1 paper · 1 filter
Igor Itkin
Token-level hallucination detectors score each token independently from a single signal, and fail exactly when the generating model is confidently wrong. This paper instead treats…