Share your thoughts, 1 month free Claude Pro on us
See more
Home
/
Benchmarks
Refusal Steering on Phi/Qwen judged grid
Loading...
33.3
Target Success Rate
GoR/DIM ablation
0.956
9.353
17.75
26.147
May 18, 2026
Target Success Rate
Benign Keep Rate
Pairwise NW Score
Harm Attack Success Rate (ASR)
Updated 13d ago
Evaluation Results
Method
Method
Links
Target Success Rate
Benign Keep Rate
Pairwise NW Score
Harm Attack Success Rate (ASR)
GoR/DIM ablation
Evidence block=Externa...
2026.05
33.3
87.6
57.3
32.5
Refusal direction ablation
Evidence block=Externa...
2026.05
32.6
87.6
56.6
33.8
Refusal direction ActAdd
Evidence block=Externa...
2026.05
30.6
90.1
62.2
39.7
GoR/DIM act-sub
Evidence block=Externa...
2026.05
29
90
62.5
36.6
GoR-RDO ablation
Evidence block=Externa...
2026.05
27.9
98.5
74.8
29.1
RepIt projection
Evidence block=Externa...
2026.05
5.3
98.5
68.9
9.8
No intervention
Evidence block=Externa...
2026.05
2.2
98.4
100
3.5
Feedback
Search any
task
Search any
task