EmoMM: Benchmarking and Steering MLLM for Multimodal Emotion Recognition under Conflict and Missingness

Jul 2, 2026·
Yueru Sun
,
Yimeng Zhang
,
Haoyu Gu
,
Nuo Chen
,
Dong She
,
Xianrong Yao
,
Yang Gao
,
Zhanpeng Jin
EmoMM modality-conflict example and the CHASE attention-steering mechanism. Figure 1 from the paper.
Abstract
EmoMM benchmarks multimodal emotion recognition when audio, video, and text agree, conflict, or are missing. The study identifies Video Contribution Collapse, in which models underuse visual evidence. Conflict-aware Head-level Attention Steering (CHASE) adjusts attention at inference time to improve robustness without retraining the model backbone.
Type
Publication
Findings of the Association for Computational Linguistics: ACL 2026 (ACL 2026 Findings), pp. 20351–20371
ACL Findings Multimodal Emotion Recognition Multimodal LLM Benchmark