Datasets:
license: cc-by-4.0
task_categories:
- zero-shot-classification
- text-classification
- text-generation
language:
- en
- zh
size_categories:
- 10K<n<100K
pretty_name: MMLA
Can Large Language Models Help Multimodal Language Analysis? MMLA: A Comprehensive Benchmark
1. Introduction
MMLA is the first comprehensive multimodal language analysis benchmark for evaluating foundation models. It has the following features:
- Large Scale: 61K+ multimodal samples.
- Various Sources: 9 datasets.
- Three Modalities: text, video, and audio
- Both Acting and Real-world Scenarios: films, TV series, YouTube, Vimeo, Bilibili, TED, improvised scripts, etc.
- Six Core Dimensions in Multimodal Language Analysis: intent, emotion, sentiment, dialogue act, speaking style, and communication behavior.
We also build baselines with three evaluation methods (i.e., zero-shot inference, supervised fine-tuning, and instruction tuning) on 8 mainstream foundation models (i.e., 5 MLLMs (Qwen2-VL, VideoLLaMA2, LLaVA-Video, LLaVA-OV, MiniCPM-V-2.6), 3 LLMs (InternLM2.5, Qwen2, LLaMA3). More details can refer to our paper.
2. Datasets
2.1 Statistics
Dataset statistics for each dimension in the MMLA benchmark. #C, #U, #Train, #Val, and #Test represent the number of label classes, utterances, training, validation, and testing samples, respectively. avg. and max. refer to the average and maximum lengths.
| Dimensions | Datasets | #C | #U | #Train | #Val | #Test | Video Hours | Source | #Video Length (avg. / max.) | #Text Length (avg. / max.) | Language |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Intent | MIntRec | 20 | 2,224 | 1,334 | 445 | 445 | 1.5 | TV series | 2.4 / 9.6 | 7.6 / 27.0 | English |
| MIntRec2.0 | 30 | 9,304 | 6,165 | 1,106 | 2,033 | 7.5 | TV series | 2.9 / 19.9 | 8.5 / 46.0 | ||
| Dialogue Act | MELD | 12 | 9,989 | 6,992 | 999 | 1,998 | 8.8 | TV series | 3.2 / 41.1 | 8.6 / 72.0 | English |
| IEMOCAP | 12 | 9,416 | 6,590 | 942 | 1,884 | 11.7 | Improvised scripts | 4.5 / 34.2 | 12.4 / 106.0 | ||
| Emotion | MELD | 7 | 13,708 | 9,989 | 1,109 | 2,610 | 12.2 | TV series | 3.2 / 305.0 | 8.7 / 72.0 | English |
| IEMOCAP | 6 | 7,532 | 5,237 | 521 | 1,622 | 9.6 | Improvised scripts | 4.6 / 34.2 | 12.8 / 106.0 | ||
| Sentiment | MOSI | 2 | 2,199 | 1,284 | 229 | 686 | 2.6 | Youtube | 4.3 / 52.5 | 12.5 / 114.0 | English |
| CH-SIMS v2.0 | 3 | 4,403 | 2,722 | 647 | 1,034 | 4.3 | TV series, films | 3.6 / 42.7 | 1.8 / 7.0 | Mandarin | |
| Speaking Style | UR-FUNNY-v2 | 2 | 9,586 | 7,612 | 980 | 994 | 12.9 | TED | 4.8 / 325.7 | 16.3 / 126.0 | English |
| MUStARD | 2 | 690 | 414 | 138 | 138 | 1.0 | TV series | 5.2 / 20.0 | 13.1 / 68.0 | ||
| Communication Behavior | Anno-MI (client) | 3 | 4,713 | 3,123 | 461 | 1,128 | 10.8 | YouTube & Vimeo | 8.2 / 600.0 | 16.3 / 266.0 | English |
| Anno-MI (therapist) | 4 | 4,773 | 3,161 | 472 | 1,139 | 12.1 | 9.1 / 1316.1 | 17.9 / 205.0 |
2.2 License
This benchmark uses nine datasets, each of which is employed strictly in accordance with its official license and exclusively for academic research purposes. We fully respect the datasets’ copyright policies, license requirements, and ethical standards. For those datasets whose licenses explicitly permit redistribution, we release the original video data (e.g., MIntRec, MIntRec2.0, MELD, UR-FUNNY-v2, MUStARD, MELD-DA, CH-SIMS v2.0, and Anno-MI. For datasets that restrict video redistribution, users should obtain the videos directly from their official repositories (e.g., MOSI, IEMOCAP and IEMOCAP-DA. In compliance with all relevant licenses, we also provide the original textual data unchanged, together with the specific dataset splits used in our experiments. This approach ensures reproducibility and academic transparency while strictly adhering to copyright obligations and protecting the privacy of individuals featured in the videos.
