MER2025 is the third year of our MER series of challenges, aiming to bring together researchers in the affective computing community to explore emerging trends and future directions in the field. Previously, MER23 Challenge@ACM MM focused on multi-label learning, noise robustness, and semi-supervised learning, while MER24 Challenge@IJCAI introduced a new track dedicated to open-vocabulary emotion recognition. This year, MER2025 centers on the theme "Affective Computing Meets Large Language Models". We aim to shift the paradigm from traditional categorical frameworks reliant on predefined emotion taxonomies to large language model (LLM)-driven generative methods, offering innovative solutions for more accurate and reliable emotion understanding. The challenge features four tracks: MER-SEMI focuses on fixed categorical emotion recognition enhanced by semi-supervised learning; MER-FG explores fine-grained emotions, expanding recognition from basic to nuanced emotional states; MER-DES incorporates multimodal cues (beyond emotion words) into predictions to enhance model interpretability; MER-PR investigates whether emotion prediction results can improve personality recognition performance. For the first three tracks, baseline code is available at MERTools and datasets can be accessed via Hugging Face. For the last track, the dataset and baseline code are available on GitHub.

Track 1. MER-SEMI. MER-SEMI spans three consecutive MER challenges, aiming to enhance the performance of categorical emotion recognition algorithms through semi-supervised learning and unlabeled data. This year, we expanded the dataset by incorporating more labeled and unlabeled samples. Participants are encouraged to leverage semi-supervised learning techniques, such as masked auto-encoders or contrastive learning, to achieve better results.

Track 2. MER-FG. Current frameworks primarily focus on basic emotions, often failing to capture the complexity and subtlety of human emotions. This track shifts the emphasis to fine-grained MER, enabling the prediction of a broader range of emotions. Following previous works [1, 2], participants are encouraged to leverage LLMs for this purpose. Given that LLMs possess extensive vocabularies, they hold the potential to generate more diverse emotion categories beyond basic labels.

Track 3. MER-DES. The first two tracks primarily focus on emotion words, neglecting the integration of multimodal clues during the inference process. This omission results in prediction outcomes that lack interpretability. Moreover, emotion words struggle to fully capture the dynamic, diverse, and sometimes ambiguous nature of human emotions. This track seeks to leverage free-form, natural language descriptions to represent emotions [2, 3], offering greater flexibility to achieve more accurate emotion representations and enhance model interpretability.

Track 4. MER-PR. Personality and emotion are deeply intertwined in human behavior and social interactions, yet current research often treats them as separate tasks, neglecting their inherent correlations. This track seeks to investigate the interplay between emotion and personality, exploring whether emotion recognition can enhance the accuracy of personality predictions. Participants are encouraged to employ techniques such as multi-task learning to analyze the influence of emotion on personality prediction.

Dataset. For the first three tracks, to download the dataset, participants must complete an EULA available on Hugging Face. The EULA clearly states that the dataset is for academic research purposes only and prohibits any modifications or uploads to the Internet. We will review the submitted EULA promptly; once approved, participants will gain access to the dataset. For the last track, the dataset and baseline code are available on GitHub.

Result submission. For MER-SEMI, MER-FG, and MER-DES, the test samples are selected from the 124k unlabeled samples. To reduce the task difficulty, we reduce the evaluation scope from 124k to 20k candidate samples . Participants must submit predictions for these 20k samples, which cover all the test samples of these tracks. Each track has distinct objectives. For MER-SEMI, participants must predict the most likely label from six predefined categories: worried, happy, neutral, angry, surprise, and sad ; For MER-FG, participants can freely predict any emotion labels without restrictions on category or quantity; For MER-DES, participants are required to submit both multimodal evidence and corresponding emotion labels to improve model interpretability. For MER-PR, we provide an official test set, and participants can directly submit predictions for the test set.

Paper submission. All participants are encouraged to submit a paper describing their solution to MRAC25 Workshop.

Baseline paper: https://arxiv.org/abs/2504.19423
Baseline code (Track1~Track3): https://github.com/zeroQiaoba/MERTools/tree/master/MER2025
Baseline code (Track4): https://github.com/cai-cong/MER25_personality
Contact email: merchallenge.contact@gmail.com; lianzheng2016@ia.ac.cn

[1] Zheng Lian, Haiyang Sun, Licai Sun, Haoyu Chen, Lan Chen, Hao Gu, Zhuofan Wen, Shun Chen, Siyuan Zhang, Hailiang Yao, Bin Liu, Rui Liu, Shan Liang, Ya Li, Jiangyan Yi, Jianhua Tao. OV-MER: Towards Open-Vocabulary Multimodal Emotion Recognition. arXiv preprint arXiv:2410.01495, 2024.
[2] Zheng Lian, Haoyu Chen, Lan Chen, Haiyang Sun, Licai Sun, Yong Ren, Zebang Cheng, Bin Liu, Rui Liu, Xiaojiang Peng, Jiangyan Yi, Jianhua Tao. AffectGPT: A New Dataset, Model, and Benchmark for Emotion Understanding with Multimodal Large Language Models. arXiv preprint arXiv:2501.16566, 2025.
[3] Zheng Lian, Haiyang Sun, Licai Sun, Hao Gu, Zhuofan Wen, Siyuan Zhang, Shun Chen, Mingyu Xu, Ke Xu, Kang Chen, Lan Chen, Shan Liang, Ya Li, Jiangyan Yi, Bin Liu, Jianhua Tao. Explainable Multimodal Emotion Recognition. arXiv preprint arXiv:2306.15401v5, 2024.