LLM Steering on Anthropic Model-Written Evaluations (MWE) In-distribution
2.363SteerabilityL2S w/ W2S
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| L2S w/ W2SBackbone=Llama-2-7B-chat, Layer Selection Strategy=W2S2026.04 | 2.363 | 91.8 | |
| L2SBackbone=Llama-2-7B-chat, Layer Selection Strategy=Fixed Layer2026.04 | 2.098 | 89.9 | |
| L2S w/ W2SBackbone=Qwen-1.5-14B-Chat, Layer Selection Strategy=W2S2026.04 | 2.071 | 91.8 | |
| L2SBackbone=Qwen-1.5-14B-Chat, Layer Selection Strategy=Fixed Layer2026.04 | 1.888 | 87.5 | |
| CAA w/ W2SBackbone=Qwen-1.5-14B-Chat, Layer Selection Strategy=W2S2026.04 | 1.675 | 85.4 | |
| CAA w/ W2SBackbone=Llama-2-7B-chat, Layer Selection Strategy=W2S2026.04 | 1.502 | 84.6 | |
| CAABackbone=Qwen-1.5-14B-Chat, Layer Selection Strategy=Fixed Layer2026.04 | 1.493 | 83.3 | |
| CAABackbone=Llama-2-7B-chat, Layer Selection Strategy=Fixed Layer2026.04 | 1.259 | 75.4 |