3. LeaderBoard
3.1 Rank of Zero-shot Inference
| RANK | Models | ACC | TYPE |
|---|---|---|---|
| 🥇 | GPT-4o | 52.60 | MLLM |
| 🥈 | Qwen2-VL-72B | 52.55 | MLLM |
| 🥉 | LLaVA-OV-72B | 52.44 | MLLM |
| 4 | LLaVA-Video-72B | 51.64 | MLLM |
| 5 | InternLM2.5-7B | 50.28 | LLM |
| 6 | Qwen2-7B | 48.45 | LLM |
| 7 | Qwen2-VL-7B | 47.12 | MLLM |
| 8 | Llama3-8B | 44.06 | LLM |
| 9 | LLaVA-Video-7B | 43.32 | MLLM |
| 10 | VideoLLaMA2-7B | 42.82 | MLLM |
| 11 | LLaVA-OV-7B | 40.65 | MLLM |
| 12 | Qwen2-1.5B | 40.61 | LLM |
| 13 | MiniCPM-V-2.6-8B | 37.03 | MLLM |
| 14 | Qwen2-0.5B | 22.14 | LLM |
3.2 Rank of Supervised Fine-tuning (SFT) and Instruction Tuning (IT)
| Rank | Models | ACC | Type |
|---|---|---|---|
| 🥇 | Qwen2-VL-72B (SFT) | 69.18 | MLLM |
| 🥈 | MiniCPM-V-2.6-8B (SFT) | 68.88 | MLLM |
| 🥉 | LLaVA-Video-72B (IT) | 68.87 | MLLM |
| 4 | LLaVA-ov-72B (SFT) | 68.67 | MLLM |
| 5 | Qwen2-VL-72B (IT) | 68.64 | MLLM |
| 6 | LLaVA-Video-72B (SFT) | 68.44 | MLLM |
| 7 | VideoLLaMA2-7B (SFT) | 68.30 | MLLM |
| 8 | Qwen2-VL-7B (SFT) | 67.60 | MLLM |
| 9 | LLaVA-ov-7B (SFT) | 67.54 | MLLM |
| 10 | LLaVA-Video-7B (SFT) | 67.47 | MLLM |
| 11 | Qwen2-VL-7B (IT) | 67.34 | MLLM |
| 12 | MiniCPM-V-2.6-8B (IT) | 67.25 | MLLM |
| 13 | Llama-3-8B (SFT) | 66.18 | LLM |
| 14 | Qwen2-7B (SFT) | 66.15 | LLM |
| 15 | Internlm-2.5-7B (SFT) | 65.72 | LLM |
| 16 | Qwen-2-7B (IT) | 64.58 | LLM |
| 17 | Internlm-2.5-7B (IT) | 64.41 | LLM |
| 18 | Llama-3-8B (IT) | 64.16 | LLM |
| 19 | Qwen2-1.5B (SFT) | 64.00 | LLM |
| 20 | Qwen2-0.5B (SFT) | 62.80 | LLM |
4. Acknowledgements
For more details, please refer to our Github repo. If our work is helpful to your research, please consider citing the following paper:
@article{zhang2025mmla,
author={Zhang, Hanlei and Li, Zhuohang and Zhu, Yeshuang and Xu, Hua and Wang, Peiwu and Zhu, Haige and Zhou, Jie and Zhang, Jinchao},
title={Can Large Language Models Help Multimodal Language Analysis? MMLA: A Comprehensive Benchmark},
year={2025},
journal={arXiv preprint arXiv:2504.16427},
}