keyword
Fast-DetectGPT
Fast-DetectGPT is a zero-shot algorithm designed to identify machine-generated text by analyzing statistical patterns in language model probabilities without requiring task-specific training data or fine-tuned classifiers. It operates on the principle of conditional probability curvature, exploiting the tendency of large language models to select tokens with consistently higher model-assigned probabilities compared to human authors. While earlier curvature-based methods like DetectGPT rely on computationally expensive passage perturbations from external models, Fast-DetectGPT evaluates probability curvature through direct conditional token sampling. This structural optimization significantly accelerates detection speed, lowers computational overhead, and maintains high classification accuracy across diverse models and writing domains.
1 item
