Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MM-IQ: Benchmarking Human-Like Abstraction and Reasoning in Multimodal Models

Huanqia Cai, Yijun Yang, Winston Hu

TL;DR

MM-IQ introduces a large-scale, eight-paradigm visual reasoning benchmark for multimodal models to quantify abstract reasoning independent of language and domain knowledge. It provides a 4,776-QA training set and a 2,710-item test set derived from authoritative sources, with careful data cleaning and leakage avoidance. Evaluations show a sizable gap between humans and current LMMs (humans ~51.27%, best LMM ~33.17%), with RL-trained baselines and long CoT prompting offering meaningful improvements, particularly on geometry and relational tasks. The work highlights the need for stronger perceptual and reasoning capabilities and proposes a route toward closer human-like multimodal intelligence via RL and structured reasoning.

Abstract

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 capabilities in multimodal systems. To address this crucial gap, we propose MM-IQ, a comprehensive evaluation framework, which comprises a large-scale training set with 4,776 visual reasoning problems and 2,710 meticulously curated test items spanning 8 distinct reasoning paradigms. Through systematic evaluation of existing 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 (33.17% vs. 25% baseline accuracy). This substantial performance chasm highlights the inadequacy of current multimodal models in approximating fundamental human reasoning capacities, underscoring the need for paradigm-shifting advancements to bridge this cognitive divide. Moreover, inspired by the recent surge of large reasoning models, we also release a multimodal reasoning model as the baseline that is trained via reinforcement learning with verifiable reward functions, reaching competitive performance to the state-of-the-art with a notably smaller model size.

MM-IQ: Benchmarking Human-Like Abstraction and Reasoning in Multimodal Models

TL;DR

MM-IQ introduces a large-scale, eight-paradigm visual reasoning benchmark for multimodal models to quantify abstract reasoning independent of language and domain knowledge. It provides a 4,776-QA training set and a 2,710-item test set derived from authoritative sources, with careful data cleaning and leakage avoidance. Evaluations show a sizable gap between humans and current LMMs (humans ~51.27%, best LMM ~33.17%), with RL-trained baselines and long CoT prompting offering meaningful improvements, particularly on geometry and relational tasks. The work highlights the need for stronger perceptual and reasoning capabilities and proposes a route toward closer human-like multimodal intelligence via RL and structured reasoning.

Abstract

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 capabilities in multimodal systems. To address this crucial gap, we propose MM-IQ, a comprehensive evaluation framework, which comprises a large-scale training set with 4,776 visual reasoning problems and 2,710 meticulously curated test items spanning 8 distinct reasoning paradigms. Through systematic evaluation of existing 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 (33.17% vs. 25% baseline accuracy). This substantial performance chasm highlights the inadequacy of current multimodal models in approximating fundamental human reasoning capacities, underscoring the need for paradigm-shifting advancements to bridge this cognitive divide. Moreover, inspired by the recent surge of large reasoning models, we also release a multimodal reasoning model as the baseline that is trained via reinforcement learning with verifiable reward functions, reaching competitive performance to the state-of-the-art with a notably smaller model size.

Paper Structure

This paper contains 20 sections, 22 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Construction of MM-IQ
  4. Data Collection
  5. Reasoning Paradigms of MM-IQ
  6. Experiments
  7. Experimental Setup
  8. Performance
  9. Empirical Analysis
  10. What Improvements does Reinforcement Learning Bring?
  11. Long CoT vs. Short CoT
  12. Failure Analysis of Non-Reasoning LMMs on MM-IQ
  13. Conclusion
  14. Appendix
  15. Experimental Setup
  16. ...and 5 more sections

Figures (22)

  • Figure 1: Left: Accuracy of large multimodal models vs. humans across eight reasoning paradigms of MM-IQ. Right: Visual examples of MM-IQ's reasoning paradigms (Detailed information can be found in Section \ref{['sec:reasoning pattern']}).
  • Figure 2: A visualized example of logical operation paradigm.
  • Figure 3: The test accuracy of Qwen2.5-VL-7B-Instruct on MM-IQ during RL training.
  • Figure 4: Left: Distribution over different error types across three representative LMMs. Right: Quantitative distribution of reasoning paradigms in MM-IQ's test set.
  • Figure 5: Proportions of incorrect visual understanding across eight reasoning paradigms.
  • ...and 17 more figures