C3-UniMM Enhances Multimodal AI with Causal Cycle Consistency.

This paper introduces C3-UniMM (Causal Cycle-Consistent Unified Multimodal Modeling), a novel framework designed to address fundamental limitations in existing unified multimodal models. Current models often struggle with issues like semantic drift, poor compositional generalization, and instability because they primarily rely on implicit statistical correlations rather than explicit structural consistency across modalities. C3-UniMM tackles these problems by integrating Causal Cycle Consistency and Super Alignment. The core of C3-UniMM involves a Structured Latent Causal Graph (SLCG) that serves as a shared semantic space across different modalities. Unified multimodal encoding blocks are designed to optimize understanding and generation synergistically within this causal structure. Additionally, a Unified Decoding Space is implemented to ensure structural preservation and semantic invertibility during the cross-modal generation process. Theoretical analysis and extensive experiments demonstrate that C3-UniMM significantly improves the invertibility and mechanism invariance of cross-modal mappings, outperforming existing baselines across various understanding, generation, and compositional generalization tasks.