The paper presents ESSC-RM, a general refinement framework for 3D Semantic Scene Completion (SSC) that can be seamlessly integrated into existing SSC models as a plug-and-play module. The framework operates in two phases: a baseline SSC network first produces a coarse voxel prediction, which is then refined by a 3D U-Net-based architecture incorporating two key components—the Prediction Noise-Aware Module (PNAM) and the Voxel-level Local Geometry Module (VLGM). PNAM enhances multi-scale voxel reasoning by combining global self-attention with localized neighborhood cross-attention, effectively capturing long-range dependencies and local geometric consistency. VLGM leverages frozen vision-language models (e.g., LLaVA, InstructBLIP) to generate free-form scene descriptions, which are encoded via JinaCLIP and Q-Former and fused into the voxel refinement pipeline through Semantic Interaction Guidance (SIGM) and Dual Cross-Attention (DCAM) modules. This injects high-level semantic priors that compensate for missing geometric cues in occluded or ambiguous regions. The refinement module is trained with a multi-scale loss combining class-weighted cross-entropy and Scene-Class Affinity Loss (SCAL), ensuring both voxel-wise accuracy and global semantic coherence. Experiments on the SemanticKITTI benchmark demonstrate consistent improvements: when integrated into CGFormer and MonoScene, the mean IoU increases from 16.87% to 17.27% and from 11.08% to 11.51%, respectively. The framework supports both joint training (co-optimizing refinement and backbone) and separate training (true plug-and-play), making it applicable to a wide range of SSC models without architectural modifications. The results validate ESSC-RM as a general, flexible, and effective refinement solution for 3D semantic scene completion.