Hallucination Detection on NQ
0.889AUCMax Pooling
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Max PoolingBackbone=LLaMA-3.3-Instruct-70B2026.05 | 0.889 | — | — | — | — | |
| HaMI (SP)Backbone=LLaMA-3.3-Instruct-70B2026.05 | 0.885 | — | — | — | — | |
| Mean PoolingBackbone=LLaMA-3.3-Instruct-70B2026.05 | 0.87 | — | — | — | — | |
| PRModel=LLama-3.2-1B2026.01 | 0.8645 | — | — | — | — | |
| HaMI (SP)Backbone=Mistral-Nemo-Instruct (12B)2026.05 | 0.864 | — | — | — | — | |
| HaMI (Original)Backbone=LLaMA-3.3-Instruct-70B2026.05 | 0.862 | — | — | — | — | |
| Max PoolingBackbone=Mistral-Nemo-Instruct (12B)2026.05 | 0.859 | — | — | — | — | |
| Mean PoolingBackbone=Mistral-Nemo-Instruct (12B)2026.05 | 0.841 | — | — | — | — | |
| HaMI (SP)Backbone=LLaMA-3.1-8B2026.05 | 0.84 | — | — | — | — | |
| HaMI (Original)Backbone=Mistral-Nemo-Instruct (12B)2026.05 | 0.83 | — | — | — | — | |
| QAODQ⊕V⊥Model=Qwen3-30B-A3B2026.05 | 0.8276 | — | — | — | 81.76 | |
| QAODQ⊕V⊥Model=Qwen3-14B2026.05 | 0.8269 | — | — | — | 81.25 | |
| PRModel=LLama-3.2-3B2026.01 | 0.8191 | — | — | — | — | |
| Attention probe, soft targetRegime=Pre-gen., Model=Qwen2.5-32B2026.06 | 0.8152 | — | — | — | — | |
| QAODQ⊕V⊥Model=gemma-2-2b2026.05 | 0.8142 | — | — | — | 85.68 | |
| Max PoolingBackbone=LLaMA-3.1-8B2026.05 | 0.814 | — | — | — | — | |
| PRModel=Mistral-7B-v0.12026.01 | 0.8092 | — | — | — | — | |
| PfalseModel=Qwen3-30B-A3B2026.05 | 0.803 | — | — | — | 81.84 | |
| QAODV⊥Model=Qwen3-14B2026.05 | 0.7998 | — | — | — | 80.25 | |
| QAODV⊥Model=Qwen3-30B-A3B2026.05 | 0.7927 | — | — | — | 80.48 | |
| Mean PoolingBackbone=LLaMA-3.1-8B2026.05 | 0.792 | — | — | — | — | |
| MoPModel=LLama-3.2-1B2026.01 | 0.7919 | — | — | — | — | |
| PRModel=LLama-3-8B2026.01 | 0.791 | — | — | — | — | |
| PRBackbone=Llama-3-8B2026.01 | 0.791 | — | — | — | — | |
| PRBackbone=Mistral-7B-v0.32026.01 | 0.7909 | — | — | — | — | |
| PRModel=Mistral-7B-v0.32026.01 | 0.7909 | — | — | — | — | |
| PfalseModel=Qwen3-14B2026.05 | 0.7905 | — | — | — | 77.71 | |
| Attention probeRegime=Pre-gen., Model=Qwen2.5-32B2026.06 | 0.7897 | — | — | — | — | |
| SAPLMAModel=Qwen3-14B2026.05 | 0.7893 | — | — | — | 79.03 | |
| SAPLMAModel=Qwen3-30B-A3B2026.05 | 0.7878 | — | — | — | 80.27 | |
| SAPLMAModel=gemma-2-2b2026.05 | 0.784 | — | — | — | 84.96 | |
| SINdexModel=Qwen3-14B2026.05 | 0.783 | — | — | — | 74.65 | |
| PRModel=LLama-3-70B2026.01 | 0.7824 | — | — | — | — | |
| MoPModel=LLama-3.2-3B2026.01 | 0.7814 | — | — | — | — | |
