The field of multimodal research is moving towards a deeper understanding of human emotions and behaviors, with a focus on developing more robust and effective models for intent detection and emotion reasoning. Recent studies have highlighted the challenges posed by modality bias in multimodal datasets and the need for unbiased datasets to evaluate multimodal models effectively. Additionally, there is a growing interest in using generative multimodal large language models to synthesize normative data for cognitive assessments, which could potentially overcome the limitations of traditional data collection methods. Noteworthy papers in this area include: Text Takes Over, which proposes a framework to debias multimodal intent detection datasets and highlights the performance advantage of text-only models in certain datasets. CMR-SPB, which introduces a novel benchmark for cross-modal multi-hop reasoning and proposes a new prompting technique to mitigate performance gaps across different reasoning paths. Beyond Emotion Recognition, which introduces a multi-turn multimodal emotion understanding and reasoning benchmark and proposes a multi-agent framework to improve reasoning capabilities. Towards Synthesizing Normative Data for Cognitive Assessments, which demonstrates the feasibility of using generative multimodal large language models to synthesize robust synthetic normative data for existing cognitive tests.