GURI boosts LLM hallucination detection to 79.09% ROC-AUC

Researchers Shaked Goldshtein, Yitzchak Grunbaum and coauthors introduced GURI, a hallucination detector that uses candidate token identities alongside their probabilities. Across three open-weight models and three QA benchmarks, mean test ROC-AUC rose from 72.09% for LOS-Net to 79.09%.

• GURI raised mean ROC-AUC from 72.09% to 79.09% across three models and benchmarks
• The method transferred in 17 of 18 directed cross-model transfers without target fine-tuning
• Fixed random identity codes matched or beat pretrained embeddings in experiments
• It requires candidate token strings and probabilities, which some commercial APIs do not expose