Can Large Language Models Help Multimodal Language Analysis? MMLA: A Comprehensive Benchmark
September 22, 2025 ยท View on GitHub
๐ค Hugging Faceย ย | ย ย ๐ Paperย ย | ย ย ๐ Datasetsย ย | ย ย ๐ฌ WeChat (ๅพฎไฟก)ย ย
Multimodal language analysis is a rapidly evolving field that leverages multiple modalities to enhance the understanding of high-level semantics underlying human conversational utterances. Despite its significance, little research has investigated the capability of multimodal large language models (MLLMs) to comprehend cognitive-level semantics. In this paper, we introduce MMLA, a comprehensive benchmark specifically designed to address this gap. MMLA comprises over 61K multimodal utterances drawn from both staged and real-world scenarios, covering six core dimensions of multimodal semantics: intent, emotion, dialogue act, sentiment, speaking style, and communication behavior. We evaluate eight mainstream branches of LLMs and MLLMs using three methods: zero-shot inference, supervised fine-tuning, and instruction tuning. Extensive experiments reveal that even fine-tuned models achieve only about 60%~70% accuracy, underscoring the limitations of current MLLMs in understanding complex human language. We believe that MMLA will serve as a solid foundation for exploring the potential of large language models in multimodal language analysis and provide valuable resources to advance this field.
Updates
- [2025.09.19]: ๐ ๐ ๐ Our work has been accepted to NeurIPS 2025! ๐ ๐ ๐
- [2025.05.16]: ๐ฅ ๐ฅ ๐ฅ The supervised fine-tuning (SFT) and instruction tuning (IT) code of the MMLA benchmark is released (Link), enjoy it!
- [2025.05.06]: ๐ฅ ๐ฅ ๐ฅ The zero-shot inference code of the MMLA benchmark is released (Link), enjoy it!
- [2025.04.29]: The datasets of the MMLA benchmark are released on Huggingface and Google Drive! The code will be released soon.
- [2025.04.24]: ๐ Our paper: Can Large Language Models Help Multimodal Language Analysis? MMLA: A Comprehensive Benchmark is released (arXiv, Huggingface, alphaXiv). The official repo is released on Github.
Overview of the MMLA Benchmark
Highlights
- Various Sources: 9 datasets, 61K+ samples, 3 modalities, 76.6 videos. Both acting and real-world scenarios (Films, TV series, YouTube, Vimeo, Bilibili, TED, Improvised scripts, etc.).
- 6 Core semantic Dimensions: Intent, Emotion, Sentiment, Dialogue Act, Speaking Style, and Communication Behavior.
- 3 Evaluation Methods: Zero-shot Inference, Supervised Fine-tuning, and Instruction Tuning.
- 8 Mainstream Foundation Models: 5 MLLMs (Qwen2-VL, VideoLLaMA2, LLaVA-Video, LLaVA-OV, MiniCPM-V-2.6), 3 LLMs (InternLM2.5, Qwen2, LLaMA3).
Supported Datasets
| Dimension | Dataset | Source | Venue |
|---|---|---|---|
| Intent | MIntRec | Paper / GitHub | ACM MM 2022 |
| Intent | MIntRec2.0 | Paper / GitHub | ICLR 2024 |
| Emotion | MELD | Paper / GitHub | ACL 2019 |
| Emotion | IEMOCAP | Paper / Website | Language Resources and Evaluation 2008 |
| Dialogue Act | MELD-DA | Paper / GitHub | ACL 2020 |
| Dialogue Act | IEMOCAP-DA | Paper / Website | ACL 2020 |
| Sentiment | MOSI | Paper / GitHub | IEEE Intelligent Systems 2016 |
| Sentiment | CH-SIMS v2.0 | Paper / GitHub | ICMI 2022 |
| Speaking Style | UR-FUNNY-v2 | Paper / GitHub | ACL 2019 |
| Speaking Style | MUStARD | Paper / GitHub | ACL 2019 |
| Communication Behavior | Anno-MI (client) | Paper / GitHub | ICASSP 2022 |
| Communication Behavior | Anno-MI (therapist) | Paper / GitHub | ICASSP 2022 |
Release
The raw text and videos of each dataset all released on Huggingface and Google Drive.
Note that for MOSI, IEMOCAP, and IEMOCAP-DA datasets, we only provide the raw texts due to their restricted license. The raw videos of IEMOCAP can be downloaded from here. The MOSI dataset cannot be released due to the privacy limitation as mentioned in CMU-MultimodalSDK.
Supported Models
| Models | Model scale and Link | Source | Type |
|---|---|---|---|
| Qwen2 | ๐ค 0.5B / 1.5B / 7B | Paper / GitHub | LLM |
| Llama3 | ๐ค 8B | Paper / GitHub | LLM |
| InternLM2.5 | ๐ค 7B | Paper / GitHub | LLM |
| VideoLLaMA2 | ๐ค 7B | Paper / GitHub | MLLM |
| Qwen2-VL | ๐ค 7B / 72B | Paper / GitHub | MLLM |
| LLaVA-Video | ๐ค 7B / 72B | Paper / GitHub | MLLM |
| LLaVA-OneVision | ๐ค 7B / 72B | Paper / GitHub | MLLM |
| MiniCPM-V-2.6 | ๐ค 8B | Paper / GitHub | MLLM |
Evaluation Results
LeaderBoard
Rank of Zero-shot Inference
| RANK | Models | ACC | TYPE |
|---|---|---|---|
| ๐ฅ | GPT-4o | 52.60 | MLLM |
| ๐ฅ | Qwen2-VL-72B | 52.55 | MLLM |
| ๐ฅ | 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 |
Rank of Supervised Fine-tuning (SFT) and Instruction Tuning (IT)
| Rank | Models | ACC | Type |
|---|---|---|---|
| ๐ฅ | Qwen2-VL-72B (SFT) | 69.18 | MLLM |
| ๐ฅ | MiniCPM-V-2.6-8B (SFT) | 68.88 | MLLM |
| ๐ฅ | 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 |
Fine-grained Performance on Each Dimension
We show the results of three evaluation methods (i.e., zero-shot inference, SFT, and IT). The performance of state-of-the-art multimodal machine learning methods and GPT-4o is also shown in the figure below.
Zero-shot Inference and Supervised Fine-tuning (SFT)
Instruction Tuning (IT)
Acknowledgements
If our work is helpful to your research, please consider giving us a star ๐ and 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},
}