Misalignment Detection on WorkBench (n=326)
52Average IRProvePrompt
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| ProvePromptModel=Gemini 3 Flash2026.05 | 52 | 0.106 | 1 | |
| ProvePromptModel=GPT-52026.05 | 49.5 | 0.044 | 1 | |
| DirectPromptModel=GPT-52026.05 | 45.2 | — | — | |
| ProvePromptModel=GPT-5 Mini2026.05 | 44.4 | 0.071 | 1 | |
| ProvePromptModel=Claude Sonnet 4.62026.05 | 43.1 | 0.037 | 1 | |
| DirectPromptModel=Gemini 3 Flash2026.05 | 41.4 | — | — | |
| ProvePromptModel=Claude Haiku 4.52026.05 | 40.7 | 0.025 | 0.978 | |
| DirectPromptModel=Claude Sonnet 4.62026.05 | 39.4 | — | — | |
| DirectPromptModel=Claude Haiku 4.52026.05 | 38.3 | — | — | |
| ProvePromptModel=Gemini 2.5 Flash2026.05 | 37.5 | 0.133 | 1 | |
| DirectPromptModel=GPT-5 Mini2026.05 | 37.3 | — | — | |
| ProvePromptModel=Gemini 3.1 Flash-Lite2026.05 | 29.4 | 0.096 | 1 | |
| ProvePromptModel=GPT-4.1 Mini2026.05 | 27.5 | 0.007 | 0.659 | |
| ProvePromptModel=GPT-4.12026.05 | 26.9 | 0.116 | 1 | |
| DirectPromptModel=GPT-4.1 Mini2026.05 | 26.7 | — | — | |
| DirectPromptModel=Gemini 2.5 Flash2026.05 | 24.2 | — | — | |
| DirectPromptModel=Gemini 3.1 Flash-Lite2026.05 | 19.8 | — | — | |
| ProvenanceGuardModel=Claude Haiku 4.52026.05 | 18.1 | -0.202 | 0.001 | |
| ProvenanceGuardModel=Gemini 3 Flash2026.05 | 17.2 | -0.242 | 0.001 | |
| DirectPromptModel=Gemini 2.5 Flash-Lite2026.05 | 16.9 | — | — | |
| DirectPromptModel=GPT-4.12026.05 | 15.3 | — | — | |
| ProvenanceGuardModel=GPT-4.12026.05 | 15.2 | -0.002 | 0.479 | |
| ProvenanceGuardModel=GPT-5 Mini2026.05 | 14.2 | -0.231 | 0.001 | |
| ProvePromptModel=Gemini 2.5 Flash-Lite2026.05 | 13.9 | -0.031 | 0.006 | |
| ProvenanceGuardModel=GPT-52026.05 | 13.2 | -0.32 | 0.001 | |
| ProvenanceGuardModel=Gemini 3.1 Flash-Lite2026.05 | 12.5 | -0.073 | 0.011 | |
| ProvenanceGuardModel=GPT-4.1 Mini2026.05 | 11.9 | -0.148 | 0.001 | |
| ProvenanceGuardModel=Gemini 2.5 Flash2026.05 | 11 | -0.132 | 0.001 | |
| ProvenanceGuardModel=Gemini 2.5 Flash-Lite2026.05 | 10 | -0.069 | 0.005 | |
| ProvenanceGuardModel=Claude Sonnet 4.62026.05 | 4.8 | -0.346 | 0.001 |