Benchmarking Bias Mitigation Toward Fairness Without Harm from Vision to LVLMs

AI summary

Overview

Research area: Machine learning fairness, computer vision, and multimodal (vision–language) model evaluation.

Technical level: Intermediate. Readers need basic familiarity with classification metrics and fairness notions such as demographic parity and equalized odds, but the paper's argument is largely about benchmarking practice rather than new mathematics.

Scope: The paper introduces NH-Fair, a unified benchmark that compares bias mitigation methods and large vision–language models (LVLMs) under shared datasets, metrics, and hyperparameter tuning protocols, with the stated goal of achieving "fairness without harm" — reduced group disparity without degrading performance for any group.

What This Paper Is About

Fairness research is hard to compare across papers because studies use different datasets, different fairness metrics, different hyperparameter settings, and evaluate vision models and multimodal models in isolation. The authors build NH-Fair to evaluate bias mitigation methods and large vision–language models on the same footing, using seven datasets, four fairness metrics, independent hyperparameter search per dataset, and a two-stage model-selection procedure. The goal is to determine whether specialized debiasing algorithms actually beat a carefully tuned baseline, and whether large pretrained multimodal models are inherently fairer than task-specific vision models.

Key Contributions

1. A tuning-aware benchmark pipeline (NH-Fair). The authors standardize data, metrics, training protocols, and model selection across seven datasets, covering supervised (ERM and debiasing) and zero-shot (multimodal, LVLM) regimes, and release code at https://github.com/osu-srml/NH-Fair.

2. A systematic ERM tuning study. By sweeping training choices (optimizer, learning rate, weight decay, pretrained vs. scratch initialization, model depth, batch size) they identify which decisions most affect utility and subgroup disparities, producing guidelines intended to shrink the hyperparameter search space for practitioners.

3. A re-evaluation of bias mitigation methods against a strong ERM baseline. They compare 12 baseline algorithms spanning data-centric and algorithmic approaches and report that many specialized methods do not reliably outperform a well-tuned ERM, while data augmentation consistently delivers parity gains without utility loss.

4. An evaluation of multimodal models and LVLMs. They assess image–text matchers (CLIP, BLIP-2, FairerCLIP, CLIP-SFID) and LVLMs (LLaVA-1.6, Qwen2.5-VL, Gemma 3, Llama) in zero-shot prediction, showing that subgroup disparities persist and that scaling alone gives smaller fairness gains than changing architecture or training protocol.

Main Findings

• Optimizer and initialization matter most for fairness. Optimizer choice (SGD, Adam, AdamW, Adagrad) and its learning rate affect both fairness and utility, as does selecting the correct pretrained weights, while model depth, batch size, and weight decay have weaker or inconsistent effects. The authors note SGD works best on CelebA and Adam on Fitz17k as examples.

• A well-tuned ERM is a strong fairness baseline. Selected via the Distance to Optimal (DTO) criterion, ERM performs competitively with — and sometimes better than — specialized methods across Gap, Worst-group accuracy, DP, and EqOdd, and no single mitigation approach dominates across all seven datasets. Friedman tests with Nemenyi post-hoc analysis (visualized as critical difference plots) support this comparison.

• Data augmentation achieves fairness without harm. RandAugment, although not designed for bias mitigation, improves both fairness and accuracy across multiple datasets, making it the recommended first strategy before exploring specialized algorithms.

• Regularization-based fairness penalties show a clear trade-off. GapReg and MCDP score strongly on EqOdd and DP because they penalize subgroup disparity in the loss, but this pulls decision boundaries away from the utility-optimal surface, often lowering accuracy and occasionally worst-group accuracy.

• Contrastive representation learning is another promising route. FSCL, which clusters same-class representations across groups, improves fairness while keeping accuracy competitive.

• Waterbirds overstates algorithmic progress. Many methods (Mixup, Resampling, BM, OxonFair) improve both utility and fairness on Waterbirds, which the authors attribute to spurious background–object correlation being easier to resolve than socially meaningful group disparities. They caution against over-relying on domain-generalization datasets for fairness evaluation.

• LVLMs are not inherently fairer. On CelebA, a tuned ERM reaches 86.57 accuracy with a 6.76 gap, while CLIP shows a 15.97 gap, BLIP-2 an 18.09 gap, and LLaVA-1.6 34B only 44.83 accuracy with a 20.75 gap. Debiasing variants FairerCLIP and CLIP-SFID, which can use the validation set for tuning, still do not significantly resolve the issue.

• Balanced datasets favor LVLMs. On UTKFace, Qwen2.5-VL 72B reaches 97.25 accuracy with a 0.85 gap and Gemma 3 27B reaches 97.25 accuracy with a 0.85 gap, versus ERM's 92.75 accuracy and 2.26 gap. Qwen2.5-VL 72B sits closest to the outer envelope of the utility–parity landscape across datasets.

• Scaling is not enough. Across model scales, larger models generally raise average accuracy, but gaps remain non-trivial and sometimes widen (Llama 3.2-11B averages a 10.08 gap versus 11.09 for Llama 3.2-90B). The authors report that fairness gains from scaling are smaller than those from switching model family, so model choice should be explored before scale.

• Substantial compute was required. The authors report spending over 10,000 A100 GPU hours to obtain the benchmarking results.

