XModBench: Benchmarking Cross-Modal Capabilities and Consistency in Omni-Language Models

XModBench is a comprehensive benchmark designed to evaluate the cross-modal capabilities and consistency of omni-language models. It systematically assesses model performance across multiple modalities (text, vision, audio) and various cognitive tasks, revealing critical gaps in current state-of-the-art models.

Key Features

• 🎯 Multi-Modal Evaluation: Comprehensive testing across text, vision, and audio modalities
• 🧩 5 Task Dimensions: Perception, Spatial, Temporal, Linguistic, and Knowledge tasks
• 📊 13 SOTA Models Evaluated: Including Gemini 2.5 Pro, Qwen2.5-Omni, EchoInk-R1, and more
• 🔄 Consistency Analysis: Measures performance stability across different modal configurations
• 👥 Human Performance Baseline: Establishes human-level benchmarks for comparison

🚀 Quick Start

Installation

# Clone the repository
git clone https://github.com/XingruiWang/XModBench.git
cd XModBench

# Install dependencies
pip install -r requirements.txt

📂 Dataset Structure

Download and Setup

After cloning from HuggingFace, you'll need to extract the data:

# Download the dataset from HuggingFace
git clone https://huggingface.co/datasets/RyanWW/XModBench

cd XModBench

# Extract the Data.zip file
unzip Data.zip

# Now you have the following structure:

Directory Structure

XModBench/
├── Data/                              # Unzipped from Data.zip
│   ├── landscape_audiobench/          # Nature sound scenes
│   ├── emotions/                      # Emotion classification data
│   ├── solos_processed/               # Musical instrument solos
│   ├── gtzan-dataset-music-genre-classification/  # Music genre data
│   ├── singers_data_processed/        # Singer identification
│   ├── temporal_audiobench/           # Temporal reasoning tasks
│   ├── urbansas_samples_videos_filtered/  # Urban 3D movements
│   ├── STARSS23_processed_augmented/  # Spatial audio panorama
│   ├── vggss_audio_bench/             # Fine-grained audio-visual
│   ├── URMP_processed/                # Musical instrument arrangements
│   ├── ExtremCountAV/                 # Counting tasks
│   ├── posters/                       # Movie posters
│   └── trailer_clips/                 # Movie trailers
│
└── tasks/                             # Task configurations (ready to use)
    ├── 01_perception/                 # Perception tasks
    │   ├── finegrained/               # Fine-grained recognition
    │   ├── natures/                   # Nature scenes
    │   ├── instruments/               # Musical instruments
    │   ├── instruments_comp/          # Instrument compositions
    │   └── general_activities/        # General activities
    ├── 02_spatial/                    # Spatial reasoning tasks
    │   ├── 3D_movements/              # 3D movement tracking
    │   ├── panaroma/                  # Panoramic spatial audio
    │   └── arrangements/              # Spatial arrangements
    ├── 03_speech/                     # Speech and language tasks
    │   ├── recognition/               # Speech recognition
    │   └── translation/               # Translation
    ├── 04_temporal/                   # Temporal reasoning tasks
    │   ├── count/                     # Temporal counting
    │   ├── order/                     # Temporal ordering
    │   └── calculation/               # Temporal calculations
    └── 05_Exteral/                    # Additional classification tasks
        ├── emotion_classification/    # Emotion recognition
        ├── music_genre_classification/ # Music genre
        ├── singer_identification/     # Singer identification
        └── movie_matching/            # Movie matching

Note: All file paths in the task JSON files use relative paths (./benchmark/Data/...), so ensure your working directory is set correctly when running evaluations.

Basic Usage

#!/bin/bash
#SBATCH --job-name=VLM_eval        
#SBATCH --output=log/job_%j.out
#SBATCH --error=log/job_%j.log                        
#SBATCH --ntasks-per-node=1
#SBATCH --gpus-per-node=4

echo "Running on host: $(hostname)"
echo "CUDA_VISIBLE_DEVICES=$CUDA_VISIBLE_DEVICES"

module load conda
# conda activate vlm
conda activate omni

export audioBench='/home/xwang378/scratch/2025/AudioBench'

# python $audioBench/scripts/run.py \
#     --model gemini \
#     --task_name perception/vggss_audio_vision \
#     --sample 1000


# python $audioBench/scripts/run.py \
#     --model gemini \
#     --task_name perception/vggss_vision_audio \
#     --sample 1000

# python $audioBench/scripts/run.py \
#     --model gemini \
#     --task_name perception/vggss_vision_text \
#     --sample 1000

# python $audioBench/scripts/run.py \
#     --model gemini \
#     --task_name perception/vggss_audio_text \
#     --sample 1000

# Qwen2.5-Omni

# python $audioBench/scripts/run.py \
#         --model qwen2.5_omni \
#         --task_name perception/vggss_audio_text \
#         --sample 1000

python $audioBench/scripts/run.py \
        --model qwen2.5_omni \
        --task_name perception/vggss_vision_text \
        --sample 1000

📈 Benchmark Results

Overall Performance Comparison

Model	Perception	Spatial	Temporal	Linguistic	Knowledge	Average
Gemini 2.5 Pro	75.9%	50.1%	60.8%	76.8%	89.3%	70.6%
Human Performance	91.0%	89.7%	88.9%	93.9%	93.9%	91.5%

Key Findings

1️⃣ Task Competence Gaps

• Strong Performance: Perception and linguistic tasks (~75% for best models)
• Weak Performance: Spatial (50.1%) and temporal reasoning (60.8%)
• Performance Drop: 15-25 points decrease in spatial/temporal vs. perception tasks

2️⃣ Modality Disparity

• Audio vs. Text: 20-49 point performance drop
• Audio vs. Vision: 33-point average gap
• Vision vs. Text: ~15-point disparity
• Consistency: Best models show 10-12 point standard deviation

3️⃣ Directional Imbalance

• Vision↔Text: 9-17 point gaps between directions
• Audio↔Text: 6-8 point asymmetries
• Root Cause: Training data imbalance favoring image-to-text over inverse directions

📝 Citation

If you use XModBench in your research, please cite our paper:

@article{wang2024xmodbench,
  title={XModBench: Benchmarking Cross-Modal Capabilities and Consistency in Omni-Language Models},
  author={Wang, Xingrui, etc.},
  journal={arXiv preprint arXiv:2510.15148},
  year={2024}
}

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

🙏 Acknowledgments

We thank all contributors and the research community for their valuable feedback and suggestions.

📧 Contact

• Project Lead: Xingrui Wang
• Email: [xwang378@jh.edu]
• Website: https://xingruiwang.github.io/projects/XModBench/

🔗 Links

• Project Website
• Paper
• Leaderboard
• Documentation

Todo

• Release Huggingface data
• Release data processing code
• Release data evaluation code

---

Note: XModBench is actively maintained and regularly updated with new models and evaluation metrics. For the latest updates, please check our releases page.