Medical LLM Risk Triage on RETINA-SAFE Stage-1
99.92Unsafe RecallSelfCheckGPT
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| SelfCheckGPTBackbone=Llama 3-8B, Target-recall policy=strict, Seed=seed132026.04 | 99.92 | 99.95 | 50 | -0.3405 | |
| EigenScoreBackbone=Llama 2-7B, Target-recall policy=strict, Seed=seed132026.04 | 99.91 | 95.82 | 72.41 | -0.1988 | |
| FocusBackbone=Llama 2-7B, Target-recall policy=strict, Seed=seed132026.04 | 99.84 | 94.12 | 52.84 | -0.3945 | |
| P(True)Backbone=Llama 2-7B, Target-recall policy=strict, Seed=seed132026.04 | 99.82 | 91.24 | 56.43 | -0.3586 | |
| SelfCheckGPTBackbone=Qwen2.5-7B, Target-recall policy=strict, Seed=seed132026.04 | 99.81 | 99.95 | 50 | -0.4679 | |
| FocusBackbone=Qwen2.5-7B, Target-recall policy=strict, Seed=seed132026.04 | 98.98 | 99.08 | 49.49 | -0.473 | |
| EigenScoreBackbone=Qwen3-8B, Target-recall policy=strict, Seed=seed132026.04 | 96.42 | 95.57 | 54.38 | -0.4202 | |
| FactoScopeBackbone=Qwen2.5-7B, Target-recall policy=strict, Seed=seed132026.04 | 96.34 | 87.12 | 94.12 | -0.0267 | |
| ReDeEPBackbone=Qwen2.5-7B, Target-recall policy=strict, Seed=seed132026.04 | 96.23 | 89.12 | 78.42 | -0.1837 | |
| UQ HeadsBackbone=Qwen2.5-7B, Target-recall policy=strict, Seed=seed132026.04 | 96.12 | 87.84 | 91.23 | -0.0556 | |
| RefCheckerBackbone=Qwen2.5-7B, Target-recall policy=strict, Seed=seed132026.04 | 96.12 | 90.12 | 72.41 | -0.2438 | |
| Lookback LensBackbone=Qwen2.5-7B, Target-recall policy=strict, Seed=seed132026.04 | 96.02 | 88.21 | 87.42 | -0.0937 | |
| LN-EntropyBackbone=Qwen2.5-7B, Target-recall policy=strict, Seed=seed132026.04 | 95.92 | 91.24 | 68.21 | -0.2858 | |
| EigenScoreBackbone=Qwen2.5-7B, Target-recall policy=strict, Seed=seed132026.04 | 95.76 | 95.93 | 49.11 | -0.4768 | |
| FocusBackbone=Qwen3-8B, Target-recall policy=strict, Seed=seed132026.04 | 95.67 | 93.45 | 61.41 | -0.3499 | |
| ECRTBackbone=Qwen2.5-7B, Target-recall policy=strict, Seed=seed132026.04 | 95.66 | 86.65 | 96.79 | — | |
| FactoScopeBackbone=Llama 2-7B, Target-recall policy=strict, Seed=seed132026.04 | 95.56 | 87.54 | 85.46 | -0.0683 | |
| FactoScopeBackbone=Mistral-7B, Target-recall policy=strict, Seed=seed132026.04 | 95.51 | 87.23 | 91.24 | -0.0186 | |
| FactoScopeBackbone=Llama 3-8B, Target-recall policy=strict, Seed=seed132026.04 | 95.42 | 88.56 | 81.25 | -0.028 | |
| ReDeEPBackbone=Mistral-7B, Target-recall policy=strict, Seed=seed132026.04 | 95.42 | 90.54 | 74.23 | -0.1887 | |
| FactoScopeBackbone=Qwen3-8B, Target-recall policy=strict, Seed=seed132026.04 | 95.42 | 86.24 | 93.42 | -0.0298 | |
| ReDeEPBackbone=Llama 2-7B, Target-recall policy=strict, Seed=seed132026.04 | 95.41 | 89.54 | 73.25 | -0.1904 | |
| RefCheckerBackbone=Llama 2-7B, Target-recall policy=strict, Seed=seed132026.04 | 95.34 | 90.87 | 69.25 | -0.2304 | |
| ReDeEPBackbone=Qwen3-8B, Target-recall policy=strict, Seed=seed132026.04 | 95.34 | 90.12 | 78.42 | -0.1798 | |
| ReDeEPBackbone=Llama 3-8B, Target-recall policy=strict, Seed=seed132026.04 | 95.31 | 90.82 | 69.23 | -0.1482 | |
| RefCheckerBackbone=Mistral-7B, Target-recall policy=strict, Seed=seed132026.04 | 95.31 | 91.23 | 70.21 | -0.2289 | |
| FocusBackbone=Llama 3-8B, Target-recall policy=strict, Seed=seed132026.04 | 95.28 | 94.11 | 54.88 | -0.2917 | |
| UQ HeadsBackbone=Llama 3-8B, Target-recall policy=strict, Seed=seed132026.04 | 95.25 | 89.24 | 78.56 | -0.0549 | |
| RefCheckerBackbone=Qwen3-8B, Target-recall policy=strict, Seed=seed132026.04 | 95.25 | 91.12 | 71.24 | -0.2516 | |
