Core Insights - The article emphasizes that "emotional intelligence" is a crucial development direction for the next generation of artificial intelligence, marking a significant step towards general artificial intelligence. It highlights the need for digital humans and robots to accurately interpret multimodal interaction information and deeply explore human emotional states for more realistic and natural human-machine dialogue [1]. Group 1: Technological Advancements - The Kuaishou team and Nankai University have made groundbreaking research in the field of "multimodal emotion understanding," identifying key shortcomings in existing multimodal large models regarding emotional cue capture [1]. - A new modular duplex attention paradigm has been proposed, leading to the development of a multimodal model named 'MODA,' which significantly enhances capabilities in perception, cognition, and emotion across various tasks [1][7]. - The 'MODA' model has shown remarkable performance improvements in 21 benchmark tests across six major task categories, including general dialogue, knowledge Q&A, table processing, visual perception, cognitive analysis, and emotional understanding [1][28]. Group 2: Attention Mechanism Challenges - Existing multimodal large models exhibit a modal bias due to a language-centric pre-training mechanism, which hampers their ability to focus on fine-grained emotional cues, resulting in poor performance in advanced tasks requiring detailed cognitive and emotional understanding [4][7]. - The study reveals that attention scores in multimodal models tend to favor text modalities, leading to significant discrepancies in attention distribution across different layers, with cross-modal attention differences reaching up to 63% [4][8]. Group 3: Performance Metrics - The introduction of the modular duplex attention paradigm has effectively mitigated attention misalignment issues, reducing cross-modal attention differences from 56% and 62% to 50% and 41% respectively [25]. - The 'MODA' model, with parameter sizes of 8 billion and 34 billion, has achieved significant performance enhancements across various tasks, demonstrating its effectiveness in content perception, role cognition, and emotional understanding [25][28]. Group 4: Practical Applications - 'MODA' has shown strong potential in human-machine dialogue scenarios, capable of real-time analysis of user micro-expressions, tone, and cultural background, thereby constructing multidimensional character profiles and understanding emotional contexts [31]. - The model has been successfully applied in Kuaishou's data perception project, significantly enhancing data analysis capabilities, particularly in emotion recognition and reasoning tasks, thereby improving the accuracy of emotional change detection and personalized recommendations [33].