Hallucination Detection on SQuAD (AUCs, AUCr, PCC Metrics)
69.9AUCsLSC
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| LSCModel=LLaMA-7B2026.01 | 69.9 | 69.4 | 0.327 | — | |
| LSCModel=LLaMA-13B2026.01 | 69.1 | 69.7 | 0.348 | — | |
| EigenScoreModel=LLaMA-7B2026.01 | 67.5 | 67.8 | 0.342 | — | |
| EigenScoreModel=LLaMA-13B2026.01 | 66.1 | 66.4 | 0.327 | — | |
| Lexical SimilarityModel=LLaMA-7B2026.01 | 64.2 | 65.4 | 0.252 | — | |
| Lexical SimilarityModel=LLaMA-13B2026.01 | 62 | 64.3 | 0.234 | — | |
| LSCModel=Qwen-3B2026.01 | 61.3 | 62.1 | 0.197 | — | |
| EigenScoreModel=Qwen-7B2026.01 | 58.5 | 60 | 0.142 | — | |
| EigenScoreModel=Qwen-3B2026.01 | 57.9 | 59.5 | 0.153 | — | |
| Lexical SimilarityModel=Qwen-7B2026.01 | 57.2 | 59.8 | 0.129 | — | |
| Lexical SimilarityModel=Qwen-3B2026.01 | 55.1 | 58.6 | 0.12 | — | |
| AGSERModel=Qwen-3B2026.01 | 54.2 | 57.3 | 0.122 | — | |
| LSCModel=Qwen-7B2026.01 | 53.8 | 56 | 0.073 | — | |
| LN-EntropyModel=LLaMA-13B2026.01 | 53.5 | 56.4 | 0.098 | — | |
| LN-EntropyModel=LLaMA-7B2026.01 | 53.5 | 56.5 | 0.072 | — | |
| AGSERModel=LLaMA-13B2026.01 | 50.8 | 51.9 | -0.015 | — | |
| LN-EntropyModel=Qwen-3B2026.01 | 49.5 | 53.6 | 0.048 | — | |
| LN-EntropyModel=Qwen-7B2026.01 | 48.9 | 51.6 | -0.012 | — | |
| AGSERModel=Qwen-7B2026.01 | 45.9 | 48.8 | -0.076 | — | |
| PerplexityModel=Qwen-3B2026.01 | 45.8 | 51.4 | 0.039 | — | |
| AGSERModel=LLaMA-7B2026.01 | 43.2 | 45.5 | -0.053 | — | |
| PerplexityModel=LLaMA-7B2026.01 | 40.6 | 46.2 | -0.055 | — | |
| PerplexityModel=LLaMA-13B2026.01 | 39.7 | 44.8 | -0.066 | — | |
| EnergyModel=Qwen-3B2026.01 | 38.9 | 42.4 | -0.07 | — | |
| PerplexityModel=Qwen-7B2026.01 | 38.3 | 42 | -0.14 | — | |
| EnergyModel=LLaMA-13B2026.01 | 35.4 | 38.5 | -0.211 | — | |
| EnergyModel=LLaMA-7B2026.01 | 32.2 | 37.4 | -0.189 | — | |
| EnergyModel=Qwen-7B2026.01 | 28.4 | 30.8 | -0.345 | — | |
| HaMI (Original)Backbone=LLaMA-3.1-8B2026.05 | — | — | — | 79.5 | |
| HaMI (Original)Backbone=Mistral-Nemo-Instruct (12B)2026.05 | — | — | — | 84 | |
| HaMI (Original)Backbone=LLaMA-3.3-Instruct-70B2026.05 | — | — | — | 84.1 | |
| HaMI (SP)Backbone=LLaMA-3.1-8B2026.05 | — | — | — | 84.4 | |
| HaMI (SP)Backbone=Mistral-Nemo-Instruct (12B)2026.05 | — | — | — | 85.8 | |
| HaMI (SP)Backbone=LLaMA-3.3-Instruct-70B2026.05 | — | — | — | 87.4 | |
| Max PoolingBackbone=LLaMA-3.1-8B2026.05 | — | — | — | 82.9 | |
| Max PoolingBackbone=Mistral-Nemo-Instruct (12B)2026.05 | — | — | — | 85.9 | |
| Max PoolingBackbone=LLaMA-3.3-Instruct-70B2026.05 | — | — | — | 87.4 | |
| Mean PoolingBackbone=LLaMA-3.1-8B2026.05 | — | — | — | 81.7 | |
| Mean PoolingBackbone=Mistral-Nemo-Instruct (12B)2026.05 | — | — | — | 84.6 | |
| Mean PoolingBackbone=LLaMA-3.3-Instruct-70B2026.05 | — | — | — | 85.4 |