Unintended behavior elicitation on AUTOELICIT-SEED OS domain
93.3Elicitation Success (Seed)Claude 4.5 Opus
Evaluation Results
| Method | Links | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Claude 4.5 OpusRefinement Model=GPT-5, # Seeds=30, # Tasks=142026.02 | 93.3 | 92.9 | 56.7 | 91.7 | 43.3 | 0 | 10 | 33.3 | 10 | 3.3 | |
| Claude 4.5 OpusRefinement Model=Claude 4.5 Haiku, # Seeds=30, # Tasks=122026.02 | 80 | 100 | 60 | 90 | 40 | 3.3 | 3.3 | 30 | 20 | 3.3 | |
| Claude 4.5 HaikuRefinement Model=GPT-5, # Seeds=109, # Tasks=142026.02 | 72.5 | 100 | — | — | 27.5 | 6.4 | 23.9 | 32.1 | 8.3 | 1.8 | |
| Claude 4.5 HaikuRefinement Model=Claude 4.5 Haiku, # Seeds=109, # Tasks=142026.02 | 70.6 | 92.9 | — | — | 29.4 | 2.8 | 20.2 | 38.5 | 6.4 | 2.8 |