Spotting LLMs With Binoculars: Zero-Shot Detection of Machine-Generated Text
Detection without training data, by contrasting two models' views of the same text. A common independent baseline.
What it is
Binoculars scores text by the ratio of two perplexities from related models, an observer and a performer. The insight is that prompts make machine text look surprising to a single model even when it is not, and the ratio cancels that out. The paper reports strong detection with no training on any detector dataset and publishes code, which is why it is a standard baseline.
When to use this entry
Cite it for zero-shot detection and as a reproducible baseline.
Questions
Does Binoculars need training data?
No. It uses two pretrained language models and a ratio of their perplexities.