HumaniBench: A Human-Centric Benchmark for Large Multimodal Models Evaluation

🌐 Website: vectorinstitute.github.io/humanibench  |  📄 Paper: arxiv.org/abs/2505.11454  |  📊 Dataset: Hugging Face

---

🧠 Overview

As multimodal generative AI systems become increasingly integrated into human-centered applications, evaluating their alignment with human values has become critical.

HumaniBench is the first comprehensive benchmark designed to evaluate Large Multimodal Models (LMMs) on seven Human-Centered AI (HCAI) principles:

• Fairness
• Ethics
• Understanding
• Reasoning
• Language Inclusivity
• Empathy
• Robustness

This repository provides code and scripts for evaluating LMMs across 7 human-aligned tasks.

---

📦 Features

• 📷 32,000+ Real-World Image–Question Pairs
• ✅ Human-Verified Ground Truth Annotations
• 🌐 Multilingual QA Support (10+ languages)
• 🧠 Open and Closed-Ended VQA Formats
• 🧪 Visual Robustness & Bias Stress Testing
• 📑 Chain-of-Thought Reasoning + Perceptual Grounding

---

📂 Evaluation Tasks Overview

Task	Focus	Folder
Task 1: Scene Understanding	Visual reasoning + bias/toxicity analysis in social attributes (gender, age, occupation, etc.)	code/task1_Scene_Understanding
Task 2: Instance Identity	Visual reasoning in culturally rich, socially grounded settings	code/task2_Instance_Identity
Task 3: Multiple Choice QA	Structured attribute recognition via multi-choice questions	code/task3_Multiple_Choice_VQA
Task 4: Multilingual Visual QA	VQA across 10+ languages, including low-resource ones	code/task4_Multilingual
Task 5: Visual Grounding	Bounding box localization of socially salient regions	code/task5_Visual_Grounding
Task 6: Empathetic Captioning	Human-style emotional captioning evaluation	code/task6_Empathetic_Captioning
Task 7: Image Resilience	Robustness testing via image perturbations	code/task7_Image_Resilience

🔍 Each task folder includes a README with setup instructions, task structure, and metrics.

---

🧬 Pipeline

Three-stage process:

1. Data Collection Curated from global news imagery, tagged by social attributes (age, gender, race, occupation, sport)

2. Annotation GPT-4o–assisted labeling + human expert verification

3. Evaluation Comprehensive scoring across:

   • Accuracy
   • Fairness
   • Robustness
   • Empathy
   • Faithfulness

---

🔑 Key Insights

• 🔍 Bias persists, especially across gender and race
• 🌐 Multilingual gaps affect low-resource language performance
• ❤️ Empathy and ethics vary significantly by model family
• 🧠 Chain-of-Thought reasoning improves performance but doesn’t fully mitigate bias
• 🧪 Robustness tests reveal fragility to noise, occlusion, and blur

---

📚 Citation

If you use HumaniBench or this evaluation suite in your work, please cite:

        @article{raza2025humanibench,
            title={Humanibench: A human-centric framework for large multimodal models evaluation},
            author={Raza, Shaina and Narayanan, Aravind and Khazaie, Vahid Reza and Vayani, Ashmal and Radwan, Ahmed Y and Chettiar, Mukund S and Singh, Amandeep and Shah, Mubarak and Pandya, Deval},
            journal={arXiv preprint arXiv:2505.11454},
            year={2025}
          }

---

📬 Contact

For questions, collaborations, or dataset access requests, please open an issue in this repository or contact the corresponding author at [email protected], as listed in the paper.

---

⚡ HumaniBench promotes trustworthy, fair, and human-centered multimodal AI.

We invite researchers, developers, and policymakers to explore, evaluate, and extend HumaniBench. 🚀

---