Fair-GPTQ: Bias-Aware Quantization for Large Language Models

arXiv:2509.15206 · cs.CL · Submitted 2025-09-18 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom, and with me are Jane, Lu, senior AI researcher at Tsinghua, Meng, lead engineer at a mysterious AI startup and Lalam, the in-house Large Language Model.

Jane: Today's paper: "Fair-GPTQ: Bias-Aware Quantization for Large Language Models".

Tom: The gist The first quantization method explicitly designed to reduce unfairness in large language models, Fair-GPTQ, introduces explicit group-fairness constraints into the quantization objective to mitigate bias during compression.

Jane: First, who's behind it and why it matters.

Paper summary: Tom: So we're talking about this new paper called Fair-GPTQ: Bias-Aware Quantization for Large Language Models. Basically, it tackles how compressing these massive language models can sometimes make them more biased.

Jane: That’s right. The authors are proposing a new quantization method that builds explicit rules into the compression process to reduce unfairness during that process. It addresses a problem where efficiency gains from quantization might come with a cost to model fairness, especially when dealing with stereotypes about gender, race, and religion.

Lu: What they claim is that they introduce group-fairness constraints directly into the quantization objective. This guides the rounding operation to produce less biased text generation for protected groups. They focus specifically on stereotype generation involving occupational bias and discriminatory language in those areas.

Meng: So, it’s not just about fitting the model into smaller memory; it's about making sure that when you shrink the model, you don't accidentally amplify existing biases in the data or the original training. That makes sense from an engineering standpoint.

Lalam: From my side, I see this as a way to improve how culture is represented by these models. If we can build in constraints that actively push against stereotypical outputs during compression, it helps make the resulting AI more aligned with fairness standards.

Tom: Exactly. The paper shows that the model quantized with Fair-GPTQ produces responses labeled “Unknown” when asked questions targeting socioeconomic and racial stereotypes, which is better than what you'd see from a standard GPTQ model.

Jane: That’s a concrete result they show in Figure one <ref:2509.15206#pg1>. It means for certain types of unfair questions, the compressed version rejects the stereotype instead of generating one. It demonstrates that the added constraints actually work to guide the AI toward unbiased text generation.

Lu: They are modifying the reconstruction objective used in compression methods like GPTQ by adding a bias-aware regularization term. They’ve derived a closed-form solution and an efficient implementation that keeps the computational complexity similar to standard GPTQ.

Meng: How do they actually implement this constraint without making the quantization process impossibly slow? That efficiency part is always tricky with these kinds of added objectives.

Lalam: The paper mentions modifying the objective function to include a term Wc, which measures how much the quantized model changes the representation gap between stereotyped and anti-stereotyped inputs. It couples minimizing reconstruction error with penalizing large errors on semantically paired inputs that differ only in stereotypical versus antistereotypical attributes.

Paper summary: Tom: So it’s not just minimizing error on any input, but specifically penalizing errors when the difference is purely about whether something is stereotyped or not, which targets the source of the bias amplification. That’s a clever way to focus the regularization.

Jane: The paper validates this approach on several recent base and instruction-tuned models, showing improvements in both fairness and performance trade-offs compared to baselines. They also did ablation studies looking at how applying Fair-GPTQ to different layers affects stereotype generation across protected groups.

Lu: They specifically analyze the effect of applying Fair-GPTQ to different layers and the role of a hyperparameter that controls the regularization strength relative to the reconstruction objective. It shows you can tune how aggressively it fights bias without completely destroying performance.

Meng: So, for someone building these models, they are giving you a way to dial in fairness rather than just accepting a trade-off where you pick one or the other blindly. That level of control over that regularization strength is what engineers need.

Lalam: For me, it means we can develop systems where the resulting AI output is intentionally steered away from harmful patterns, not just passively hoping it doesn't happen due to the compression process. It’s about building a culture into the quantization step.

Tom: So, we’ve heard that Fair-GPTQ is designed to reduce unfairness by explicitly constraining the quantization objective with group-fairness terms, and they show it improves performance while mitigating bias on stereotype generation tasks. Now we move into what this actually means for the broader picture of using these powerful language models.

Jane: The authors introduce Fair-GPTQ as a way to bridge the gap between making AI models efficient enough to run widely and ensuring those models are also fairer in their outputs. It moves the conversation from simply measuring bias after quantization to actively designing fairness into the compression itself.

Lu: They are taking existing techniques, like Self-Debias approach from Schick et al., and adapting them within the context of quantizing, showing how these concepts can be combined effectively. This suggests a path for incorporating social awareness directly into model optimization workflows.

