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

TL;DR

This paper introduces MM-IQ, a benchmark for evaluating human-like abstraction and reasoning in multimodal models, revealing current models' significant limitations and proposing a new baseline trained with reinforcement learning.

Contribution

The paper presents MM-IQ, a large-scale multimodal reasoning benchmark, and provides a baseline model trained with reinforcement learning to advance AI reasoning capabilities.

Findings

Current models perform only slightly better than chance on MM-IQ.

Existing models show significant gaps in human-like reasoning abilities.

A new reinforcement learning-based baseline achieves competitive performance with smaller size.

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.