naplab/AVMeme-Exam

A Multimodal Multilingual Multicultural Benchmark for LLMs' Contextual and Cultural Knowledge and Thinking Beyond Text

[arXiv]  |  [demo]

Overview

AVMeme Exam is a multimodal benchmark for testing whether multimodal large language models (MLLMs) can really understand audio-visual memes, the short, widespread clips you see every day on YouTube, TikTok, Bilibili, X, etc. Beyond recognizing “what is said” or “what object is shown”, it evaluates deeper, richer multimedia knowledge and thinking: cultural context, implied meaning, induced emotion, use of audio/video, and world knowledge.

Motivation

We believe that AGI should recognize media beyond text and understand the context, emotion, culture, and usage of audio and video, just like humans. Existing multimodal benchmarks primarily focus on standard ASR, sound and object recognition, and NLP tasks. Even advanced multimodal understanding largely evaluates explicit content hearable or visible within audio or video frames.

Therefore, here we test whether MLLMs can understand audio-visual memes, to think and feel above audio and video content, and where they succeed or fail.

Dataset

• Multimodal: Audio, video, and text
• Multilingual: Multiple languages, including English, Chinese, Japanese, Korean, Hindi, Persian, and others
• Multicultural: Diverse cultural contexts and references

There are two JSON files: avmeme_full is our full test set of 1032 meme clips. avmeme_main is a harder subset of 846 meme clips.

Videos are downsampled to 360p at 1 FPS. If you need higher resolution or frame rates, please download the original videos from the URLs with appropriate tools.

Data Schema

Field	Description
category	Type of meme audio (e.g., clean speech, noisy speech, song, music, sound effect)
name	Meme identifier or commonly used title
url	Original source URL of the meme
onset	Start time of the selected segment (in seconds)
offset	End time of the selected segment (in seconds)
original_date	Year when the meme or source content first appeared
language	Primary spoken or sung language (or N/A for non-verbal audio)
transcription	Speech transcription if available; null for non-verbal audio
summary	Description of the clip content and scenario
emotion	Annotated emotional cues conveyed by the meme (e.g., sarcastic, happy, nostalgic)
sensitivity	Content safety tags (e.g., none, sex, violence, drug/alcohol)
usage	Typical meme usage or communicative intent in online contexts
question	Multiple-choice question designed to test meme understanding
choices	List of candidate answer options
solution	Correct answer
question_type	Question category in Language Analysis, Audio Analysis, Contextual Inference, Emotion Analysis, Humor & Popularity, Usage & Application, and World Knowledge
video_path	Relative path to the trimmed video clip. If visual_cheat=True, this field is None, meaning LLM should NOT see the visual.
audio_path	Relative path to the trimmed audio clip
visual_hint	Optional visual cues available to the model (e.g., transcription, title text, irrelevant text)
visual_cheat	Boolean flag indicating whether visual information directly reveals the answer
question_MC	The exact prompt to test LLM in multiple-choice format
solution_MC	The letter (e.g., A, B, C) of solution in multiple-choice format

Audio and video file names contain clip UIDs. For testing models with file names coded as text tokens and commercial models (e.g., Gemini), we recommend using temporary copies with anonymized file names.

Example

{
    "category": "sound effect",
    "name": "Goat Talking to Clueless Huh Cat",
    "url": "https://www.youtube.com/watch?v=B7aSDu6Qh6U",
    "onset": 0.0,
    "offset": 9.0,
    "original_date": 2024,
    "language": "N/A",
    "transcription": null,
    "summary": "A goat talking to a cat.",
    "emotion": [
        "surprised/shocked"
    ],
    "sensitivity": [
        "none"
    ],
    "usage": "Used to depict a situation where one person passionately and intensely explains a complex or niche topic to another person who is utterly confused, overwhelmed, or uncomprehending.",
    "question": "What kind of conversational dynamic are they used to represent?",
    "choices": [
        "One person chaotically spouting nonsense to a listener who responds with confusion.",
        "A deep, meaningful conversation where both parties are fully engaged.",
        "An angry argument between two equally aggressive individuals.",
        "A calm and patient mentor teaching a highly attentive student."
    ],
    "solution": "One person chaotically spouting nonsense to a listener who responds with confusion.",
    "question_type": "Contextual Inference",
    "video_path": "video/B7aSDu6Qh6U_0.0_9.0.mp4",
    "audio_path": "audio/B7aSDu6Qh6U_0.0_9.0.wav",
    "visual_hint": [
        "No text"
    ],
    "visual_cheat": false
}

Usage

from datasets import load_dataset
full_memes = load_dataset("naplab/AVMeme-Exam", "full", split="test") # for full 1032 memes
main_memes = load_dataset("naplab/AVMeme-Exam", "main", split="test") # for harder 846 memes

# Audio is loaded as a 16kHz waveform. Video (if not "visual_cheat") is returned as a path.
full_memes[0]

Citation

@misc{
  jiang2026avmemeexammultimodalmultilingual,
  title={AVMeme Exam: A Multimodal Multilingual Multicultural Benchmark for LLMs' Contextual and Cultural Knowledge and Thinking}, 
  author={Xilin Jiang and Qiaolin Wang and Junkai Wu and Xiaomin He and Zhongweiyang Xu and Yinghao Ma and Minshuo Piao and Kaiyi Yang and Xiuwen Zheng and Riki Shimizu and Yicong Chen and Arsalan Firoozi and Gavin Mischler and Sukru Samet Dindar and Richard Antonello and Linyang He and Tsun-An Hsieh and Xulin Fan and Yulun Wu and Yuesheng Ma and Chaitanya Amballa and Weixiong Chen and Jiarui Hai and Ruisi Li and Vishal Choudhari and Cong Han and Yinghao Aaron Li and Adeen Flinker and Mounya Elhilali and Emmanouil Benetos and Mark Hasegawa-Johnson and Romit Roy Choudhury and Nima Mesgarani},
  year={2026},
  eprint={2601.17645},
  archivePrefix={arXiv},
  primaryClass={cs.SD},
  url={https://arxiv.org/abs/2601.17645}, 
}

License

Annotations and metadata are released under CC BY 4.0.

The associated audio and video clips are collected from publicly available online sources and are provided solely for academic research.

Users are responsible for complying with the original content licenses.

Ethical Considerations

• Content includes sensitivity tags for mature themes
• Reflects online meme culture with inherent biases
• Requires cultural familiarity for appropriate judgment

---