Activation Steering on Matched-bucket Diagnostic Buckets comparator export protocol
86.8Edit Reference RateEdit-target ActAdd
Evaluation Results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Edit-target ActAddScale=22026.05 | 86.8 | 100 | 99.9 | 99.9 | 55.3 | 52.6 | -2.6 | 100 | 100 | |
| Edit-target ActAdd + vetoScale=22026.05 | 86.8 | 100 | 100 | 100 | 55.3 | 55.3 | 0 | 0 | 5.3 | |
| Edit-target oracle ActAddScale=22026.05 | 86.8 | 100 | 100 | 100 | 55.3 | 55.3 | 0 | 0 | 0 | |
| DIM-style refusal directionScale=82026.05 | 78.9 | 100 | 99.8 | 99.8 | 55.3 | 52.6 | -2.6 | 100 | 100 | |
| DIM-style refusal direction + vetoScale=82026.05 | 78.9 | 100 | 100 | 100 | 55.3 | 55.3 | 0 | 0 | 5.3 | |
| DIM-style oracle refusal directionScale=82026.05 | 78.9 | 100 | 100 | 100 | 55.3 | 55.3 | 0 | 0 | 0 | |
| Probe-gated ActAdd + vetoScale=16, Model=Gemma-2-2B-IT2026.05 | 78.9 | — | 100 | 100 | — | 78.9 | 0 | — | — | |
| Oracle-gated ActAddScale=16, Model=Gemma-2-2B-IT2026.05 | 76.3 | — | 100 | 100 | — | 78.9 | 0 | — | — | |
| Probe-gated ActAdd + vetoScale=8, Model=Qwen2.5-7B-Instruct2026.05 | 15.8 | — | 100 | 90.4 | — | 47.4 | -2.6 | — | — | |
| Probe-gated ActAdd + vetoScale=4, Model=Qwen2.5-1.5B-Instruct2026.05 | 10.5 | — | 100 | 95.9 | — | 84.2 | -2.6 | — | — | |
| Oracle-gated ActAddScale=16, Model=Qwen2.5-7B-Instruct2026.05 | 5.3 | — | 100 | 100 | — | 50 | 0 | — | — | |
| Oracle-gated ActAddScale=16, Model=Qwen2.5-1.5B-Instruct2026.05 | 0 | — | 100 | 100 | — | 86.8 | 0 | — | — |