LLM Hallucination Detection on HaluEval (random sample of 1,000 text pairs)
95.3RecallLAG-XAI (cheap geometric check)
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| LAG-XAI (cheap geometric check)Threshold=1.108, Evaluation protocol=cross-corpus transfer, unsupervised, no fine-tuning, Calibration source=90th percentile of the PIT-2015 validation set, Sample composition=1,000 hallucinations vs. 1,000 legitimate semantic transitions2026.04 | 95.3 | 10 | 90.5 | 92.8 |