We are pleased to announce that COSMOS’s research has been published at the 39th Annual Conference on Neural Information Processing Systems (NeurIPS 2025) considered the most prestigious AI conference worldwide. Acceptance to NeurIPS represents one of the highest levels of recognition in the AI research community, and this achievement highlights the global impact and excellence of the work emerging from our center.

The paper, authored by Prof. Agarwal and collaborators from University of Arkansas Fayetteville and University of Florida Gainesville,  introduces MANGO: a Multimodal Attention-based Normalizing Flow approach that redefines how modern AI systems learn from and fuse multiple data sources. Multimodal learning integrating information such as images, depth maps, and text is essential for building AI systems that perceive the world in a more human-like and comprehensive way. However, existing fusion techniques often rely on implicit attention mechanisms, which can struggle to capture deeper structural relationships or provide interpretable insights into how different modalities interact.

MANGO addresses these gaps through a novel explicit modeling framework built on Invertible Cross-Attention (ICA), enabling interpretable, mathematically tractable, and highly expressive multimodal fusion. This new architecture introduces three advanced cross-attention strategies Modality-to-Modality, Inter-Modality, and Learnable Inter-Modality Cross-Attention that allow the model to capture intricate correlations and complementary information across modalities with greater precision than prior methods.

To further enhance scalability and efficiency, the framework integrates a latent normalizing flow design, significantly reducing computational overhead while improving semantic understanding of multimodal inputs. Through extensive experimentation across diverse benchmarks including semantic segmentation, image-to-image translation, and multimodal movie genre classification MANGO consistently outperforms existing state-of-the-art approaches.

This publication at NeurIPS 2025 reinforces COSMOS’s commitment to pioneering research in transparent, reliable, and high-impact AI systems. It showcases our leadership in advancing next-generation multimodal information environment analysis and demonstrates how innovations developed at COSMOS continue to shape the global AI landscape.