Language-Centric Framework for Multimodal AI Intelligence Proposed

This paper introduces a novel language-centric framework designed to unify multimodal data understanding in AI. The core idea is to convert all forms of observation—whether from images, videos, or text—into a collection of "atomic propositions." These are simple, factual statements describing entities, actions, and relationships within a scene. A global semantic codebook then standardizes these propositions into a shared, interpretable vocabulary. This creates a single conceptual space that spans fine-grained details to high-level concepts, allowing for richer composition and understanding across different modalities. The framework promises improved interpretability, enabling AI systems to reason more effectively and perform cross-modal retrieval. It also supports compositional understanding, facilitating complex multimodal tasks and structured data curation. The authors demonstrate the framework's capabilities using examples from autonomous driving and open-world data.