| QAODQ⊕V⊥Model=Llama-2-7B-chat2026.05 | 0.7794 | — | — | — | 81.68 | |
| QAODV⊥Model=gemma-2-2b2026.05 | 0.7791 | — | — | — | 85.19 | |
| HaMI (Original)Backbone=LLaMA-3.1-8B2026.05 | 0.777 | — | — | — | — | |
| SINdexModel=Qwen3-30B-A3B2026.05 | 0.7733 | — | — | — | 79.43 | |
| Attention probe, soft targetRegime=Post-gen., Model=Qwen2.5-32B2026.06 | 0.7728 | — | — | — | — | |
| MoPBackbone=Mistral-7B-v0.32026.01 | 0.7687 | — | — | — | — | |
| MoPModel=Mistral-7B-v0.32026.01 | 0.7687 | — | — | — | — | |
| Attention probeRegime=Post-gen., Model=Qwen2.5-32B2026.06 | 0.7661 | — | — | — | — | |
| MoPModel=Mistral-7B-v0.12026.01 | 0.7638 | — | — | — | — | |
| MoPModel=LLama-3-70B2026.01 | 0.7612 | — | — | — | — | |
| MoP-VanillaExpertsModel=LLama-3.2-1B2026.01 | 0.7573 | — | — | — | — | |
| MoP-VanillaExpertsModel=LLama-3.2-3B2026.01 | 0.753 | — | — | — | — | |
| LSCModel=Qwen-32B2026.01 | 0.752 | 38.7 | — | — | — | |
| SINdexModel=gemma-2-2b2026.05 | 0.752 | — | — | — | 85.05 | |
| QAODV⊥Model=Llama-2-7B-chat2026.05 | 0.7508 | — | — | — | 81.2 | |
| Attention probeRegime=Post-gen., Model=Llama-3-8B2026.06 | 0.75 | — | — | — | — | |
| EigenScoreModel=Qwen-0.5B2026.01 | 0.749 | 30.4 | — | — | — | |
| Probing BaselineModel=LLama-3.2-1B2026.01 | 0.7483 | — | — | — | — | |
| MoP-VanillaExpertsModel=Mistral-7B-v0.12026.01 | 0.7478 | — | — | — | — | |
| MoPModel=LLama-3-8B2026.01 | 0.7464 | — | — | — | — | |
| MoPBackbone=Llama-3-8B2026.01 | 0.7464 | — | — | — | — | |
| Attention probe, soft targetRegime=Post-gen., Model=Qwen2.5-7B2026.06 | 0.7443 | — | — | — | — | |
| Probing BaselineModel=LLama-3-70B2026.01 | 0.7439 | — | — | — | — | |
| Attention probe, soft targetRegime=Post-gen., Model=Llama-3-8B2026.06 | 0.7436 | — | — | — | — | |
| Probing BaselineModel=LLama-3.2-3B2026.01 | 0.742 | — | — | — | — | |
| Attention probe, soft targetRegime=Pre-gen., Model=Llama-3-8B2026.06 | 0.7413 | — | — | — | — | |
| Probing BaselineModel=Mistral-7B-v0.12026.01 | 0.741 | — | — | — | — | |
| SAPLMAModel=Llama-2-7B-chat2026.05 | 0.74 | — | — | — | 80.9 | |
| MoP-VanillaExpertsModel=LLama-3-70B2026.01 | 0.7385 | — | — | — | — | |
| SINdexModel=Llama-2-7B-chat2026.05 | 0.7378 | — | — | — | 79.72 | |
| MoP-VanillaExpertsBackbone=Mistral-7B-v0.32026.01 | 0.7377 | — | — | — | — | |
| MoP-VanillaExpertsModel=Mistral-7B-v0.32026.01 | 0.7377 | — | — | — | — | |
| MINDModel=Qwen3-14B2026.05 | 0.7369 | — | — | — | 78.34 | |
