Hallucination Detection on MetaQA 1hop (Qwen2.5-7B)
85.28AUCGGA (PRD + SAS)
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| GGA (PRD + SAS)Category=Our Method, LLM=Qwen2.5-7B2025.12 | 85.28 | 46.06 | 70.83 | 39.74 | 54.77 | |
| SSR + SASCategory=Mechanistic Detector2026.05 | 85.28 | 46.06 | 70.83 | 39.74 | 54.77 | |
| NLI ContradictionCategory=Semantic Consistency, LLM=Qwen2.5-7B2025.12 | 50.44 | 11.5 | 26.25 | 6.23 | 74.19 | |
| NLI ContradictionCategory=Semantic Consistency2026.05 | 50.44 | 11.5 | 26.25 | 6.23 | 74.19 | |
| Token confidenceCategory=Model-Internal Confidence, LLM=Qwen2.5-7B2025.12 | 49.66 | 10.84 | 36.77 | 6.06 | 51.61 | |
| Token confidenceCategory=Model-Internal Confidence2026.05 | 49.66 | 10.84 | 36.77 | 6.06 | 51.61 | |
| PerplexityCategory=Model-Internal Confidence, LLM=Qwen2.5-7B2025.12 | 49.35 | 10.86 | 35.12 | 6.02 | 54.84 | |
| PerplexityCategory=Model-Internal Confidence2026.05 | 49.35 | 10.86 | 35.12 | 6.02 | 54.84 | |
| Embedding DivergenceCategory=Semantic Consistency, LLM=Qwen2.5-7B2025.12 | 49 | 11.34 | 5.88 | 6.02 | 96.77 | |
| Embedding DivergenceCategory=Semantic Consistency2026.05 | 49 | 11.34 | 5.88 | 6.02 | 96.77 | |
| BERTScoreCategory=Semantic Consistency, LLM=Qwen2.5-7B2025.12 | 48.67 | 11.15 | 7.45 | 5.93 | 93.55 | |
| BERTScoreCategory=Semantic Consistency2026.05 | 48.67 | 11.15 | 7.45 | 5.93 | 93.55 | |
| Max token probabilityCategory=Model-Internal Confidence, LLM=Qwen2.5-7B2025.12 | 44.19 | 12.97 | 19.04 | 7.1 | 75 | |
| Max token probabilityCategory=Model-Internal Confidence2026.05 | 44.19 | 12.97 | 19.04 | 7.1 | 75 |