AVI-Bench: Toward Human-like Audio-Visual Intelligence of Omni-MLLMs

A new benchmark called AVI-Bench systematically evaluates the audio-visual intelligence of Omni-Multimodal Large Language Models, revealing substantial limitations in current systems, according to research submitted to arXiv on June 1, 2026 [1][2]. The benchmark, introduced by researchers affiliated with Fudan University, assesses models across three cognitively inspired stages: perception, understanding, and reasoning [1][2]. It requires models to perform cross-modal tasks that demand joint interpretation of audio and visual inputs, enabling a fine-grained diagnosis of both capabilities and failure modes [2]. The work arrives as Omni-MLLMs — models integrating vision, audio, and language — have drawn increasing attention, though systematic evaluation of their combined audio-visual faculties has lagged [2]. To probe robustness beyond familiar training domains, the team also developed AVI-Bench-PriSe, an extension that tests models' primitive audio-visual sensation using unfamiliar, low-semantic stimuli [1][2]. This design is intended to measure generalization beyond common training distributions [2]. Based on experimental results across both open-source and closed-source models, the authors present a four-level taxonomy of audio-visual intelligence [1][2]. The paper appears on arXiv, a preprint repository that has integrated with platforms such as Hugging Face to make research more accessible. Since November 2022, arXiv has collaborated with Hugging Face to embed interactive demos directly alongside paper abstracts, allowing users to test models without writing code [5]. Authors can link their work to demos by including a paper citation in a Hugging Face Space's README file or by associating a model on the Hugging Face Hub with the paper [6]. The AVI-Bench project page is hosted at fudancvl.github.io/AVI-Bench [2]. Large language models have proliferated rapidly. DeepSeek, a Chinese AI firm founded in July 2023, launched its R1 model in January 2025 with performance comparable to OpenAI's GPT-4, while reporting training costs of $6 million — far below the estimated $100 million for GPT-4 [7]. Alibaba Cloud's Qwen family of models is distributed under open-source licenses including Apache 2.0 [9]. The AVI-Bench authors argue that current Omni-MLLMs still fall short of robust audio-visual intelligence, and they position their benchmark as a principled framework to guide development of more generalizable systems [2].