Steering on Corrigibility
3.22LLM Judge ScoreDISCO-Q
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| DISCO-QBackbone=LLaMA-3.1-8B-Instruct2026.05 | 3.22 | 4.75 | |
| SKOPBackbone=LLaMA-3.1-8B-Instruct2026.05 | 3.19 | 2.69 | |
| Comm SteerBackbone=LLaMA-3.1-8B-Instruct2026.05 | 3.01 | 4.63 | |
| CAABackbone=LLaMA-3.1-8B-Instruct2026.05 | 2.79 | 5.44 | |
| LoRABackbone=LLaMA-3.1-8B-Instruct2026.05 | 2.68 | — | |
| ITIBackbone=LLaMA-3.1-8B-Instruct2026.05 | 2.6 | 4.75 | |
| SADIBackbone=LLaMA-3.1-8B-Instruct2026.05 | 2.49 | 3.38 | |
| CASTBackbone=LLaMA-3.1-8B-Instruct2026.05 | 2.28 | 5.63 | |
| Angular SteerBackbone=LLaMA-3.1-8B-Instruct2026.05 | 2.18 | 4.75 | |
| BaselineBackbone=LLaMA-3.1-8B-Instruct2026.05 | 1.94 | — |