One email each morning — yesterday's AI, sortedGet it in your inbox

METAL LAB

Temporal Multi-Signal Fusion for Token-Level Hallucination Detection

arXiv:2608.181152026-08-20

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 hallucination as a temporally extended span and detects it by sequence labeling: each token is scored from a 33-dimensional feature stream that fuses text statistics, Natural Language Inference (NLI) entailment, and language model surprisal, with no access to model intern

Authors · Igor Itkin

Read on arXiv

Latest papers

All papers →

Latest from METAL LAB