| Probing BaselineBackbone=Mistral-7B-v0.32026.01 | 0.736 | — | — | — | — | |
| Probing BaselineModel=Mistral-7B-v0.32026.01 | 0.736 | — | — | — | — | |
| LSCModel=Qwen-0.5B2026.01 | 0.732 | 34 | — | — | — | |
| PerplexityModel=Qwen-0.5B2026.01 | 0.73 | 19.5 | — | — | — | |
| Attention probeRegime=Pre-gen., Model=Llama-3-8B2026.06 | 0.7294 | — | — | — | — | |
| MINDModel=Qwen3-30B-A3B2026.05 | 0.7269 | — | — | — | 78.3 | |
| Attention probeRegime=Post-gen., Model=Qwen2.5-7B2026.06 | 0.7264 | — | — | — | — | |
| Attention probe, soft targetRegime=Post-gen., Model=Gemma-2-9B2026.06 | 0.7245 | — | — | — | — | |
| MINDModel=gemma-2-2b2026.05 | 0.7241 | — | — | — | 84 | |
| Attention probeRegime=Post-gen., Model=Gemma-2-9B2026.06 | 0.7222 | — | — | — | — | |
| PerplexityModel=Qwen-32B2026.01 | 0.72 | 24.5 | — | — | — | |
| PikModel=gemma-2-2b2026.05 | 0.7158 | — | — | — | 83.72 | |
| Attention probe, soft targetRegime=Pre-gen., Model=Gemma-2-9B2026.06 | 0.7156 | — | — | — | — | |
| Linear probeRegime=Post-gen., Model=Qwen2.5-32B2026.06 | 0.7137 | — | — | — | — | |
| MoP-VanillaExpertsModel=LLama-3-8B2026.01 | 0.7121 | — | — | — | — | |
| MoP-VanillaExpertsBackbone=Llama-3-8B2026.01 | 0.7121 | — | — | — | — | |
| Attention probe, soft targetRegime=Pre-gen., Model=Qwen2.5-7B2026.06 | 0.7118 | — | — | — | — | |
| LN-EntropyModel=Qwen-0.5B2026.01 | 0.703 | 9.1 | — | — | — | |
| Attention probeRegime=Pre-gen., Model=Gemma-2-9B2026.06 | 0.7022 | — | — | — | — | |
| Probing BaselineModel=LLama-3-8B2026.01 | 0.702 | — | — | — | — | |
| Probing BaselineBackbone=Llama-3-8B2026.01 | 0.702 | — | — | — | — | |
| LN-EntropyModel=Qwen-32B2026.01 | 0.702 | 25.4 | — | — | — | |
| Linear probeRegime=Post-gen., Model=Qwen2.5-7B2026.06 | 0.7006 | — | — | — | — | |
| PikModel=Qwen3-30B-A3B2026.05 | 0.6999 | — | — | — | 77.61 | |
| PfalseModel=gemma-2-2b2026.05 | 0.6989 | — | — | — | 83.73 | |
| Linear probeRegime=Post-gen., Model=Gemma-2-9B2026.06 | 0.6968 | — | — | — | — | |
| Logit EntropyModel=gemma-2-2b2026.05 | 0.6914 | — | — | — | 84.44 | |
| Attention probeRegime=Pre-gen., Model=Qwen2.5-7B2026.06 | 0.6898 | — | — | — | — | |
| MoP-RandomGateModel=LLama-3.2-1B2026.01 | 0.6876 | — | — | — | — | |
| Lexical SimilarityModel=Qwen-0.5B2026.01 | 0.682 | 14.8 | — | — | — | |
| EigenScoreModel=Qwen-32B2026.01 | 0.681 | 31.9 | — | — | — | |
| PfalseModel=Llama-2-7B-chat2026.05 | 0.6808 | — | — | — | 79.57 | |
| MoP-RandomGateModel=Mistral-7B-v0.12026.01 | 0.6805 | — | — | — | — |