This paper presents our approach to Binary and Multiclass Hope Speech Detection within the IberLEF 2025 shared task. The objective of this study is to assess the effectiveness of various large language models (LLMs)-including GPT (ChatGPT o3-mini), Claude (3.5 Sonnet), Llama (4), and DeepSeek(R1)-by comparing their performance against XLM-RoBERTa, which we use as a baseline multilingual transformer model. To enhance the performance of these models, we leverage advanced techniques such as one-shot and few-shot learning prompting strategies, as well as data preprocessing and augmentation methods. Our results provide empirical insights into the comparative strengths and limitations of these models in detecting hope speech and distinguishing it from sarcasm across multiple languages. In Task 1, our best model achieved F1 scores of 0.8584 for English and 0.8435 for Spanish, while in Task 2, it attained 0.7563 for English and 0.7540 for Spanish.
