🤖 AI

Demographic Fairness in Multimodal LLMs: A Benchmark of Gender and Ethnicity Bias in Face Verification

This paper presents a benchmarking study evaluating the gender and ethnicity bias of nine open-source Multimodal Large Language Models in face verification tasks, revealing that specialized models outperform general-purpose ones while demonstrating that higher accuracy does not guarantee demographic fairness and that bias patterns differ significantly from traditional face recognition systems.

Original authors: Ünsal Öztürk, Hatef Otroshi Shahreza, Sébastien Marcel

Published 2026-03-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ünsal Öztürk, Hatef Otroshi Shahreza, Sébastien Marcel

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a new, super-smart robot assistant that can look at pictures and answer questions about them. You might ask it, "Is this person the same as that person?" This is called face verification.

For years, we've used specialized "face detectives" (traditional systems) for this job. They are like master fingerprint analysts: they turn a face into a secret code and compare the codes. But recently, we've started using these giant, general-purpose "brainy robots" (called Multimodal Large Language Models or MLLMs) to do the same thing. Instead of using secret codes, you just show them two photos and ask, "Are these the same person?" and they give you a score.

This paper is a big report card on how fair these new "brainy robots" are when they try to identify people from different backgrounds.

The Big Question: Are They Fair?

We know that old-school face detectives sometimes get it wrong more often for people with darker skin or specific ethnicities. It's like a security guard who is great at spotting one type of face but keeps misidentifying others.

The researchers wanted to know: Do these new, general-purpose AI robots have the same bias, or is it different?

The Experiment: A Test Drive

The researchers took 9 different AI models (ranging from small, 2-billion-parameter "puppies" to larger, 8-billion-parameter "dogs") and put them through a driving test.

The Course: They used two famous driving tracks (datasets): IJB-C (a huge, messy track with lots of different people) and RFW (a track specifically designed to test racial bias).
The Drivers: They tested how well these models could tell if two photos were the same person, breaking the results down by Ethnicity (African, East Asian, South Asian, Caucasian) and Gender (Male, Female).
The Scorecard: They didn't just look at who got the most answers right (accuracy). They also looked at Fairness. Did the robot make mistakes equally for everyone, or did it fail more often for specific groups?

The Surprising Results

1. The "Specialist" vs. The "Generalist"
There was one model in the test called FaceLLM-8B. Think of this model as a specialist detective who only studies faces. The other eight models were generalists (like a Swiss Army knife) that can do many things but aren't experts at faces.

Result: The specialist detective (FaceLLM) was way better at the job than the generalists. It made far fewer mistakes.
The Catch: Even the specialist wasn't as good as the old-school "code-based" systems. The generalist robots were often just guessing, performing no better than flipping a coin.

2. The Bias Flip-Flop
Here is the most interesting part. In traditional systems, people with darker skin often get the worst results.

The Twist: On the big, messy track (IJB-C), the new AI models actually struggled more with Caucasian faces than with others in many cases! It's like a security guard who is usually great at spotting everyone but suddenly gets confused by the people they know best.
On the other track (RFW): The bias flipped back. The models struggled most with African faces, which is what we usually expect.
Lesson: You can't assume these new AIs will have the same biases as the old ones. They are unpredictable.

3. The "Bad but Fair" Trap
The researchers found a tricky situation. Sometimes, a model is so bad at its job that it fails everyone equally.

Analogy: Imagine a terrible archer who misses the target every single time, no matter who is standing there. If they miss everyone equally, they look "fair" because the error rate is the same for all groups.
Reality: Being "fair" when you are useless isn't actually good! The most accurate models (like the specialist) weren't always the fairest, and the "fair" ones were often just the ones that were too dumb to get it right for anyone.

4. Gender vs. Race
The models were generally much fairer regarding gender (Male vs. Female) than they were regarding ethnicity. The gap between how well they treated men and women was tiny, but the gap between how they treated different ethnic groups was much wider.

The Bottom Line

This paper tells us that while these new "brainy robots" are exciting, we can't just trust them to do face verification yet.

They aren't as accurate as the old, specialized systems.
Their biases are weird and different from what we are used to (sometimes they mess up the "majority" group more!).
A model can look "fair" just because it's bad at its job.

The Takeaway: If we want to use these powerful new tools for security or ID checks, we need to be very careful. We can't just assume they are neutral. We need to test them constantly, just like we test a new car before letting it drive on the highway.

1. Problem Statement

While Multimodal Large Language Models (MLLMs) have demonstrated remarkable capabilities in visual reasoning and are increasingly being explored for face verification (determining if two face images belong to the same person), their demographic fairness remains largely unexplored.

The Gap: Traditional face recognition systems (embedding-based) are known to exhibit demographic bias (e.g., higher error rates for darker-skinned females). However, it is unclear if MLLMs, which rely on general visual reasoning and prompting rather than specialized identity embeddings, inherit, amplify, or exhibit different bias patterns.
The Risk: As MLLMs are considered for high-stakes applications (border control, law enforcement), understanding their error distribution across gender and ethnicity groups is critical to prevent discriminatory outcomes.

2. Methodology

The authors established a systematic benchmarking pipeline to evaluate nine open-source MLLMs across two standard face verification protocols.

