adjusts per-layer sparsity based on the 99th percentile of activation magnitudes, bounded to $\\pm 5\\%$ around the target ratio. On LLaMA-2-7B at 50\\% sparsity, PALS achieves 10.96 WikiText-2 perplexity versus 12.92 for uniform Wanda (mean over 9 runs, $p \u003c 0.001$). The benefit is architecture-dependent: LLaMA-3-8B shows marginal gains and Mistral-7B shows none. We also find that gradient-based allocation -- the seemingly more principled approach -- produces results worse than random, suggesting that gradient magnitude does not predict the impact of discrete weight removal. PALS adds negligible cost to the pruning pipeline and requires no fine-tuning.","short_abstract":"One-shot pruning methods like Wanda and SparseGPT apply the same sparsity ratio to every layer of a transformer, ignoring known variation in layer importance. We propose PALS (Percentile-Aware Layerwise Sparsity), which adjusts per-layer sparsity based on the 99th percentile of activation magnitudes, bounded to $\\pm 5\\...","url_abs":"https://arxiv.org/abs/2607.07557","url_pdf":"https://arxiv.org/pdf/2607.07557v1","authors":"[\"Yazdan Jamshidi\",\"Alexey Shvets\"]","published":"2026-07-08T15:51:57Z","proceeding":"cs.CL","tasks":"[\"cs.CL\",\"cs.LG\"]","methods":"[\"Transformer\",\"Large Language Model\"]","has_code":false}