In this paper, we introduce Debias-SparseGPT, a pruning-time debiasing compression method that explicitly incorporates group differences into the compression objective.<br>The method is motivated by an extensive body of work showing that aggressive pruning can amplify toxic generations and stereotypical biased outputs after compression.</p>\n","updatedAt":"2026-09-03T11:52:00.542Z","author":{"_id":"629a3dbcd496c6dcdebf41cc","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1655113762275-629a3dbcd496c6dcdebf41cc.jpeg","fullname":"Irina Proskurina","name":"iproskurina","type":"user","isPro":true,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":9,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.8721577525138855},"editors":["iproskurina"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/1655113762275-629a3dbcd496c6dcdebf41cc.jpeg"],"reactions":[],"isReport":false}},{"id":"6a996223b54c7bf304753cd2","author":{"_id":"629a3dbcd496c6dcdebf41cc","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1655113762275-629a3dbcd496c6dcdebf41cc.jpeg","fullname":"Irina Proskurina","name":"iproskurina","type":"user","isPro":true,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":9,"isUserFollowing":false},"createdAt":"2026-09-03T12:03:47.000Z","type":"comment","data":{"edited":false,"hidden":false,"latest":{"raw":"Code: https://github.com/upunaprosk/debias-llm-compressor","html":"<p>Code: <a href=\"https://github.com/upunaprosk/debias-llm-compressor\" rel=\"nofollow\">https://github.com/upunaprosk/debias-llm-compressor</a></p>\n","updatedAt":"2026-09-03T12:03:47.624Z","author":{"_id":"629a3dbcd496c6dcdebf41cc","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1655113762275-629a3dbcd496c6dcdebf41cc.jpeg","fullname":"Irina Proskurina","name":"iproskurina","type":"user","isPro":true,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":9,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.45858070254325867},"editors":["iproskurina"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/1655113762275-629a3dbcd496c6dcdebf41cc.jpeg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2609.02496","authors":[{"_id":"6a995e5ac74240b752d78810","name":"Irina Proskurina","hidden":false},{"_id":"6a995e5ac74240b752d78811","name":"Guillaume Metzler","hidden":false},{"_id":"6a995e5ac74240b752d78812","name":"Antoine Gourru","hidden":false},{"_id":"6a995e5ac74240b752d78813","name":"Julien Velcin","hidden":false}],"publishedAt":"2026-09-02T00:00:00.000Z","submittedOnDailyAt":"2026-09-03T00:00:00.000Z","title":"Debias-SparseGPT: Bias-Aware Pruning for Large Language Models","submittedOnDailyBy":{"_id":"629a3dbcd496c6dcdebf41cc","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1655113762275-629a3dbcd496c6dcdebf41cc.jpeg","isPro":true,"fullname":"Irina Proskurina","user":"iproskurina","type":"user","name":"iproskurina"},"summary":"Model compression techniques such as pruning and quantization facilitate the efficient deployment and acceleration of Large Language Models (LLMs). However, recent studies show that weight sparsification methods, such as SparseGPT, can amplify existing biases in models, with outputs varying significantly depending on persona cues in the prompt. In this paper, we introduce Debias-SparseGPT, a post-training pruning method incorporating representational debiasing using a second-order term defined over demographically contrasting inputs. We perform empirical validation of our method over a wide range of generative LLMs. Across models and sparsity regimes (25%, 50%, and structured 2:4 sparsity), Debias-SparseGPT consistently reduces pruning-induced bias compared to SparseGPT while preserving model perplexity and zero-shot accuracy. Under the most restrictive 2:4 structured sparsity pattern, which most aggressively degrades model quality, augmenting the calibration set with long-context, content-rich examples further improves both downstream performance and fairness. Overall, Debias-SparseGPT advances the bias-performance trade-off while preserving the computational efficiency of sparse models.","upvotes":1,"discussionId":"6a995e5ac74240b752d78814","ai_summary":"Debias-SparseGPT reduces pruning-induced demographic bias in compressed large language models by incorporating representational debiasing during post-training sparsification.","ai_keywords":["SparseGPT","Debias-SparseGPT","post-training pruning","weight sparsification","representational debiasing","second-order term","structured 2:4 sparsity","calibration set","large language models"],"ai_summary_model":"thinkingmachines/Inkling-Small"},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"629a3dbcd496c6dcdebf41cc","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1655113762275-629a3dbcd496c6dcdebf41cc.jpeg","isPro":true,"fullname":"Irina Proskurina","user":"iproskurina","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2609/2609.02496.md","query":{}}">
Debias-SparseGPT: Bias-Aware Pruning for Large Language Models
Abstract
Debias-SparseGPT reduces pruning-induced demographic bias in compressed large language models by incorporating representational debiasing during post-training sparsification.
Model compression techniques such as pruning and quantization facilitate the efficient deployment and acceleration of Large Language Models (LLMs). However, recent studies show that weight sparsification methods, such as SparseGPT, can amplify existing biases in models, with outputs varying significantly depending on persona cues in the prompt. In this paper, we introduce Debias-SparseGPT, a post-training pruning method incorporating representational debiasing using a second-order term defined over demographically contrasting inputs. We perform empirical validation of our method over a wide range of generative LLMs. Across models and sparsity regimes (25%, 50%, and structured 2:4 sparsity), Debias-SparseGPT consistently reduces pruning-induced bias compared to SparseGPT while preserving model perplexity and zero-shot accuracy. Under the most restrictive 2:4 structured sparsity pattern, which most aggressively degrades model quality, augmenting the calibration set with long-context, content-rich examples further improves both downstream performance and fairness. Overall, Debias-SparseGPT advances the bias-performance trade-off while preserving the computational efficiency of sparse models.
Community
In this paper, we introduce Debias-SparseGPT, a pruning-time debiasing compression method that explicitly incorporates group differences into the compression objective.
The method is motivated by an extensive body of work showing that aggressive pruning can amplify toxic generations and stereotypical biased outputs after compression.
Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.
Tap or paste here to upload images
Cite arxiv.org/abs/2609.02496 in a model README.md to link it from this page.
Cite arxiv.org/abs/2609.02496 in a dataset README.md to link it from this page.
Cite arxiv.org/abs/2609.02496 in a Space README.md to link it from this page.
Discussion (0)
Sign in to join the discussion. Free account, 30 seconds — email code or GitHub.
Sign in →No comments yet. Sign in and be the first to say something.