FIBER Detects LLM-Generated Text Using Bayesian Evidence Fusion

FIBER combines Fast-DetectGPT scores with Item Response Theory (IRT) to improve LLM-generated text detection across domain shifts. It uses Bayesian evidence fusion to weigh historical detector performance with current statistical signals. A variant called FIBER-Stream allows for estimation using unlabeled batches when historical scores are missing.