💡 This paper proposes DetectRL, a benchmark for LLM-generated text detection in real-world misuse scenarios, which simulates adversarial perturbations like human revisions and spelling errors to systema