Unified multimodal models aim for any‑to‑any understanding and generation across arbitrary modalities. Existing approaches largely rely on implicit statistical correlations and lack cross‑modal structural consistency, which causes semantic drift, poor compositional generalization, and instability under interventions.\
We introduce C3‑UniMM, a framework built on Causal Cycle Consistency and Super Alignment. A Structured Latent Causal Graph (SLCG) serves as a shared cross‑modal semantic space, and unified multimodal encoding blocks enable understanding and generation to be jointly optimized within the same causal semantic structure.\
For generation we propose a Unified Decoding Space that enforces structural preservation and semantic invertibility. Formally, let $f: X \rightarrow Z$ be the encoder and $g: Z \rightarrow Y$ the decoder; cycle consistency requires $g(f(x)) \approx x$, and both $f$ and $g$ must preserve the causal mechanism on the SLCG. Theoretical analysis shows that these constraints markedly improve invertibility and mechanism invariance of cross‑modal mappings.\
Extensive experiments on understanding, generation, and compositional generalization tasks demonstrate that C3‑UniMM consistently outperforms current unified multimodal baselines in accuracy, BLEU, and especially in robustness to compositional reasoning scenarios.\
Review: By embedding explicit causal structure and a shared decoding space, C3‑UniMM addresses the structural shortcomings of prior multimodal models and offers a practical route toward genuine cross‑modal causal inference.