<p>Identifying causal relationships from observational data in complex environments remains a formidable challenge, particularly in systems where interactions are influenced by latent confounders. Conventional causal discovery approaches for nonlinear data often rely on the causal sufficiency assumption, limiting their applicability in real-world scenarios. In this work, we propose VC-flow, a novel generative structural causal model that integrates machine learning techniques to address causal direction identification in the presence of latent variables. By relaxing both linearity and additive noise assumptions, our approach provides a more general framework for understanding system-environment interactions. Leveraging a VFlow-based variational inference module, we explicitly identify latent variables and recover causal structures, enabling improved communication and interaction between systems. Experimental results on both simulated and real-world datasets demonstrate the superior accuracy of our method.</p>