VietBinoculars: A Zero-Shot Approach for Detecting Vietnamese LLM-Generated Text
VietBinoculars is a zero-shot method to detect Vietnamese LLM-generated text by adapting the Binoculars technique with tuned thresholds. It uses two related LLMs to calculate a surprise score based on perplexity ratios. This approach achieved over 99% accuracy across multiple Vietnamese datasets, significantly outperforming existing tools for AI content detection.