PALS: Percentile-Aware Layerwise Sparsity for LLM Pruning
Quick Answer
This paper shows that PALS (Percentile-Aware Layerwise Sparsity) improves transformer pruning by adjusting layer-specific sparsity based on activation magnitudes, achieving 10.96 perplexity on LLaMA-2-7B at 50% sparsity, outperforming uniform methods like Wanda (12.92 perplexity).
Quick Take
PALS (Percentile-Aware Layerwise Sparsity) improves transformer pruning by adjusting layer-specific sparsity based on activation magnitudes, achieving 10.96 perplexity on LLaMA-2-7B at 50% sparsity, outperforming uniform methods like Wanda (12.92 perplexity). The approach shows variable benefits across architectures and incurs negligible costs without requiring fine-tuning.
Key Points
- PALS adjusts layer sparsity based on the 99th percentile of activation magnitudes.
- Achieved 10.96 perplexity on LLaMA-2-7B at 50% sparsity, significantly better than Wanda.
- Architecture-dependent results: LLaMA-3-8B shows marginal gains, Mistral-7B shows none.
- Gradient-based allocation performed worse than random, questioning its effectiveness.
- PALS incurs negligible cost and requires no fine-tuning.
Paper Resources
📖 Reader Mode
~2 min readAbstract: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, $p < 0.001$). 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.
| Subjects: | Computation and Language (cs.CL); Machine Learning (cs.LG) |
| Cite as: | arXiv:2607.07557 [cs.CL] |
| (or arXiv:2607.07557v1 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2607.07557 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Alexey Shvets [view email]
[v1]
Wed, 8 Jul 2026 15:51:57 UTC (49 KB)
— Originally published at arxiv.org
Want this in your inbox every morning?
Daily brief at your local 8am — bilingual EN/中文, free.
More from arXiv cs.CL
See more →Quantifying Prior Dominance in Systems
The study introduces the Normalized Context Utilization (NCU) metric to evaluate Retrieval-Augmented Generation (RAG) systems, revealing that Small Language Models (SLMs) outperform larger models in factual extraction. The findings indicate that traditional scaling laws yield diminishing returns, with a commercial API frequently failing against adversarial evidence due to systemic confidence collapse.