🤖 AI

VLAgeBench: Benchmarking Large Vision-Language Models for Zero-Shot Human Age Estimation

This paper introduces VLAgeBench, a comprehensive zero-shot benchmark evaluating state-of-the-art Large Vision-Language Models (GPT-4o, Claude 3.5 Sonnet, and LLaMA 3.2 Vision) on facial age estimation across UTKFace and FG-NET datasets, demonstrating their competitive performance while highlighting critical challenges in fairness, prompt sensitivity, and interpretability.

Original authors: Rakib Hossain Sajib, Md Kishor Morol, Rajan Das Gupta, Mohammad Sakib Mahmood, Shuvra Smaran Das

Published 2026-03-30

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Rakib Hossain Sajib, Md Kishor Morol, Rajan Das Gupta, Mohammad Sakib Mahmood, Shuvra Smaran Das

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 are at a party, and someone asks you to guess the age of a stranger just by looking at their face. You might squint at their wrinkles, check their skin texture, and maybe guess, "Hmm, they look about 35." You aren't a doctor, and you haven't studied that specific person before, but you use your general life experience to make a good guess.

This paper is about teaching super-smart AI computers to do exactly that, but with a twist: they have to guess the age of anyone they've never seen before, without any special training.

Here is the breakdown of the research, explained simply:

1. The Big Idea: The "Universal Guessers"

For years, computers needed to be "taught" how to guess ages. It was like hiring a tutor to study thousands of photos of 20-year-olds, then thousands of 50-year-olds, until the computer memorized the patterns. This is called supervised learning.

But recently, a new type of AI called a Large Vision-Language Model (LVLM) has emerged. Think of these models as super-readers who have also seen the whole internet. They know what a face looks like, what wrinkles mean, and how skin changes over time, just by reading and seeing everything online.

The researchers asked: Can these super-readers guess a person's age just by looking at a photo, without us teaching them specifically how to do it? This is called Zero-Shot Learning (meaning "zero practice shots").

2. The Test: The "Age Guessing Olympics"

To find out, the researchers set up a competition. They took two famous photo albums of faces (called UTKFace and FG-NET) containing people of all ages, races, and genders.

They gave these photos to three top-tier AI "guessers":

GPT-4o (The current heavyweight champion from OpenAI).
Claude 3.5 Sonnet (A very sharp competitor from Anthropic).
LLaMA 3.2 Vision (A powerful open-source model, like a community-built tool).

They didn't tweak the settings or teach them anything new. They just handed them a photo and said, "Tell me the number of years this person has been alive."

3. The Results: Who Won?

The results were surprisingly good! These general-purpose AI models didn't need to be trained to be decent at this task.

The Gold Medalist: GPT-4o was the most accurate. On the big, messy dataset (UTKFace), it guessed the age within about 5 years of the real age on average. On the smaller dataset, it was even better.
The Silver Medalist: Claude 3.5 Sonnet did very well too, especially on the smaller dataset where the age range was narrower.
The Bronze Medalist: LLaMA 3.2 Vision did a solid job, proving that open-source models are catching up, though it made slightly more mistakes than the big commercial ones.

The Analogy: Imagine a general knowledge quiz. GPT-4o is like the person who reads every book in the library and can guess your age just by looking at you. LLaMA is like a very smart student who read a lot of books but maybe missed a few chapters. They both did better than expected for not having studied the specific quiz beforehand.

4. The Catch: It's Not Perfect Yet

While the AI did well, the researchers found some "glitches" in the system:

The "Baby" Problem: The AI sometimes struggles with very young children. If a baby is 2 years old, a small error (guessing 4) feels huge in percentage terms. The math gets tricky here.
The "Bias" Problem: Just like humans, these AIs can have hidden biases. They might guess ages differently depending on a person's race or gender because the data they learned from (the internet) has those biases built-in.
The "Prompt" Sensitivity: How you ask the question matters. If you ask the AI, "How old is this?" it might give a different answer than if you ask, "Estimate the age." The AI is very sensitive to the exact words used.

5. Why Does This Matter?

This research is a big deal because it shows we might not need to build a new, specialized computer brain for every single task (like age guessing, gender detection, or emotion reading).

Instead, we can use these general-purpose "super-brains" and just ask them to do the job. This is faster, cheaper, and easier.

Real-world uses could include:

Healthcare: Quickly estimating a patient's age from a photo to check if their biological age matches their calendar age.
Security: Verifying if someone looks old enough to enter a club or buy something, without needing a database of their specific face.
Forensics: Helping investigators estimate the age of a person in an old, blurry photo.

The Bottom Line

The paper concludes that these new AI models are like gifted amateurs who can guess ages almost as well as professional experts, even without training. They aren't perfect yet—they need to be fairer and more consistent—but they are a huge step forward. It's like discovering that your smart assistant can suddenly do your taxes just by reading a manual, without you hiring an accountant first.

