<p>Identifying causal direction between two variables remainsś challenging when data exhibit strong nonlinearity, discreteness, or mixed continuous–discrete structures. Existing approaches often rely on restrictive functional assumptions, are limited to specific data types, or degrade substantially under complex noise and non-additive interactions. Motivated by recent advances in representation learning for sufficient dimension reduction, we propose a unified and highly flexible framework for bivariate causal direction discovery based on the belted and ensembled neural network (BENN). The key idea is to learn a low-dimensional sufficient predictor of the response through a bottleneck architecture that captures essential information in the conditional distribution, thereby simplifying the subsequent regression or classification task used to construct residuals. These residuals approximate the noise component of the underlying additive or latent-utility model, enabling reliable independence testing in both directions. The proposed framework naturally accommodates continuous, discrete, and mixed-type variable pairs without requiring any modification to the overall pipeline. Extensive synthetic experiments across a wide variety of nonlinear mechanisms, noise distributions, and response types, along with evaluations on real-world datasets, demonstrate that the BENN-based approach achieves state-of-the-art accuracy and stability. It performs favorably compared to existing information-theoretic methods, ANM variants, and recent models for mixed-type data. Owing to its modular structure, we discuss the method’s future potential to be integrated as a preprocessing step in multivariate causal discovery algorithms, such as RESIT-type procedures.</p>