Logo MM-IQ

Benchmarking Human-Like Abstraction and Reasoning in Multimodal Models

Huanqia Cai, Yijun Yang, Winston Hu
Tencent Hunyuan Team
caihuanqia19@mails.ucas.ac.cn
Paper Code

🤗

 Dataset

🏆

 Leaderboard
combined reasoning

Introduction

IQ testing has served as a foundational methodology for evaluating human cognitive capabilities, deliberately decoupling assessment from linguistic background, language proficiency, or domain-specific knowledge to isolate core competencies in abstraction and reasoning. Yet, artificial intelligence research currently lacks systematic benchmarks to quantify these critical cognitive dimensions in multimodal systems. To address this critical gap, we propose Logo MM-IQ, a comprehensive evaluation framework comprising 2,710 meticulously curated test items spanning 8 distinct reasoning paradigms.

Through systematic evaluation of leading open-source and proprietary multimodal models, our benchmark reveals striking limitations: even state-of-the-art architectures achieve only marginally superior performance to random chance (27.49% vs. 25% baseline accuracy). This substantial performance chasm highlights the inadequacy of current multimodal systems in approximating fundamental human reasoning capacities, underscoring the need for paradigm-shifting advancements to bridge this cognitive divide.

Leaderboard on Logo MM-IQ

Accuracy scores on the test set (2,710 problems) of Logo MM-IQ.

# Model Type Source Date Mean LO Math 2D-G 3D-G VI TM SR CO
1 Human Performance* 🥇 - Link 2024-09-04 51.27 61.36 45.03 60.11 47.48 46.67 55.61 36.63 65.79
2 Claude-3.5-sonnet 🥈 Proprietary Link 2024-11-14 27.49 23.41 29.48 26.60 24.37 35.56 25.69 27.72 42.11
3 QVQ-72B-Preview 🥉 Open-Source 🖼️ Link 2024-11-22 26.94 28.91 25.59 29.23 26.38 26.67 25.43 22.77 34.21
4 GPT-4o Proprietary Link 2024-11-12 26.87 25.52 25.70 28.32 27.64 26.67 25.69 27.72 50.00
5 Gemini-1.5-pro-002 Proprietary Link 2024-11-09 26.86 19.53 27.43 28.03 25.88 24.44 31.17 25.74 39.47
6 Qwen2-VL-72B-Instruct Open-Source 🖼️ Link 2024-11-13 26.38 24.74 24.40 28.60 27.39 24.44 26.93 32.67 23.68
7 Deepseek-vl-7b-chat Open-Source 🖼️ Link 2024-11-14 22.17 19.53 20.30 22.25 27.39 35.56 23.72 24.75 15.79
8 LLaVA-1.6-7B Open-Source 🖼️ Link 2024-11-03 19.45 24.22 20.34 17.92 15.83 20.00 18.23 17.82 18.42
Human Performance*: Average human performance from annotators who have high school diplomas or above.
Reasoning Paradigm: LO: Logical Operation, 2D-G: 2D-Geometry, 3D-G: 3D-Geometry,
VI: Visual Instruction, TM: Temporal Movement, SR: Spatial Relationship, CO: Concrete Object.

🚨 To submit your results to the leaderboard, please send to this email with your result json files.

Logo MM-IQ Dataset

Examples



BibTeX

@article{cai2025mm,
  title={MM-IQ: Benchmarking Human-Like Abstraction and Reasoning in Multimodal Models},
  author={Cai, Huanqia and Yang, Yijun and Hu, Winston},
  journal={arXiv preprint arXiv:2502.00698},
  year={2025}
}