Unsupervised multi-omics integration methods are increasingly used to uncover latent structure in high-dimensional biological data, yet their behaviour and interpretability can vary substantially depending on the underlying inference strategy. In this study, two unsupervised multi-omics factorisation frameworks, Multi-Omics Factor Analysis (MOFA2) and Group Factor Analysis (GFA), were systematically benchmarked using paired RNA sequencing and promoter-level DNA methylation data from glioblastoma tumours as a representative use case. The methods were compared with respect to latent factor structure, partitioning of shared versus modality-specific variation, variance distribution across factors, and biological interpretability based on functional enrichment analyses. Although both frameworks captured complementary transcriptomic and epigenetic signals, they differed markedly in how variation was organised within the latent space. MOFA2 produced a compact and strongly regularised representation with a small number of dominant factors, whereas GFA retained a more distributed latent structure that preserved weaker sources of variation. Overall, this study highlights fundamental methodological trade-offs between interpretability and completeness in unsupervised multi-omics integration. The results emphasise that method selection should be guided by analytical objectives rather than biological context alone, and demonstrate the value of comparative benchmarking for informed application of unsupervised integration frameworks.