MER2026 Challenge & MRAC2026 Workshop

About

What is MER2026?

MER2026 marks the fourth edition of the MER series of challenges. The MER series provides valuable data resources to the research community and offers tasks centered on recent research trends, establishing itself as one of the largest challenges in the field. Throughout its history, the focus of MER has shifted from discriminative emotion recognition to generative emotion understanding. Specifically, MER2023 concentrated on discriminative emotion recognition, restricting the scope to fixed basic labels. In MER2024 and MER2025, we transitioned to generative emotion understanding, leveraging the extensive vocabulary and multimodal understanding capabilities of MLLMs to facilitate fine-grained and explainable emotion recognition. Building on this trajectory, MER2026 contains four tracks: MER-Cross, MER-FG, MER-Prefer, and MER-PS.

Contact: merchallenge.contact@gmail.com · lianzheng@tongji.edu.cn

Updates

News

Apr 30, 2026 We establish an initial website for MER2026 Challenge and MRAC2026 Workshop.

Jun 22, 2026 CodaBench links for Track 1–4 are now open for result submission.

Jul 03, 2026 To better assess model robustness, the final ranking for Track 3 will now be determined by the average score of Stage 1 (the current test set) and Stage 2 (a supplementary test set). Stage 2 will be released on July 11, AoE, and each team is allowed strictly FIVE submission via a dedicated submission channel to be provided shortly. The final submission deadline remains unchanged at July 14, AoE.

Jul 16, 2026 All tracks have now concluded. We have released test set labels on HuggingFace. The next phase is paper writing, with a deadline of July 22, 23:59 AoE. To facilitate your paper writing, we provide additional CodaBench links:
• Track 1: codabench.org/competitions/17536
• Track 2: codabench.org/competitions/17537/
• Track 3 (Stage1): codabench.org/competitions/17539
• Track 3 (Stage2): codabench.org/competitions/17538

Aug 4, 2026 Workshop paper decisions have been announced, and notification emails have been sent to the authors.

MER26 Challenge @ ACM MM

Four Challenge Tracks

🗣️

Track 1

MER-Cross: Interlocutor Emotion

A newly introduced track that shifts focus from individual scenarios to dyadic interactions. When speaker s₁ is talking, MER-Cross targets the emotion of the listening s₂ (interlocutor) rather than the speaker, enabling capture of both sides of a conversation.

MER-Cross illustration — MER-Cross: predicting the interlocutor's emotion in dyadic interactions.

🔍

Track 2

MER-FG: Fine-grained Emotion

Human emotion extends far beyond basic labels. In this track, participants can predict any number of emotion labels across diverse categories, expanding recognition scope from basic to more nuanced emotions.

MER-FG illustration — MER-FG: open-vocabulary emotion recognition beyond basic categories.

⚖️

Track 3

MER-Prefer: Emotion Preference

A newly introduced track predicting which of two emotion descriptions is preferred by human annotators for a given video — a critical component for training reward models in emotion understanding.

MER-Prefer illustration — MER-Prefer: determine which emotion description humans prefer.

🧠

Track 4

MER-PS: Physiological Signal Emotion

Shifts emotion recognition from observable behaviors to physiological evidence. Using synchronized EEG (64-ch, 1000 Hz) and fNIRS (51-ch, 47.62 Hz) signals from 30 participants watching 15 emotion-eliciting video clips.

A key feature is real-time dynamic emotion annotation — subjects continuously report valence and arousal via joystick at 1 Hz. The goal is to estimate the valence–arousal trajectory from brain signals.

Resources & Submission

📄

Baseline Paper

arxiv.org/abs/2604.19417

💻

Tracks 1-3 Baseline Code

https://github.com/zeroQiaoba/MERTools

📦

Tracks 1-3 Register&Data

https://huggingface.co/datasets/MERChallenge/MER2026

🏆

Track 1 CodaBench

codabench.org/competitions/17195

🏆

Track 2 CodaBench

codabench.org/competitions/17196

🏆

Track 3 CodaBench (stage1)

codabench.org/competitions/17197

🏆

Track 3 CodaBench (stage2)

codabench.org/competitions/17459

💻

Track 4 Baseline Code&CodaBench

https://www.codabench.org/competitions/15992/

📦

Track 4 Register&Data

https://huggingface.co/datasets/MER-PS/MER-PS-trainval

[1] Zheng Lian et al. MER 2023: Multi-label Learning, Modality Robustness, and Semi-supervised Learning. Proceedings of the 31st ACM International Conference on Multimedia, 2023.

[2] Zheng Lian et al. MER 2024: Semi-supervised Learning, Noise Robustness, and Open-vocabulary Multimodal Emotion Recognition. Proceedings of the 2nd International Workshop on Multimodal and Responsible Affective Computing, 2024.

[3] Zheng Lian et al. MER 2025: When Affective Computing Meets Large Language Models. Proceedings of the 33rd ACM International Conference on Multimedia, 2025.