Meng: Practically speaking, if this works well across various models they tested, it means that deploying powerful AI systems won't automatically mean we have to deal with amplified societal biases hidden in the compression layer. It’s a design consideration that needs to be front and center during deployment planning.

Lalam: If this method becomes standard, then the expectation for large language models shifts from just being accurate predictors to being actively vetted for fairness during their creation pipeline, starting right when you compress them for use.

Paper summary: Tom: So, Fair-GPTQ is a concrete step toward making model efficiency and ethical behavior go hand in hand by building fairness directly into the quantization process. We’ve seen how they modify the objective and how they validate it on various models.

Jane: It’s a practical technique that shows we can target specific biases like occupational bias or gender discrimination through these mathematical constraints, rather than just hoping for the best with post-hoc checks.

Lu: The theoretical framework is quite robust because they derived a closed-form solution and an efficient implementation that maintains the complexity of GPTQ. That makes it accessible to use in production environments.

Meng: For an engineer, knowing there’s a way to enforce these specific constraints on the rounding operation without adding massive overhead is crucial for practical adoption across different hardware setups.

Lalam: This research helps create a more responsible AI ecosystem where efficiency and ethical alignment are engineered into the core of model optimization from the start.

Tom: So, to wrap up this segment on Fair-GPTQ: it’s a method that adds group-fairness constraints to the quantization objective to reduce bias during compression, showing measurable improvements in avoiding stereotypical outputs.

Jane: The paper by Irina Proskurina et al., titled "Fair-GPTQ: Bias-Aware Quantization for Large Language Models", provides a clear path for incorporating fairness directly into the model compression pipeline. It shows that we can guide the rounding operation toward less biased text generation for protected groups, like those facing occupational bias or gender discrimination.

Lu: The key is modifying the reconstruction objective by introducing a bias penalty that measures how much the quantized model changes the representation gap between stereotyped and anti-stereotyped inputs, which they do by considering paired inputs that differ only in a single protected-attribute token.

Meng: The practical implication here is that deployment teams can now choose to use this method if mitigating certain types of stereotype amplification during compression is a priority, rather than just accepting the standard GPTQ path.

Lalam: This research helps create a more responsible AI ecosystem where efficiency and ethical alignment are engineered into the core of model optimization from the start.

Tom: That’s all we have time for on this paper. We’ve covered what Fair-GPTQ is, how it works conceptually, and what the authors claim about its impact on stereotype generation. Next up, we’re going to look at the broader implications for how we think about deploying these models responsibly.

Conclusion: Tom: So we're wrapping up this deep dive into Fair-GPTQ, which is about bringing fairness directly into how we compress big language models.

Jane: Right, so essentially, this paper is introducing a new way to do quantization—shrinking these massive models—that actively tries to stop them from amplifying existing social biases.

Lu: The authors are taking the standard GPTQ process and adding a specific regularization term into the objective function that focuses on fairness constraints during the rounding step.

Meng: From an engineering standpoint, they’ve done something interesting by deriving a closed-form solution so this method doesn't introduce huge computational headaches compared to just using regular GPTQ.

Lalam: What I find really compelling is how they define the penalty—it forces the quantization to be sensitive to whether two inputs are stereotyped or anti-stereotyped when they only differ by one attribute token.

Tom: So, if you think of it simply, instead of just making the model smaller and faster, this method makes it smarter about avoiding certain kinds of unfair outputs during that shrinking process.

Jane: Exactly. The authors show that when they test it on models like GPTQ, the compressed versions produce fewer stereotypical responses when asked about things like job types or gender roles.

Lu: They validated this by testing it on a range of instruction-tuned models and showed improvements in both how fair the model is and how well it still performs its general tasks.

Meng: It’s good that they looked at different layers too, figuring out which parts of the model are most sensitive to these bias penalties so you know where to apply this technique for maximum effect.

Lalam: For me, this means we can start designing deployment workflows where fairness isn't just a check you do later; it’s baked into the compression step itself.

Tom: So, the title Fair-GPTQ really sums up what they did: taking quantization and adding bias awareness to make sure the resulting model output is less biased.

Jane: It shifts the focus from fixing bias after training to designing systems that are inherently fairer during optimization and compression.

Lu: The paper shows a strong link between preserving reconstruction fidelity and controlling the fairness objective, which is a key piece of theoretical insight here.

Tom: So this isn't just a tweak; it’s a way to build social awareness right into the technical process of making these powerful language models usable.

Jane: And what they show suggests that we can start measuring and mitigating specific types of group bias much earlier in the lifecycle of an AI system.

Meng: This has big implications for how companies think about deploying these models; it means fairness becomes a necessary feature, not just an afterthought.

