Unintended behavior elicitation on AUTOELICIT-SEED Multi-Apps domain
90Elicitation Success Rate (Seed)Claude 4.5 Opus
Evaluation Results
| Method | Links | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Claude 4.5 OpusRefinement Model=GPT-5, # Seeds=30, # Tasks=202026.02 | 90 | 90 | 80 | 75 | 20 | 0 | 13.3 | 40 | 20 | 6.7 | |
| Claude 4.5 OpusRefinement Model=Claude 4.5 Haiku, # Seeds=30, # Tasks=182026.02 | 80 | 83.3 | 66.7 | 72.2 | 33.3 | 3.3 | 3.3 | 30 | 23.3 | 6.7 | |
| Claude 4.5 HaikuRefinement Model=GPT-5, # Seeds=194, # Tasks=442026.02 | 60.8 | 81.8 | — | — | 38.1 | 3.6 | 18.6 | 30.4 | 6.2 | — | |
| Claude 4.5 HaikuRefinement Model=Claude 4.5 Haiku, # Seeds=194, # Tasks=442026.02 | 58.2 | 95.5 | — | — | 41.2 | 2.1 | 13.9 | 33 | 6.2 | 3.6 |