Lightweight Haar Wavelet Subband Pruning for LLMs - ACL AnthologyLightweight Haar Wavelet Subband Pruning for LLMs<br>Jiang Li,<br>Pengfei Cao,<br>Chenxi Zhou,<br>Tian Lan,<br>Xiangdong Su,<br>Kang Liu,<br>Jun Zhao,<br>Guanglai Gao
Correct Metadata for
Use this form to create a GitHub issue with structured data describing the correction. You will need a GitHub account.<br>Once you create that issue, the correction will be reviewed by a staff member.<br>⚠️ Mobile Users: Submitting this form to create a new issue will only work with github.com, not the GitHub Mobile app.<br>Important : The Anthology treat PDFs as authoritative. Please use this form only to correct data<br>that is out of line with the PDF. See our corrections<br>guidelines if you need to change the PDF.<br>Title<br>Adjust the title. Retain tags such as
Authors<br>Adjust author names and order to match the<br>PDF.<br>Add AuthorAbstract<br>Correct abstract if needed. Retain XML formatting tags such as . You may use ... for bold , ... for italic, ... for underline, ... for small-caps, ... for typewriter text, ... for URLs, for hyperlinks, and for paragraph breaks.
Verification against PDF<br>Ensure that the new title/authors match the snapshot below. (If there<br>is no snapshot or it is too small, consult the PDF.)<br>Authors concatenated from the text boxes above:
ALL author names match the snapshot above—including<br>middle initials, hyphens, and accents.<br>Create GitHub issue for staff review
Abstract<br>Large language models (LLMs) reach state-of-the-art performance across many NLP tasks, but their large parameter counts introduce heavy computational and memory overhead, which complicates deployment in resource-constrained settings. Pruning is a standard compression strategy that induces sparsity to lower these costs. However, most pruning methods for LLMs depend on calibration data and expensive weight updates, which limits practical scalability. To address these limitations, we introduce H aar W avelet S ubband P runing (), a post-training framework that requires no calibration data and no weight updates. applies a two-dimensional Haar wavelet transform to each weight matrix and decomposes it into four frequency subbands. It then assigns a uniform sparsity ratio to all subbands so that both low- and high-frequency components are retained in a balanced manner. Our theoretical analysis shows that the subband design of provides a deterministic per-subband retention guarantee, which helps mitigate the potential bias of global magnitude pruning toward dominant frequency components. Experiments on the LLaMA, OPT and Qwen model families show that achieves competitive accuracy relative to strong pruning baselines while substantially reducing pruning time. Compared with magnitude pruning, which serves as a simple calibration-free baseline, generally achieves better downstream performance across a wide range of sparsity levels and model scales.
Anthology ID:2026.findings-acl.798Volume:Findings of the Association for Computational Linguistics: ACL 2026Month:JulyYear:2026Address:San Diego, California, United StatesEditors:Maria Liakata,<br>Viviane P. Moreira,<br>Jiajun Zhang,<br>David JurgensVenue:FindingsSIG:Publisher:Association for Computational LinguisticsNote:Pages:16242–16259Language:URL:https://aclanthology.org/2026.findings-acl.798/DOI:10.18653/v1/2026.findings-acl.798Bibkey:li-etal-2026-lightweightCite (ACL):Jiang Li, Pengfei Cao, Chenxi Zhou, Tian Lan, Xiangdong Su, Kang Liu, Jun Zhao, and Guanglai Gao. 2026. Lightweight Haar Wavelet Subband Pruning for LLMs. In Findings of the Association for Computational Linguistics: ACL 2026, pages 16242–16259, San Diego, California, United States. Association for Computational Linguistics.Cite (Informal):Lightweight Haar Wavelet Subband Pruning for LLMs (Li et al., Findings 2026)Copy Citation:BibTeX<br>Markdown<br>MODS XML<br>Endnote<br>More<br>options…PDF:https://aclanthology.org/2026.findings-acl.798.pdfChecklist:<br>2026.findings-acl.798.checklist.pdf<br>PDF<br>Cite<br>Search
Checklist
Fix data
Export citation
BibTeX<br>MODS XML<br>Endnote<br>Preformatted<br>@inproceedings{li-etal-2026-lightweight,<br>title = "Lightweight Haar Wavelet Subband Pruning for {LLM}s",<br>author = "Li, Jiang and<br>Cao, Pengfei and<br>Zhou, Chenxi and<br>Lan, Tian and<br>Su, Xiangdong and<br>Liu, Kang and<br>Zhao, Jun and<br>Gao, Guanglai",<br>editor = "Liakata, Maria and<br>Moreira, Viviane P. and<br>Zhang, Jiajun and<br>Jurgens, David",<br>booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {ACL} 2026",<br>month = jul,<br>year = "2026",<br>address = "San Diego, California, United States",<br>publisher = "Association for Computational Linguistics",<br>url = "https://aclanthology.org/2026.findings-acl.798/",<br>doi = "10.18653/v1/2026.findings-acl.798",<br>pages = "16242--16259",<br>ISBN = "979-8-89176-395-1",<br>abstract = "Large language models (LLMs) reach state-of-the-art performance across many NLP tasks, but their large parameter counts introduce heavy computational and memory overhead, which complicates deployment in resource-constrained settings. Pruning is a...