Methodology in Plain English

The authors assemble seven public image datasets that include demographic annotations (CelebA, UTKFace, FairFace, Facet, HAM10000, Fitz17k) plus Waterbirds, which lacks demographic labels but is widely used for spurious-correlation testing. Each dataset has a fixed classification target and sensitive attribute — for example, CelebA predicts wavy hair with gender as the sensitive attribute, and HAM10000 predicts malignancy with age as the sensitive attribute.

They train a common ERM baseline with careful hyperparameter search per dataset, then select the best baseline using a two-stage procedure. First, the DTO rule picks the ERM model whose subgroup accuracies sit closest in Euclidean distance to a "utopia point" where both groups achieve their best observed performance. Second, mitigation methods are classified into four zones relative to that baseline: Optimal (better for both groups), Sub-optimal (fairness gained by reducing the advantaged group's performance), Degradation (both groups harmed), and Unwanted (disparity widened). Selection runs Optimal → Sub-optimal → Degradation, and every dataset–method pair is averaged over five runs.

For fairness they report Overall Accuracy Parity (the accuracy gap), Max–Min Fairness (worst-group performance), Demographic Parity, and Equalized Odds; utility is measured by accuracy except on the medical datasets, where AUC is used because of class imbalance. They also run a Friedman test with Nemenyi post-hoc analysis to check whether method differences are statistically meaningful, and evaluate zero-shot multimodal models with prompt templates that frame each task as binary questions.

Why This Matters

Impact on research. The paper argues that much prior fairness work fixes core training parameters across baselines and datasets, which can produce unfair comparisons and inflated claims of state-of-the-art fairness. It supplies a reproducible, tuning-aware protocol and challenges the field to report against a properly tuned ERM baseline rather than an undertrained one, while also flagging that spurious-correlation benchmarks may overstate how easy fairness problems are.

Real-world applications:

• Medical imaging. HAM10000 and Fitz17k evaluate malignancy prediction across age and skin-type groups, where unequal error rates carry direct clinical consequences.
• Facial analysis systems. CelebA, UTKFace, FairFace, and Facet cover hair attributes, gender, ethnicity, and face visibility across gender and race groups — settings relevant to face-related deployment.
• Multimodal assistants. The LVLM evaluation speaks to deployed vision–language systems such as LLaVA, Qwen2.5-VL, Gemma 3, and Llama, which the paper shows still produce subgroup disparities.
• Model selection and tuning workflows. The finding that optimizer, learning rate, and pretrained-weight choice dominate fairness outcomes gives teams a way to concentrate limited compute on the settings that matter.

Industry relevance. The paper explicitly frames the ERM study around industrial practice, noting that real workflows often prioritize hyperparameter optimization with tools like Ray Tune or Optuna on incumbent models rather than testing many fairness algorithms. The takeaway — tune heavily, try augmentation early, and evaluate architecture and pretraining choices before scaling — maps directly onto how production teams allocate compute.

Future Directions

1. Better benchmarks for true fairness challenges. The authors recommend datasets with coherent, well-justified, socially meaningful groupings and caution against substituting domain-generalization datasets such as Waterbirds or Colored MNIST, which may be over-simplistic.

2. Extending beyond group fairness. The benchmark deliberately excludes individual fairness (which needs a well-defined image similarity function) and counterfactual fairness (which needs causal graphs), leaving both as open evaluation problems.

3. Explaining why LVLM disparities persist. The paper shows pretraining choices can still impart bias but does not isolate which architectural or training-protocol factors drive the differences, pointing to follow-up work on what to change besides scale.

4. Broadening coverage. The authors note they did not exhaustively use all available sensitive attributes (for example, gender in UTKFace) because of computational cost, and they omit HAM10000 and Fitz17k from LVLM evaluation since those models do not yield calibrated probabilities for AUC — both are gaps later work could close.

Target Audience

Machine learning fairness researchers and benchmark builders; computer vision and multimodal practitioners deploying classifiers or LVLMs in high-stakes settings; ML engineers responsible for hyperparameter tuning and model selection; and policy or applied-ethics readers who need concrete evidence on whether larger pretrained models reduce or reproduce subgroup disparities.

Authors’ abstract

Machine learning models trained on real-world data often inherit and amplify biases against certain social groups, raising urgent concerns about their deployment at scale. While numerous bias mitigation methods have been proposed, comparing the effectiveness of bias mitigation methods remains difficult due to heterogeneous datasets, inconsistent fairness metrics, isolated evaluation of vision versus multi-modal models, and insufficient hyperparameter tuning that undermines fair comparisons. We introduce NH-Fair, a unified benchmark for fairness without harm that spans both vision models and large vision-language models (LVLMs) under standardized data, metrics, and training protocols, covering supervised and zero-shot regimes. Our key contributions are: (1) a systematic ERM tuning study that identifies training choices with large influence on both utility and disparities, yielding empirically grounded guidelines to help practitioners reduce expensive hyperparameter tuning space in achieving strong fairness and accuracy; (2) evidence that many debiasing methods do not reliably outperform a well-tuned ERM baseline, whereas a composite data-augmentation method consistently delivers parity gains without sacrificing utility, emerging as a promising practical strategy. (3) an analysis showing that while LVLMs achieve higher average accuracy, they still exhibit subgroup disparities, and gains from scaling are typically smaller than those from architectural or training-protocol choices. NH-Fair provides a reproducible, tuning-aware pipeline for rigorous, harm-aware fairness evaluation.