We describe our submission to SemEval-2026 Task 9 Subtask 1, which focuses on multilingual polarization detection over the POLAR dataset. We compare three adaptation paradigms: fully fine-tuned multilingual encoders, frozen encoders augmented with lightweight residual heads, and inference-only multilingual LLM prompting in zero-shot and few-shot settings. For few-shot prompting, we evaluate both random and similarity-based support example selection. Similarity-based few-shot prompting with a multilingual LLM competes with our fine-tuned encoder baselines while requiring no task-specific training. We further analyze energy usage, stability across prompt selections and per-language behavior to characterize trade-offs between architectural adaptation and prompt-based inference. While our submission uses a fully fine tuned XLM-RoBERTa Large, the results indicate that inference-only prompting can be a competitive and energy-efficient alternative to task-specific fine-tuning in multilingual classification.