INSCONE: Unknown-Aware Detection of LLM-Generated Text via Informed Wild Data

Abstract

AI-generated text is increasingly difficult to distinguish from human writing, creating risks in academic integrity, medical misinformation, and social media disinformation. While recent work has reframed MGT detection as an out-of-distribution problem with strong in-distribution and zero-shot results, generalization to LLMs unseen during training remains underexplored. We propose INSCONE (Informed SCONE), which adapts the SCONE wild-data framework to the text domain by exploiting curated wild data with known mixing proportions (πid, πc, πs) to stabilize the energy geometry around seen and unseen LLM families. INSCONE achieves a 6.1-point FPR95 improvement over a competitive baseline on the RAID benchmark. We additionally release RAID+, an extended evaluation set regenerating RAID prompts with contemporary frontier models.