Current evaluations of audio-visual intelligence in Omni-MLLMs typically rely on one or more isolated tasks,
overlooking the homogeneity and heterogeneity  among these tasks. 

This limitation hinders a deeper assessment of Omni-MLLMs’ audio-visual intelligence and restricts further exploration of the underlying capabilities. 

Inspired by human perceptual processes, we categorize the evaluation of audio-visual tasks into three stages:
Perception, Understanding, and Reasoning. 

In addition, to assess whether Omni-MLLMs exhibit generalization abilities akin to human audio-visual intelligence, we additionally construct out-of-domain datasets to evaluate their Primitive Sensation capabilities. 

Experimental results on Omni-MLLMs demonstrate that the (1) reasoning abilities are constrained by their perception and understanding capacities. Moreover, all tested Omni-MLLMs perform poorly in the primitive sensation evaluation, indicating that (2) out-of-domain generalization of audio-visual intelligence remains a significant challenge for current Omni-MLLMs.