1. Problem Statement

Human age estimation from facial images is a critical task in biometrics, healthcare, and human-computer interaction. Traditionally, this problem is addressed using supervised deep learning models (e.g., CNNs, ViTs) trained on large, labeled datasets like UTKFace and FG-NET. However, these traditional approaches suffer from:

Poor Generalization: They often fail to generalize across diverse ethnicities, lighting conditions, and facial expressions when test data differs from training data.
Data Dependency: They require extensive labeled datasets and domain-specific fine-tuning.
Rigid Architecture: They are typically designed for classification or specific regression tasks, lacking the flexibility of general-purpose models.

While Large Vision-Language Models (LVLMs) have demonstrated strong zero-shot capabilities in image captioning and classification, their efficacy in precise, regression-based tasks like predicting continuous numerical age values without any fine-tuning remains largely unexplored.

2. Methodology

The study introduces VLAgeBench, a comprehensive zero-shot benchmark to evaluate state-of-the-art LVLMs on facial age estimation.

Models Evaluated:
- GPT-4o (Commercial, gpt-4o-2024-08-06)
- Claude 3.5 Sonnet (Commercial, claude-3-5-sonnet-20241022)
- LLaMA 3.2 Vision (Open-source, LLaMA-3.2-11B-Vision-Instruct)
Datasets:
- UTKFace: >20,000 images, ages 0–116, diverse demographics (gender, ethnicity).
- FG-NET: 1,002 images from 82 individuals, ages 0–69, known for lower resolution and historical lighting challenges.
Experimental Setup (Zero-Shot):
- No Fine-Tuning: Models were evaluated strictly in a zero-shot setting with no task-specific training or parameter updates.
- Standardized Prompting: A uniform prompt was used across all models: "Estimate the age as a numeric value only. If the age cannot be estimated, return your best guess. Do not include any additional text, explanations, or characters."
- Hardware: LLaMA 3.2 Vision was run on an NVIDIA A100 GPU; commercial models were accessed via API.
Evaluation Metrics:
The study utilized eight metrics to capture both accuracy and agreement:
1. Mean Absolute Error (MAE)
2. Mean Squared Error (MSE)
3. Root Mean Squared Error (RMSE)
4. Mean Absolute Percentage Error (MAPE)
5. Mean Bias Error (MBE)
6. Coefficient of Determination ( $R^2$ )
7. Concordance Correlation Coefficient (CCC)
8. Accuracy within ±5 years

3. Key Contributions

First Systematic Zero-Shot Benchmark: This work presents the first comprehensive evaluation of LVLMs for direct numeric age regression without fine-tuning.
Multi-Metric Analysis: It employs a rigorous set of eight complementary metrics, moving beyond simple error rates to assess prediction agreement and variance explanation.
Demographic Fairness Analysis: The study analyzes performance disparities across different population subgroups, highlighting issues related to bias and robustness in real-world deployment.
Reproducibility: The authors provide open-source experimental protocols and implementation details to facilitate future research in multimodal biometric AI.

4. Key Results

The performance varied significantly based on the model and the dataset characteristics:

Overall Performance:
- GPT-4o achieved the best overall performance, particularly on the diverse UTKFace dataset.
  - UTKFace: MAE = 4.93, $R^2$ = 0.88, CCC = 0.94, ±5yr Accuracy = 66.42%.
  - FG-NET: MAE = 3.73, ±5yr Accuracy = 74.80%.
- Claude 3.5 Sonnet performed competitively, excelling on the smaller FG-NET dataset.
  - FG-NET: MAE = 3.40 (Best overall), ±5yr Accuracy = 80.97%.
- LLaMA 3.2 Vision (Open-source) showed solid but lower performance compared to commercial models.
  - UTKFace: MAE = 7.58, $R^2$ = 0.72.
Dataset Sensitivity:
- All models performed better on FG-NET (narrower age range 0–69) than on UTKFace (wider range 0–116), indicating that broader age ranges introduce higher complexity and error.
- MAPE Limitations: The study found that MAPE values were inflated for young subjects (especially in FG-NET), where small absolute errors result in disproportionately high percentage errors. This suggests MAPE is a poor metric for datasets with many young individuals.
Observations:
- Prompt Sensitivity: Performance is highly dependent on prompt design.
- Bias: Models exhibited demographic biases, with performance disparities linked to ethnicity and gender, underscoring the need for fairness-aware inference.
- Computational Cost: LVLM inference is significantly more computationally expensive than traditional CNN-based estimators.

5. Significance and Future Directions

Emergent Capabilities: The study proves that general-purpose LVLMs possess emergent capabilities for precise biometric regression tasks without specialized training, challenging the dominance of domain-specific CNNs.
Practical Viability: With accuracies exceeding 66% (within ±5 years) on diverse datasets, LVLMs are positioned as viable tools for forensic science, healthcare monitoring, and HCI.
Future Work: The authors suggest future research should focus on:
- Fine-tuning LVLMs for specific biometric tasks.
- Developing bias-aware prompting and post-hoc calibration.
- Exploring few-shot learning settings.
- Addressing interpretability (currently limited by the "black-box" nature of commercial APIs).

In conclusion, VLAgeBench establishes a new baseline for age estimation, demonstrating that while LVLMs are powerful zero-shot tools, challenges regarding fairness, prompt sensitivity, and computational efficiency must be addressed for widespread real-world adoption.