Multiple-choice Question Answering on Anthropic Behavioral Concepts CAA-style (Negative Steering)
8AI CoordinationSAS
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| SASSteering Method=SAS, Steering Strength=alpha < 0 (minimum over {-2, -1}), System Prompt Condition=no system prompt, Steering Layer=22, Model=GEMMA-2-9B-IT2026.06 | 8 | — | 27 | 28 | 16 | 67 | 44 | |
| CAASteering Method=CAA, Steering Strength=alpha < 0 (minimum over {-2, -1}), System Prompt Condition=no system prompt, Steering Layer=22, Model=GEMMA-2-9B-IT2026.06 | 6 | — | 19 | 24 | 17 | 35 | 35 | |
| No steeringSteering Method=None, Model=GEMMA-2-9B-IT2026.06 | 6 | — | 44 | 29 | 24 | 67 | 53 | |
| HiDRASteering Method=HiDRA, Steering Strength=alpha < 0 (minimum over {-2, -1}), System Prompt Condition=no system prompt, Steering Layer=22, Model=GEMMA-2-9B-IT2026.06 | 4 | — | 9 | 8 | 7 | 29 | 27 |