Skip to content

Benchmarking Bias Mitigation Toward Fairness Without Harm from Vision to LVLMs

🕒 Published (v1): 2026-01-01 · Source: ICLR · Venue: ICLR 2026 · link

Why this paper was selected

ICLR 2026; first comprehensive benchmark comparing bias mitigation in VLMs

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

NH-Fair is a unified benchmark that evaluates bias mitigation methods under standardized data, metrics, and training protocols spanning classical vision models and large vision–language models (LVLMs). Key findings are that well-tuned ERM frequently matches or outperforms specialized debiasing algorithms, data augmentation is the most robust practical strategy for fairness without harm, and LVLMs—despite scale—remain susceptible to subgroup disparities comparable to task-specific vision models.

Problem

Prior fairness benchmarks (MEDFAIR, FFB, ABCFair) are siloed by domain, omit recent methods, or compare methods under inconsistent hyperparameter settings, making it impossible to accurately assess true utility–fairness trade-offs. No existing benchmark jointly evaluates classical supervised vision models and modern LVLMs under a shared evaluation protocol, leaving open questions about whether multimodal pretraining at scale inherently reduces disparities.

Method

NH-Fair operationalizes fairness without harm: a fairness-enhanced classifier \(h\) must satisfy, for every group \(a \in A\), $\(\mathbb{E}_{X,Y|A=a}[R(h(X),Y)] \leq \mathbb{E}_{X,Y|A=a}[R(h_{\text{erm}}(X),Y)]\)$ so that no subgroup is made worse off relative to an unconstrained ERM baseline.

Datasets: 7 image datasets covering facial attributes (CelebA, UTKFace, FairFace, Facet), medical imaging (HAM10000, Fitz17k), and spurious correlations (Waterbirds), with sensitive attributes including gender, race, age, and skin type.

Methods: 12 baselines spanning data-centric approaches (RandAugment, Mixup, Resampling, Bias Mimicking, FIS) and algorithmic approaches (LAFTR, FSCL, GapReg, MCDP, GroupDRO, DFR, OxonFair, Decoupled Classifier).

Zero-shot evaluation: CLIP and BLIP2 via text-image matching prompts; LVLMs (LLaVA v1.6 7B/13B/34B, Qwen2.5-VL 7B/32B/72B, Gemma 3 4B/12B/27B, Llama 3.2–11B/90B and 4-Scout–109B) via binary image-text-to-text prompts.

Model selection pipeline (two-stage): 1. DTO-based ERM selection: choose the ERM checkpoint minimizing Euclidean distance to the utopia point (each subgroup at its observed maximum accuracy). 2. Fairness-without-harm (FWH) selection: classify mitigation method checkpoints into Optimal / Sub-optimal / Degradation / Unwanted zones relative to the selected ERM; select in priority order Optimal → Sub-optimal → Degradation.

Hyperparameter study: Grid search over optimizer (SGD, Adam, AdamW, Adagrad), learning rate, weight decay, model depth, batch size, and pretrained initialization; >10,000 A100 GPU-hours expended.

Fairness metrics reported: Overall Accuracy Parity (Gap \(\downarrow\)), Max-Min Fairness (Worst \(\uparrow\)), Demographic Parity (DP \(\uparrow\)), Equalized Odds (EqOdd \(\uparrow\)).

Key Contributions

  • NH-Fair benchmark: first unified pipeline jointly evaluating supervised vision models and LVLMs under standardized data, metrics, and tuning protocols with a reproducible codebase.
  • ERM tuning study: identifies optimizer choice and pretrained initialization as the training decisions with the largest impact on the fairness–utility trade-off; batch size, weight decay, and model depth have weak or inconsistent effects.
  • Debiasing methods meta-analysis: Friedman + Nemenyi post-hoc analysis across 7 datasets showing that most specialized algorithms are statistically indistinguishable from a well-tuned ERM under FWH selection criteria.
  • Data augmentation finding: RandAugment—not designed for bias mitigation—consistently improves both fairness and utility across multiple datasets, emerging as the strongest practical strategy.
  • LVLM fairness analysis: zero-shot evaluation across 10+ LVLM variants showing that model scale correlates weakly with fairness; the Llama series in particular exhibits high sensitivity to dataset-specific biases despite large capacity.

Results

  • Well-tuned ERM is highly competitive: across 7 datasets, ERM selected via DTO matches or exceeds most specialized fairness algorithms on Gap, Worst, DP, and EqOdd; no single mitigation method dominates across all datasets.
  • RandAugment is consistently effective: improves both fairness and utility on multiple datasets without explicit fairness constraints, outperforming regularization-based methods that trade accuracy for parity.
  • Regularization methods (GapReg, MCDP) achieve strong fairness metric scores but incur utility–fairness trade-offs, placing models in Sub-optimal or Degradation zones under FWH selection.
  • LVLMs are not inherently fairer: LLaVA, Qwen2.5-VL, Gemma 3, and Llama variants achieve higher average accuracy than task-specific classifiers but exhibit comparable or similarly structured subgroup disparities; scaling from 7B → 90B+ shows weak correlation with fairness improvement.
  • Training choices dominate scaling: gains from switching optimizer or pretrained weights are typically larger than gains from increasing model size.
  • (Note: full per-dataset numeric tables in Table 2/3 are truncated in the provided text.)

Limitations

  • Compute constraints prevented exhaustive evaluation of all sensitive attributes within each dataset (e.g., gender in UTKFace omitted).
  • Waterbirds is acknowledged as potentially over-simplistic as a fairness test bed, since spurious correlations in synthetic datasets may not reflect real-world bias complexity.
  • Causal fairness approaches are excluded because they require causal graphs or intervention variables unavailable in the evaluated datasets.
  • Zero-shot LVLM evaluation uses fixed prompt templates; prompt sensitivity is not systematically analyzed.
  • Individual fairness and counterfactual fairness are out of scope.
  • Results cover binary sensitive attributes; intersectional fairness across combinations of attributes is not addressed.

Relevance to Vision-Language Models

NH-Fair provides the first large-scale, tuning-controlled head-to-head comparison of LVLMs (LLaVA, Qwen2.5-VL, Gemma 3, Llama 4) on group fairness tasks, directly challenging the assumption that diverse multimodal pretraining produces fairer representations. The finding that architectural and training-protocol choices matter more than scale for fairness has immediate implications for VLM practitioners choosing between model families for sensitive applications. The zero-shot LVLM evaluation protocol—framing binary visual classification as image-text-to-text generation—also establishes a reusable methodology for probing fairness in generative VLMs without supervised fine-tuning. More broadly, the benchmark contextualizes where LVLMs stand relative to the classical debiasing literature, filling a gap that prior VLM benchmarks (focused on capability) have left open.