Hallucination Detection on BoolQ (ROC-AUC)
94.76ROC-AUCLinear probe
Evaluation Results
| Method | Links | |
|---|---|---|
| Linear probeModel=Qwen2.5-7B, Generation Stage=Post-generation2026.06 | 94.76 | |
| Attention probeModel=Qwen2.5-7B, Generation Stage=Post-generation2026.06 | 94.46 | |
| Attention probeModel=Mistral-7B, Generation Stage=Pre-generation2026.06 | 93.21 | |
| Attention probeModel=Mistral-7B, Generation Stage=Post-generation2026.06 | 93.18 | |
| Linear probeModel=Mistral-7B, Generation Stage=Post-generation2026.06 | 92.24 | |
| Attention probeModel=Qwen2.5-3B, Generation Stage=Post-generation2026.06 | 92.18 | |
| Attention probeModel=Qwen2.5-3B, Generation Stage=Pre-generation2026.06 | 92.03 | |
| Linear probeModel=Qwen2.5-3B, Generation Stage=Post-generation2026.06 | 91.42 | |
| Attention probeModel=Llama-2-7B, Generation Stage=Post-generation2026.06 | 90.94 | |
| Linear probeModel=Llama-2-7B, Generation Stage=Post-generation2026.06 | 90.66 | |
| Attention probeModel=Llama-2-7B, Generation Stage=Pre-generation2026.06 | 89.88 | |
| Attention probeModel=Qwen2.5-7B, Generation Stage=Pre-generation2026.06 | 84.23 | |
| Linear probeModel=Qwen2.5-7B, Generation Stage=Pre-generation2026.06 | 82.25 | |
| Linear probeModel=Mistral-7B, Generation Stage=Pre-generation2026.06 | 78.99 | |
| EntropyModel=Qwen2.5-7B, Generation Stage=Post-generation2026.06 | 78.35 | |
| Linear probeModel=Qwen2.5-3B, Generation Stage=Pre-generation2026.06 | 69.65 | |
| Linear probeModel=Llama-2-7B, Generation Stage=Pre-generation2026.06 | 68.64 | |
| EntropyModel=Qwen2.5-3B, Generation Stage=Post-generation2026.06 | 68.58 | |
| Self-check, zero-shotModel=Qwen2.5-7B, Generation Stage=Pre-generation2026.06 | 64.19 | |
| Self-check, zero-shotModel=Qwen2.5-3B, Generation Stage=Pre-generation2026.06 | 63.36 | |
| EntropyModel=Llama-2-7B, Generation Stage=Post-generation2026.06 | 62.63 | |
| Self-check, zero-shotModel=Mistral-7B, Generation Stage=Pre-generation2026.06 | 61.11 | |
| Self-check, zero-shotModel=Llama-2-7B, Generation Stage=Pre-generation2026.06 | 52.65 | |
| Question lengthModel=Mistral-7B, Generation Stage=Pre-generation2026.06 | 52.52 | |
| EntropyModel=Mistral-7B, Generation Stage=Post-generation2026.06 | 51.08 | |
| Question lengthModel=Qwen2.5-3B, Generation Stage=Pre-generation2026.06 | 51.08 | |
| Question lengthModel=Llama-2-7B, Generation Stage=Pre-generation2026.06 | 51 | |
| Question lengthModel=Qwen2.5-7B, Generation Stage=Pre-generation2026.06 | 50.85 |