[4] Zheng Lian et al. Explainable Multimodal Emotion Reasoning. arXiv:2306.15401, 2023.

[5] Zheng Lian et al. OV-MER: Towards Open-Vocabulary Multimodal Emotion Recognition. Forty-second International Conference on Machine Learning (ICML), 2025.

[6] Zheng Lian et al. AffectGPT: A New Dataset, Model, and Benchmark for Emotion Understanding with Multimodal Large Language Models. Forty-second International Conference on Machine Learning (ICML), 2025.

[7] Zheng Lian et al. EmoPrefer: Can Large Language Models Understand Human Emotion Preferences? International Conference on Learning Representations (ICLR), 2026.

[8] Zheng Lian et al. AffectGPT-R1: Leveraging Reinforcement Learning for Open-Vocabulary Multimodal Emotion Recognition. arXiv:2508.01318, 2026.

MRAC2026 Workshop @ ACM MM

Workshop Topics & Submission

🎯 Research Topics

Large-scale data generation / inexpensive annotation for Affective Computing
Generative AI for Affective Computing using multimodal signals
Multi-modal methods for emotion recognition
Privacy-preserving large-scale emotion recognition in the wild
Affective Computing in education, entertainment & healthcare
Explainable or Privacy-Preserving AI in affective computing
Semi-/weak-/un-/self-supervised learning for Affective Computing
Physiological signal-based emotion recognition (EEG, fNIRS, etc.)
Dyadic and group emotion recognition in interaction scenarios
Reward modeling and preference learning for emotion understanding

📋 Submission Guidelines

Format: ACM double-column template (paper template). Double-blind review — anonymize all author information.

Length:

Option A: 4 pages + 1-page references
Option B: 8 pages + up to 2-page references

Review: All papers peer-reviewed; accepted papers published in the ACM Digital Library.

Submit Your Paper to MRAC2026

Challenge solutions and any work on multimodal affective computing are welcome.

Submit Paper

Fairness & Integrity

Method Validity and Reproducibility

Participating teams may apply reasonable preprocessing methods, priors constructed from training samples, or other approaches that help the model learn generalizable structures, provided that all components belong to a complete, reproducible, and auditable modeling pipeline.

Final submissions must be produced by a learnable, reproducible model pipeline, rather than relying on competition-specific non-generalizable shortcuts. In particular, submissions should not primarily generate predictions using sample identifiers, file names, test-sample ordering, fixed lookup tables, hand-crafted scaling, constant biases, or rules iteratively calibrated against the leaderboard.

If any prior, template, or calibration term derived from training-data statistics is used, the participating team must be able to explain its role in the training pipeline and demonstrate its contribution through necessary ablation experiments.

All submissions must strictly follow the submission template, entry function, file organization, and output format specified in the official starting_kit. The final evaluation environment will follow the official submission interface and may adopt stricter input encapsulation and template constraints. The sample_id serves only as a sample index identifier and should not be treated as a stable semantic information interface.

🔍 Integrity & Audit

The organizers will conduct random checks on submitted results during and after the competition. For submissions with anomalous performance, unusual historical variation, or prediction behavior exhibiting excessive reliance on fixed priors, template lookups, or hand-crafted calibration terms, the organizers may contact the team for additional reproducibility evidence.

📦 Required Materials

To verify method validity and reproducibility, the organizers require teams to provide trained model weights, source code, model files, training scripts, configuration files, random seeds, preprocessing instructions, inference instructions, ablation study results, and any necessary materials to reproduce the submitted results.

⚠️ Note: The organizers reserve the right to conduct consistency checks on submission files, historical submission records, reproduction materials, model input dependencies, and prediction behavior. If a submission cannot demonstrate that its results come from a reproducible modeling pipeline, or is found to primarily rely on non-generalizable competition-specific shortcuts to generate predictions, the submission may be deemed invalid and may not be included in the final ranking.

Acknowledgement

We would like to express our sincere gratitude to the Dataset Chairs for their support during the dataset construction process: Xiaojiang Peng (Shenzhen Technology University), Kele Xu (National University of Defense Technology), Fei Ma (Guangdong Lab of Artificial Intelligence and Digital Economy (SZ)), Ziyu Jia (Institute of Automation, CAS), Laizhong Cui (Shenzhen University), and Zebang Cheng (Shenzhen University). We also thank the members of our Annotation Team for their efforts: Zelin Yan (Tongji University), Liyi Liu (Tongji University), Chenxi Zhou (National University of Defense Technology), Yuan Cao (National University of Defense Technology), Kaiyao Li (Shenzhen University), Dawei Huang (Shenzhen University), and Hanwen Du (Shenzhen University). Finally, we thank the ACM Multimedia 2026 Challenge Chairs for giving us the opportunity to organize MER2026 at this premier conference.

MER2026 Challenge
& MRAC2026 Workshop