Language Steering on WikiMatrix
44.2ES ScoreDiffMean
Evaluation Results
| Method | Links | |||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| DiffMeanModel=Llama 3.1 8B2026.04 | 44.2 | 43.1 | 25.3 | 31.5 | 13.1 | 32.6 | 12.3 | 6.9 | 17.6 | 9.3 | 26.6 | 28.4 | 24.2 | |
| LangFIRModel=Llama 3.1 8B2026.04 | 42.5 | 48.7 | 32.2 | 41.2 | 26.9 | 52.9 | 1 | 18.9 | 24 | 10.6 | 28.2 | 0.2 | 27.3 | |
| LangFIRModel=Gemma 3 4B2026.04 | 40.6 | 40.1 | 31.9 | 33.2 | 31.8 | 50.7 | 24 | 22.7 | 18.3 | 15.4 | 19.3 | 42.2 | 30.9 | |
| SAE-DMModel=Gemma 3 4B2026.04 | 40.2 | 39 | 30.5 | 26.2 | 29.6 | 44.1 | 29.6 | 15.7 | 19.7 | 12.3 | 9.1 | 20.9 | 26.4 | |
| LDAModel=Llama 3.1 8B2026.04 | 39.5 | 42.6 | 31.7 | 26.7 | 24.4 | 30.9 | 7.1 | 16.5 | 25.5 | 13.7 | 12.1 | 36 | 25.6 | |
| DiffMeanModel=Gemma 3 4B2026.04 | 38.2 | 37.8 | 30.7 | 36.1 | 29.1 | 47.1 | 30.5 | 17 | 23.3 | 10.6 | 10.8 | 24.1 | 27.9 | |
| Gated-DMModel=Llama 3.1 8B2026.04 | 37.7 | 39.7 | 32.8 | 28 | 12.8 | 38.7 | 15.8 | 10.2 | 20.3 | 10.4 | 28.5 | 34.8 | 25.8 | |
| LDAModel=Gemma 3 4B2026.04 | 36 | 31.2 | 18.9 | 26 | 21.7 | 38.9 | 17.6 | 5.9 | 9.4 | 4.3 | 3.3 | 22.2 | 19.6 | |
| DiffMeanModel=Gemma 3 1B2026.04 | 34.9 | 26.9 | 28.1 | 18.7 | 29.9 | 22.3 | 15.7 | 2.4 | 5.6 | 4.5 | 12.5 | 4.8 | 17.2 | |
| SAE-DMModel=Gemma 3 1B2026.04 | 34.8 | 25.7 | 27.7 | 18.6 | 29 | 21.9 | 16.2 | 2.4 | 5.6 | 3.6 | 11.8 | 4.5 | 16.8 | |
| LangFIRModel=Gemma 3 1B2026.04 | 32.5 | 30.5 | 27.4 | 23.4 | 21.5 | 28 | 17.8 | 1.3 | 11.4 | 4.8 | 10 | 20.4 | 19.1 | |
| Zhong (Mono)Model=Gemma 3 4B2026.04 | 18.7 | 21.7 | 14 | 15.4 | 17.2 | 28 | 7.9 | 9.1 | 6.4 | 4.7 | 13.9 | 18.9 | 14.7 | |
| Gated-DMModel=Gemma 3 4B2026.04 | 18.6 | 12.4 | 12.4 | 9.1 | 10.7 | 7 | 3 | 4.5 | 3.9 | 2.5 | 16.8 | 4.2 | 8.8 | |
| Zhong (Para)Model=Gemma 3 1B2026.04 | 16 | 13.6 | 13.6 | 9.4 | 9.1 | 22.6 | 14 | 0.8 | 5.8 | 1.8 | 5.8 | 9.5 | 10.2 | |
| Gated-DMModel=Gemma 3 1B2026.04 | 15.1 | 11.5 | 21 | 11 | 18.4 | 13.7 | 2.4 | 1 | 3.9 | 2.8 | 0.6 | 3.4 | 8.7 | |
| Zhong (Para)Model=Gemma 3 4B2026.04 | 12.1 | 26.5 | 28.5 | 11 | 27.2 | 20 | 12.1 | 11.8 | 14.3 | 9.5 | 19.7 | 22.7 | 18 | |
| No SteeringModel=Gemma 3 4B2026.04 | 11.7 | 8.2 | 5 | 2.5 | 2.7 | 1.8 | 0.6 | 1.5 | 3.7 | 1 | 2.9 | 1.4 | 3.6 | |
| Zhong (Para)Model=Llama 3.1 8B2026.04 | 11 | 19.6 | 11 | 8 | 1 | 17.4 | 0.6 | 6.1 | 8.2 | 2.6 | 0.6 | 28.7 | 9.6 | |
| Zhong (Mono)Model=Llama 3.1 8B2026.04 | 10.6 | 13.8 | 9.3 | 3.1 | 2.1 | 10.9 | 0.6 | 4.7 | 9.3 | 2.6 | 0.3 | 1.9 | 5.8 | |
| PCAModel=Gemma 3 4B2026.04 | 10.1 | 13.5 | 6.6 | 1.1 | 2.5 | 4 | 0.6 | 2.3 | 2.8 | 0.1 | 3 | 0.6 | 3.9 | |
| SAE-DMModel=Llama 3.1 8B2026.04 | 9.7 | 15.6 | 11.4 | 3 | 1.7 | 9 | 0.6 | 4.8 | 8.4 | 3 | 2 | 1.5 | 5.9 | |
| PCAModel=Llama 3.1 8B2026.04 | 5.7 | 10.4 | 6.5 | 3.9 | 0.2 | 4.6 | 0.6 | 3.6 | 7.4 | 2.6 | 1.1 | 1.2 | 4 | |
| No SteeringModel=Llama 3.1 8B2026.04 | 4.5 | 6.9 | 5.5 | 2.2 | 0.1 | 3.7 | 0.6 | 3.7 | 6.3 | 1.9 | 0.5 | 0.4 | 3 | |
| Zhong (Mono)Model=Gemma 3 1B2026.04 | 2.8 | 5.7 | 9.2 | 3.4 | 3.5 | 8.6 | 3 | 0.8 | 5.5 | 2.1 | 3 | 5.1 | 4.4 | |
| No SteeringModel=Gemma 3 1B2026.04 | 2.6 | 2.4 | 9.4 | 2.3 | 0.4 | 0.8 | 0.6 | 0.6 | 1.2 | 1.5 | 0.6 | 0.9 | 1.9 | |
| PCAModel=Gemma 3 1B2026.04 | 1.8 | 2.3 | 9.8 | 1.3 | 0.4 | 0.2 | 0 | 1.7 | 2.1 | 0.1 | 0.6 | 0.9 | 1.9 | |
| LDAModel=Gemma 3 1B2026.04 | 1.3 | 2.1 | 3.7 | 2.1 | 1.4 | 0.1 | 0 | 0.2 | 0.9 | 0.7 | 0.1 | 1.7 | 1.2 |