Lalam: If this approach scales well, it could really help us build a generation pipeline where we can intentionally steer the AI away from harmful patterns before they even get out.

Tom: So that’s our time on Fair-GPTQ for now. Next up, we'll talk about how this affects the bigger picture of building responsible AI systems.

Université Claude Bernard Lyon 1 · Université Lumière Lyon 2 · ERIC

cs.CL

Submitted: 2025-09-18

Updated: 2026-10-08

Comments: Accepted for publication in TACL. Pre-MIT Press publication version

Code: https://github.com/ModelCloud/GPTQMode

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 82/100

The gist: The gist The first quantization method explicitly designed to reduce unfairness in large language models, Fair-GPTQ, introduces explicit group-fairness constraints into the quantization objective to

Key concepts

Group Fairness
This concept defines fairness as achieving approximate parity in statistical outcomes across different groups. In the context of language models, it means ensuring the model's performance or generated content is similar for different demographic groups, preventing systematic bias against specific attributes like race or gender.
Stereotype Likelihood Bias
This measures how much a model systematically favors generating stereotypical continuations when given certain prompts. A positive score indicates the model has a learned tendency to produce biased responses related to social stereotypes in its text generation.
Quantization Objective Modification
The paper changes the mathematical goal of quantization by adding an extra penalty term. This term forces the quantization process to consider not just overall fidelity, but also how much it distorts the difference between inputs that are semantically similar except for a single protected attribute.
Fair-GPTQ Algorithm
This is a three-step procedure: first, an update step to debias the model; second, the actual weight quantization; and third, a compensation step. The key innovation is integrating the bias penalty into the reconstruction optimization during this process to ensure fairness is maintained alongside compression.

Terminology

Summary

The gist The first quantization method explicitly designed to reduce unfairness in large language models, Fair-GPTQ, introduces explicit group-fairness constraints into the quantization objective to mitigate bias during compression.

Introduction and Motivation

Autoregressive (causal) language models have demonstrated strong efficacy across a variety of conditional generative tasks, including question answering and commonsense reasoning Recent studies show that attaining state-of-the-art performance in these tasks generally requires scaling up both the volume of training data and the number of trainable parameters As model scale grows to meet accuracy requirements, posttraining compression techniques such as quantization are used to control memory and computational overhead However, “ex nihilo nihil fit”1, the efficiency gains come at a cost Recent studies have shown that quantization can amplify biases a priori present in large language models (LLMs) Existing work evaluates such biases empirically after quantization, without accounting for fairness during compression We address this gap by introducing FairGPTQ, a quantization method that incorporates group-fairness constraints to reduce bias during compression As shown in Figure 1, the model quantized with Fair-GPTQ produces unbiased responses (“Unknown”) to questions targeting socioeconomic and racial stereotypes, and rejects stereotypical statements about women, compared to GPTQ Our main contributions are as follows: (i) We modify the quantization objective used in compression approaches such as GPTQ (Frantar et al., 2022) by introducing a bias-aware regularization term (§3.1), we derive a closed-form solution and provide an efficient implementation that preserves the computational complexity of GPTQ (§3.2). (ii) We validate the implementation of the theoretical solution, Fair-GPTQ, on a range of recent base and instruction-tuned models, demonstrating improvements over baselines in fairness–performance and fairness-efficiency trade-offs (§5.1). (iii) We perform ablation studies to analyze the effect of applying Fair-GPTQ to different layers on stereotype generation across protected groups and the role of the introduced hyperparameter in controlling the regularization strength relative to the reconstruction objective (§5.2).

Background on Bias and Quantization

Group fairness is defined as approximate parity between statistical outcome measures across groups: µY (Mθ, Ga) − µY (Mθ, Gb) < ε, ∀ a, b ∈ A We consider two types of group bias evaluation: (i) likelihood-based measurement of stereotypical generations and (ii) generative bias in question answering Stereotype Likelihood Bias measures the average log-likelihood difference µlik(Mθ, Ga) = E log pθ(yst x) − E log pθ(yanti x), where positive values indicate a systematic shift toward the generation of stereotypical continuations Generative QA bias uses the benchmark-specific scoring function sθ(x, y) to compute µqa(Mθ, Ga) = E[sθ(x, y)], where differences across groups indicate disparities in how the model selects or scores answers across social groups Empirical studies discussed in §2.1 report that quantization can significantly degrade LLM accuracy on fairness tasks for group bias evaluation (§2.2) Figure 2 illustrates the restructured results from Xu et al. (2024), showing an increase in generative QA bias scores, for instance, on BBQ and UnQover benchmarks, measured as 1− accuracy in selecting “Unknown” answers We hypothesize that this increase in bias can be mitigated during quantization if the objective preserves not only the similarity between the pretrained and quantized model outputs (reconstruction fidelity), as in GPTQ, but also penalizes large reconstruction errors on semantically paired inputs that differ only in stereotypical versus antistereotypical attributes as in the question illustrated in Figure 1.

