Debias-SparseGPT: Bias-Aware Pruning for Large Language Models
Irina Proskurina, Guillaume Metzler, Antoine Gourru, Julien Velcin
EMNLP 2026
cs.CL
2026-09-02
Debias-SparseGPT adds a paired stereotype term to SparseGPT's Hessian. On LLaMA-3.1-8B at 1:4 sparsity, UnQover rises from 35.60% to 60.46% with matched MMLU and perplexity.
Post-training pruners such as SparseGPT treat layer reconstruction as a second-order problem: approximate a Hessian from calibration activations, drop low-saliency weights, then compensate the rest. Aggregate scores (perplexity, MMLU) often hold up. Generative fairness benchmarks do not. With a persona or demographic cue in the prompt, pruned models more often pick an unsupported stereotype instead of "not stated."
Prior papers document the damage. They rarely change the pruning objective. Debias-SparseGPT writes the representational gap between paired stereotypical and anti-stereotypical inputs into the Hessian, so both the mask and the OBS-style update see that gap. Runtime stays in the same cubic class as SparseGPT.
For a weight matrix W, take paired activations X0 and X1 and set ΔX = X0 − X1. The layer objective keeps outputs on both inputs and adds ‖WΔX − W̃ΔX‖², which penalizes stretching the pair gap. The input-space Hessian becomes
H = X0 X0ᵀ + X1 X1ᵀ + 2 ΔX ΔXᵀ.
Saliency is the OBS increase εp = wp² / (2 [Hw⁻¹]pp). The mask keeps the top (1−s) fraction, then block-wise Cholesky updates compensate the survivors. Relative to SparseGPT, the only structural change is the extra ΔX term.
Calibration is 4,212 paired StereoSet development sentences. Under 2:4 structured sparsity the authors add 256 UltraChat dialogues (15k tokens) for coverage; those unpaired texts contribute ordinary XXᵀ, not the ΔX term. Baselines are magnitude pruning, Wanda, SparseGPT, and the dense model. Metrics: WikiText-2 perplexity, zero-shot HellaSwag and MMLU, UnQover/BBQ accuracy at answering "unknown/not stated," CrowS-Pairs stereotype likelihood, and DTO, the Euclidean distance to the (100%, 100%) utopia in normalized MMLU–UnQover space. Nine models, including LLaMA-3.1-8B-IT, Qwen-2.5-7B-IT, and Vicuna-7B.
The headline setting is 1:4 semi-structured sparsity. On LLaMA-3.1-8B, UnQover rises from SparseGPT's 35.60% to 60.46%, BBQ from 67.10% to 70.70% (dense BBQ is 76.40%), and DTO falls from 0.539 to 0.399. MMLU is 59.76% versus 59.11%, perplexity 8.19 versus 8.17. On Qwen-2.5-7B, UnQover moves 70.60% → 74.41%, DTO 0.311 → 0.291, MMLU 67.35% → 67.73%. Vicuna's UnQover only goes 17.82% → 21.54%; the dense model already sits at 17.44%, below the 33.33% three-way chance baseline.
| Model (1:4) | UnQover SparseGPT | UnQover Debias | DTO SparseGPT | DTO Debias |
| LLaMA-3.1-8B | 35.60 | 60.46 | 0.539 | 0.399 |
| Qwen-2.5-7B | 70.60 | 74.41 | 0.311 | 0.291 |
| Vicuna-1.5-7B | 17.82 | 21.54 | 0.695 | 0.674 |
Unstructured 25% and 50% on Qwen keep the same pattern (UnQover 73.43% and 80.35% versus 72.35% and 78.35%). 2:4 is the exception. StereoSet alone leaves Debias worse: UnQover 24.94%, MMLU 48.16%, DTO 0.645, against SparseGPT at 28.45% / 50.41% / 0.616. Adding UltraChat flips it: Debias 47.26% / 54.17% / 0.494 versus SparseGPT 42.46% / 53.84% / 0.522. On an A100 with vLLM, 2:4 throughput is 73.05 tok/s for both pruners, against 27.54 for the dense model.
This is a drop-in Hessian change for people already running SparseGPT, including 2:4 kernels, without a second fine-tune. The calibration corpus is the real knob. LLaMA, whose dense UnQover entropy on correct answers is 0.94 against Qwen's 0.11, gains the most from the pair term. Hard 2:4 patterns still need long, diverse text; a 4k-token stereotype set is not enough and will punch through both fairness and MMLU.
Calibration and evaluation are English-only. The headline metric is abstention under representational bias; toxicity is a side check on RealToxicityPrompts and HarmBench in an appendix, not a full safety audit. CrowS-Pairs hovers near the dense baseline. Six models move closer to the 50% ideal, the rest do not, so likelihood-style stereotype scores barely move. The abstract's "consistent" win across sparsity regimes does not hold for StereoSet-only 2:4. Sparsity masks are barely inspected, and category-specific StereoSet subsets (religion vs gender) shift UnQover enough to show calibration composition matters. There is no theoretical bound on how the bias-aware Hessian moves with calibration size.