| LN-EntropyBackbone=Mistral-7B, Target-recall policy=strict, Seed=seed132026.04 | 95.24 | 92.54 | 64.21 | -0.2889 | |
| LN-EntropyBackbone=Llama 2-7B, Target-recall policy=strict, Seed=seed132026.04 | 95.22 | 91.56 | 65.42 | -0.2687 | |
| UQ HeadsBackbone=Mistral-7B, Target-recall policy=strict, Seed=seed132026.04 | 95.21 | 88.12 | 86.23 | -0.0687 | |
| Lookback LensBackbone=Llama 3-8B, Target-recall policy=strict, Seed=seed132026.04 | 95.18 | 91.25 | 71.52 | -0.1253 | |
| Lookback LensBackbone=Mistral-7B, Target-recall policy=strict, Seed=seed132026.04 | 95.18 | 89.12 | 81.24 | -0.1186 | |
| RefCheckerBackbone=Llama 3-8B, Target-recall policy=strict, Seed=seed132026.04 | 95.12 | 92.48 | 64.55 | -0.195 | |
| UQ HeadsBackbone=Llama 2-7B, Target-recall policy=strict, Seed=seed132026.04 | 95.12 | 88.24 | 82.45 | -0.0984 | |
| PerplexityBackbone=Mistral-7B, Target-recall policy=strict, Seed=seed132026.04 | 95.12 | 94.23 | 52.14 | -0.4096 | |
| UQ HeadsBackbone=Qwen3-8B, Target-recall policy=strict, Seed=seed132026.04 | 95.12 | 87.12 | 89.24 | -0.0716 | |
| LN-EntropyBackbone=Qwen3-8B, Target-recall policy=strict, Seed=seed132026.04 | 95.11 | 92.14 | 65.42 | -0.3098 | |
| P(True)Backbone=Llama 3-8B, Target-recall policy=strict, Seed=seed132026.04 | 95.1 | 94.9 | 50.84 | -0.3321 | |
| SelfCheckGPTBackbone=Llama 2-7B, Target-recall policy=strict, Seed=seed132026.04 | 95.1 | 93.25 | 61.25 | -0.3104 | |
| Lookback LensBackbone=Llama 2-7B, Target-recall policy=strict, Seed=seed132026.04 | 95.08 | 89.12 | 78.54 | -0.1375 | |
| SelfCheckGPTBackbone=Mistral-7B, Target-recall policy=strict, Seed=seed132026.04 | 95.08 | 93.12 | 60.21 | -0.3289 | |
| LN-EntropyBackbone=Llama 3-8B, Target-recall policy=strict, Seed=seed132026.04 | 95.06 | 93.15 | 61.24 | -0.2281 | |
| ECRTBackbone=Llama 3-8B, Target-recall policy=strict, Seed=seed132026.04 | 95.05 | 87.12 | 84.05 | — | |
| PerplexityBackbone=Qwen2.5-7B, Target-recall policy=strict, Seed=seed132026.04 | 95.05 | 95.13 | 49.58 | -0.4721 | |
| Lookback LensBackbone=Qwen3-8B, Target-recall policy=strict, Seed=seed132026.04 | 95.05 | 88.12 | 84.23 | -0.1217 | |
| PerplexityBackbone=Qwen3-8B, Target-recall policy=strict, Seed=seed132026.04 | 95.02 | 94.12 | 54.23 | -0.4217 | |
| P(True)Backbone=Qwen3-8B, Target-recall policy=strict, Seed=seed132026.04 | 95 | 95.49 | 47.5 | -0.489 | |
| ECRTBackbone=Qwen3-8B, Target-recall policy=strict, Seed=seed132026.04 | 94.88 | 85.95 | 96.4 | — | |
| PerplexityBackbone=Llama 2-7B, Target-recall policy=strict, Seed=seed132026.04 | 94.85 | 92.54 | 58.42 | -0.3387 | |
| SelfCheckGPTBackbone=Qwen3-8B, Target-recall policy=strict, Seed=seed132026.04 | 94.82 | 93.24 | 62.41 | -0.3399 | |
| EigenScoreBackbone=Mistral-7B, Target-recall policy=strict, Seed=seed132026.04 | 94.78 | 92.61 | 61.18 | -0.3192 | |
| EigenScoreBackbone=Llama 3-8B, Target-recall policy=strict, Seed=seed132026.04 | 94.74 | 92.13 | 63.21 | -0.2084 | |
| ECRTBackbone=Mistral-7B, Target-recall policy=strict, Seed=seed132026.04 | 94.66 | 86.36 | 93.1 | — | |
| PerplexityBackbone=Llama 3-8B, Target-recall policy=strict, Seed=seed132026.04 | 94.65 | 92.97 | 58.64 | -0.2541 | |
| P(True)Backbone=Mistral-7B, Target-recall policy=strict, Seed=seed132026.04 | 94.61 | 94.69 | 49.57 | -0.4353 | |
| FocusBackbone=Mistral-7B, Target-recall policy=strict, Seed=seed132026.04 | 94.56 | 93.37 | 56.13 | -0.3697 | |
| P(True)Backbone=Qwen2.5-7B, Target-recall policy=strict, Seed=seed132026.04 | 94.52 | 94.33 | 50.96 | -0.4583 | |
| ECRTBackbone=Llama 2-7B, Target-recall policy=strict, Seed=seed132026.04 | 94.3 | 86.37 | 92.29 | — |