Fair-GPTQ Algorithm

The Fair-GPTQ procedure consists of three steps: (i) debiasing update, (ii) weight quantization, and (iii) quantization-error compensation We begin by reformulating the reconstruction optimization problem used in a range of quantization methods, including GPTQ (Frantar et al., 2022), Optimal Brain Surgeon (Hassibi et al., 1993), and other compression approaches In our formulation, this constraint appears as an additional regularization term in the compression objective, allowing the quantization procedure to account for differences in the representations of paired inputs We consider two matrices X0, X1 ∈ R d×m representing a pair of input texts of length m that differ only in a single protected-attribute token To make the quantization step sensitive to potential stereotypes, we introduce a bias penalty that measures how much the quantized model changes the representation gap between the stereotyped (X0) and anti-stereotyped (X1) inputs Formally, this can be restated as follows: Wc = arg min W′∥WX0 − W′X0∥2 +∥WX1 − W′X1∥2 + α∥W′(X0 − X1)∥2 In the following, for clarity, we define ∆W = W′ −W and ∆X = X0 − X1 We use row-wise flattening because the objective in Eq. (3) couples parameters only within the same row wi = Wi,: (the loss terms are sums of in-row squares) Consequently, the cross-row second derivatives vanish, ∂ 2f / ∂wi ∂w⊤ k = 0 for i ≠ k, where f denotes the objective function in Eq. (3).

Experimental Setup and Results

We apply Fair-GPTQ to the attention output projection and the output fully connected matrices in each layer, following Algorithm 1 We target these matrices because they contribute directly to the residual stream and strongly influence both bias and token generation, as shown in prior work (Elhage et al., 2021; Geva et al., 2021; Prakash and Lee, 2023; Zhou et al., 2024) All remaining matrices that do not contribute directly to the residual stream are quantized using GPTQ For quantization, we use a block size of B = 128 and a group size of g = 128, with each group associated with a scaling factor sg under a symmetric quantization grid All models are quantized to 4 bits (b = 4) on two 80 GB NVIDIA A100 GPUs We report (i) stereotype-bias scores, (ii) perplexity on WikiText-2, and (iii) zero-shot accuracy on scientific factual knowledge task ARC EASY and natural text entailment task HELLASWAG For group fairness evaluation (§2.2), we use two likelihood-based benchmarks, CrowS-Pairs (CP; Nangia et al. (2020)) and Co-occurrence tests (CooC; Brown et al.

Improvements for AI systems

  1. Bold header: Bias-Aware Quantization Integration

This improvement integrates explicit group-fairness constraints into the quantization objective, adding explicit group-fairness constraints to the quantization objective, guiding the learning of the rounding operation toward less-biased text generation for protected groups. This enables AI systems to generate outputs that are demonstrably less stereotypical across gender, race, and religion while maintaining 4-bit memory efficiency.

  1. Bold header: Performance Preservation via Debiasing Update

The Fair-GPTQ algorithm uses a debiasing update derived from the Jacobian, W ← W − (H−1HbiasW⊤), which is shown to preserve at least 90% of baseline accuracy on zero-shot benchmarks while actively reducing unfairness relative to a standard GPTQ model.

  1. Bold header: Targeted Layer-wise Bias Mitigation

By applying Fair-GPTQ to specific layer subsets, the system can achieve targeted, category-specific bias mitigation, such as increasing accuracy on religion-related questions by applying debiasing to lower layers while improving performance on race and nationality contexts via upper layers.

Abstract

The high memory demands of generative language models have drawn attention to quantization, which reduces memory usage by mapping model weights to lower-precision integers. However, recent empirical studies show that, while efficient, quantization can increase the likelihood of generating biased outputs and degrade performance on fairness benchmarks. In this work, we draw new links between quantization and model fairness by adding explicit group-fairness constraints to the quantization objective and introduce Fair-GPTQ, the first quantization method explicitly designed to reduce unfairness in large language models. The added constraints guide the learning of the rounding operation toward less-biased text generation for protected groups. Specifically, we focus on stereotype generation involving occupational bias and discriminatory language spanning gender, race, and religion. Fair-GPTQ has minimal impact on performance, preserving at least 90% of baseline accuracy on zero-shot benchmarks, reduces unfairness relative to a half-precision model, and retains the memory and speed benefits of 4-bit quantization.

Sources

Related papers