Debias-SparseGPT: Bias-Aware Pruning for Large Language Models
Irina Proskurina, Guillaume Metzler, Antoine Gourru, Julien Velcin
Laboratoire Hubert Curien · Université Claude Bernard Lyon 1 · Université Lumière Lyon 2 · École Centrale de Lyon
Published September 2, 2026 · on the feed September 3, 2026 · ▲ 2 · arXiv 2609.02496
The abstract
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.
The Brief
2:06A short podcast summary of the paper. Written and voiced by AI.
Transcript
Debias-SparseGPT. Bias-Aware Pruning for Large Language Models. By Irina Proskurina and three others, from the Hubert Curien Laboratory, Claude Bernard University Lyon 1, and two other institutions. Published on the second of September. No code repository is linked to it.
Pruning and quantization are how large language models get cheap enough to deploy and fast enough to serve. Compression has a side effect, though. Recent studies show that weight sparsification methods such as SparseGPT can amplify biases a model already carries. After pruning, outputs can vary significantly depending on persona cues in the prompt, meaning that who the prompt says is asking changes what comes back more than it did before.
Debias-SparseGPT is a post-training pruning method that folds debiasing into the pruning itself. It incorporates representational debiasing through a second-order term, and that term is defined over demographically contrasting inputs: inputs chosen so that they differ along demographic lines. Bias is addressed inside the pruning step, rather than in a separate pass before or after it.
The method is validated empirically over a wide range of generative language models, and across three sparsity regimes. Two are unstructured, removing twenty-five percent and fifty percent of weights. The third is the structured two-out-of-four pattern, the most restrictive of the three and the one that most aggressively degrades model quality.
Across models and regimes, Debias-SparseGPT consistently reduces the bias that pruning introduces, compared with SparseGPT, while preserving perplexity and zero-shot accuracy. The structured pattern gets an extra finding. There, augmenting the calibration set with long-context, content-rich examples further improves both downstream performance and fairness, so the hardest setting is also where the choice of calibration data pays off on both counts at once.
The result is framed as a better trade-off between bias and performance, not as a cost paid for fairness. Sparse models keep their computational efficiency, and the step that makes them sparse stops being the step that makes them more biased.
Paper Quest · 7 chapters · about 41 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Compression is cheap until it costs you fairness. Here is the wall.
You compress a large language model by zeroing out weights after training, with a second-order method that keeps each layer's outputs close to the original on a few calibration inputs.
The road ahead
- 1 · The problem ~7 min Removing weights after training keeps a language model fluent, but it can make the model lean harder on stereotypes about who is asking.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Add a second-order term over paired, demographically contrasting inputs to the pruning objective, so pruning keeps how each layer tells the pair apart.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Gather paired demographic inputs, build a bias-aware Hessian from them, prune each layer to the target sparsity, then check bias and quality.
- 6 · The experiment ~6 min You’ll guess: Under the strictest 2:4 sparsity, the authors added long-context, content-rich examples to the calibration set. What happened?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.