Datasets and Protocols

IJB-C: A large-scale benchmark containing 3,531 subjects. The authors subsampled 10,000 template pairs. Labels include Gender (Male/Female) and Ethnicity (African, East Asian, South Asian, Caucasian).
RFW (Racial Faces in-the-Wild): A balanced benchmark specifically designed for ethnic bias, containing 24,000 pairs across four ethnic groups (African, Asian, Caucasian, Indian). It lacks gender annotations.
Protocol: 1:1 verification (genuine vs. impostor pairs).

Models Evaluated

Nine MLLMs from six families (2B to 8B parameters) were tested in a zero-shot setting (no fine-tuning on face data):

General-Purpose: Idefics3-8B, LLaVA-NeXT-Mistral-7B, Ovis1.5-Llama3-8B, Qwen2-VL (2B/7B), Qwen2.5-VL (3B/7B), Valley2.
Specialized: FaceLLM-8B (fine-tuned for face analysis).

Evaluation Pipeline

Prompting: MLLMs are prompted with two face images and asked to output a similarity score (0–100) on a single number.
Template Scoring: For templates containing multiple images/frames, all pairwise combinations are scored, and the final score is the arithmetic mean of these pairs.
Metrics:
- Accuracy: Equal Error Rate (EER) and True Match Rate (TMR) at fixed False Match Rates (10%, 1%, 0.1%).
- Fairness: Four FMR-based metrics calculated per demographic group:
  - $\Delta_F$ (Differential): Max minus Min FMR (Absolute gap).
  - $R_F$ (Ratio): Max divided by Min FMR (Relative gap).
  - $M_F$ (Ratio to Geometric Mean): Normalizes the worst-case group against the geometric mean of all groups.
  - $G_F$ (Gini Coefficient): Measures overall inequality across groups.
- Decidability ( $d'$ ): A threshold-independent measure of the separation between genuine and impostor score distributions.

3. Key Contributions

First Fairness Benchmark for MLLMs: This is the first study to systematically evaluate demographic bias in MLLMs for face verification, contrasting them with traditional embedding-based systems.
Discovery of Divergent Bias Patterns: The study reveals that bias patterns in MLLMs differ significantly from traditional systems. Notably, on the IJB-C dataset, Caucasian groups were often the most disadvantaged (highest error rates) in five of the eight models, whereas traditional systems typically show higher errors for darker-skinned groups.
Accuracy-Fairness Trade-off Analysis: The paper demonstrates that high accuracy does not guarantee fairness, and conversely, low accuracy can mask unfairness. Models with poor discrimination ability often produce uniformly high error rates across all groups, appearing "fair" (low Gini) simply because they fail everyone equally.
Specialized vs. General Models: It highlights that the only face-specialized model (FaceLLM-8B) significantly outperforms general-purpose MLLMs in both accuracy and specific fairness metrics (particularly for gender).

4. Key Results

Accuracy Performance

FaceLLM-8B is the clear leader, achieving a global EER of 5.13% on IJB-C and 29.46% on RFW.
General-purpose models performed significantly worse. On RFW, several models (e.g., Ovis1.5, Qwen2-VL-2B) approached 50% EER, equivalent to random guessing.
Template Size Impact: Performance was better on IJB-C than RFW because IJB-C templates contain more images (median 9 vs. 3), allowing the MLLM to average out noise more effectively.

Fairness Findings

Ethnicity Bias (IJB-C):
- In 5 out of 8 models, the Caucasian group had the highest EER, reversing the typical bias seen in traditional systems.
- FaceLLM-8B showed the lowest absolute differential ( $\Delta_F = 0.020$ ) but a higher Gini coefficient than some less accurate models, illustrating the complexity of fairness metrics.
Ethnicity Bias (RFW):
- Disparities were much larger on RFW. The African group was consistently the most disadvantaged across all models.
- FaceLLM-8B had the highest differential ( $\Delta_F = 0.478$ ) on RFW, despite having the best accuracy.
Gender Bias:
- Gender disparities were consistently smaller than ethnicity disparities across all models and operating points.
- FaceLLM-8B achieved near-perfect gender fairness on IJB-C ( $\Delta_F = 0.001$ , $G_F = 0.004$ ).

The "False Fairness" Phenomenon

Models with poor overall accuracy (e.g., Ovis1.5) often exhibited the lowest Gini coefficients (appearing fairest) because they failed to distinguish between genuine and impostor pairs for any group, resulting in uniformly high error rates.
Conversely, the most accurate model (FaceLLM-8B) showed higher relative disparities on RFW because it successfully separated scores for some groups better than others.

5. Significance and Conclusion

Paradigm Shift: The study challenges the assumption that bias in face recognition is universal. MLLMs introduce new, unpredictable bias patterns that do not simply mirror traditional embedding-based systems.
Metric Caution: It underscores the danger of relying on a single fairness metric. A model can appear fair due to low accuracy (uniform failure) or unfair due to high accuracy (differential success). A combination of absolute and relative metrics is required.
Practical Implications: Current general-purpose MLLMs are not yet ready for high-stakes face verification due to low accuracy and unpredictable bias. Specialized fine-tuning (as seen with FaceLLM-8B) is necessary to bridge the gap, though fairness must still be rigorously monitored.
Future Work: The authors suggest extending evaluations to intersectional attributes (e.g., ethnicity $\times$ gender), testing larger models (>13B), and analyzing the sensitivity of results to prompt engineering.

Project Page: https://www.idiap.ch/paper/mllm-fairness