In real-world scenarios, referring video object segmentation (RVOS) remains constrained by the scarcity of annotated data, which impedes effective model training and severely limits generalization capabilities to unseen environments. To address these persistent challenges, we propose an innovative framework that integrates meta-learning with graph neural networks (GNNs) through seamless incorporation of Laplacian regularization for robust cross-set relational reasoning, coupled with a Bayesian neural networks(BNNs) segmentation head for theoretically grounded uncertainty quantification. The graph Laplacian term in our framework effectively imposes smoothness priors on feature maps, thereby preserving the inherent manifold structure of vision-language joint representations in RVOS tasks. This regularization significantly reduces noise sensitivity while mitigating overfitting in few-shot scenarios. Meanwhile, the pixel-level uncertainty estimates generated by the Bayesian segmentation head adaptively refine object boundaries and enhance reliability in semantically ambiguous regions. Although existing studies have explored few-shot learning in RVOS, our approach pioneers the synergistic integration of GNNs-based cross-set relational reasoning with Bayesian uncertainty modeling. Experiments on the Mini-Ref-YouTube-VOS and Mini-Ref-SAIL-VOS few-shot benchmarks demonstrate that our method consistently outperforms baseline models, achieving superior segmentation accuracy and robustness. By addressing data scarcity while enhancing interpretability through uncertainty visualization and structured reasoning, this framework significantly advances the practical application value of RVOS technology in real-world video understanding scenarios.