This paper comprehensively evaluates the performance of large-scale language models (LLMs), which are attracting attention as a countermeasure against the proliferation of fake news, extreme bias, and harmful content on online platforms, across various models, usage methods, and languages. Using ten datasets and five languages (English, Spanish, Portuguese, Arabic, and Bulgarian), we experimentally compared and analyzed LLM adaptation paradigms in binary and multi-class scenarios. We tested various in-context learning strategies, including parameter-efficient fine-tuning, zero-shot prompts, codebooks, few-shot learning (including random selection and DPP-based multiple-choice examples), and chain-of-thought.
