Misalignment detection on Taylor
99AccuracyTELLME
Evaluation Results
| Method | Links | |
|---|---|---|
| TELLMEBackbone=Mistral-7B-v0.3, Evaluation Protocol=Self-Sim2025.02 | 99 | |
| TELLMEBackbone=Mistral-7B-v0.3, Evaluation Protocol=Linear Probe2025.02 | 99 | |
| TELLMEBackbone=Llama-3.1-8B, Evaluation Protocol=Linear Probe2025.02 | 96.3 | |
| TELLMEBackbone=Qwen2.5-7B, Evaluation Protocol=Linear Probe2025.02 | 96.3 | |
| SAE-Multi-TopKBackbone=Llama-3.1-8B, Evaluation Protocol=Linear Probe2025.02 | 94.3 | |
| SAE-KLBackbone=Llama-3.1-8B, Evaluation Protocol=Linear Probe2025.02 | 94 | |
| SAEBackbone=Llama-3.1-8B, Evaluation Protocol=Linear Probe2025.02 | 93.5 | |
| OriginBackbone=Llama-3.1-8B, Evaluation Protocol=Linear Probe2025.02 | 92.8 | |
| SAE-Multi-TopKBackbone=Qwen2.5-7B, Evaluation Protocol=Linear Probe2025.02 | 92.8 | |
| SAEBackbone=Qwen2.5-7B, Evaluation Protocol=Linear Probe2025.02 | 91.5 | |
| SAE-Multi-TopKBackbone=Mistral-7B-v0.3, Evaluation Protocol=Linear Probe2025.02 | 91.3 | |
| TELLMEBackbone=Qwen2.5-7B, Evaluation Protocol=Self-Sim2025.02 | 91.3 | |
| SAEBackbone=Mistral-7B-v0.3, Evaluation Protocol=Linear Probe2025.02 | 91 | |
| SAE-KLBackbone=Mistral-7B-v0.3, Evaluation Protocol=Linear Probe2025.02 | 90.8 | |
| Skip TranscoderBackbone=Mistral-7B-v0.3, Evaluation Protocol=Linear Probe2025.02 | 90.8 | |
| SAE-KLBackbone=Qwen2.5-7B, Evaluation Protocol=Linear Probe2025.02 | 90.8 | |
| TELLMEBackbone=Llama-3.1-8B, Evaluation Protocol=Self-Sim2025.02 | 89 | |
| Skip TranscoderBackbone=Qwen2.5-7B, Evaluation Protocol=Linear Probe2025.02 | 88.8 | |
| TELLMEBackbone=Llama-3.1-8B, Evaluation Protocol=Judge2025.02 | 88.5 | |
| OriginBackbone=Llama-3.1-8B, Evaluation Protocol=Judge2025.02 | 87.8 | |
| SAEBackbone=Llama-3.1-8B, Evaluation Protocol=Judge2025.02 | 86.2 | |
| SAE-KLBackbone=Llama-3.1-8B, Evaluation Protocol=Judge2025.02 | 86 | |
| OriginBackbone=Mistral-7B-v0.3, Evaluation Protocol=Linear Probe2025.02 | 86 | |
| SAE-Multi-TopKBackbone=Llama-3.1-8B, Evaluation Protocol=Judge2025.02 | 85.8 | |
| Skip TranscoderBackbone=Llama-3.1-8B, Evaluation Protocol=Judge2025.02 | 85.5 | |
| TranscoderBackbone=Llama-3.1-8B, Evaluation Protocol=Judge2025.02 | 85.2 | |
| Skip TranscoderBackbone=Llama-3.1-8B, Evaluation Protocol=Linear Probe2025.02 | 84.5 | |
| TranscoderBackbone=Llama-3.1-8B, Evaluation Protocol=Linear Probe2025.02 | 83.5 | |
| OriginBackbone=Qwen2.5-7B, Evaluation Protocol=Linear Probe2025.02 | 83.3 | |
| TELLMEBackbone=Qwen2.5-7B, Evaluation Protocol=Judge2025.02 | 82 | |
| OriginBackbone=Qwen2.5-7B, Evaluation Protocol=Self-Sim2025.02 | 80.5 | |
| OriginBackbone=Qwen2.5-7B, Evaluation Protocol=Judge2025.02 | 80.5 | |
| SAEBackbone=Qwen2.5-7B, Evaluation Protocol=Self-Sim2025.02 | 80.3 | |
| SAE-Multi-TopKBackbone=Qwen2.5-7B, Evaluation Protocol=Judge2025.02 | 80 | |
| SAE-KLBackbone=Qwen2.5-7B, Evaluation Protocol=Self-Sim2025.02 | 79.8 | |
| SAE-Multi-TopKBackbone=Qwen2.5-7B, Evaluation Protocol=Self-Sim2025.02 | 79.8 | |
| Skip TranscoderBackbone=Qwen2.5-7B, Evaluation Protocol=Self-Sim2025.02 | 78.8 | |
| TranscoderBackbone=Qwen2.5-7B, Evaluation Protocol=Self-Sim2025.02 | 78.5 | |
| Skip TranscoderBackbone=Qwen2.5-7B, Evaluation Protocol=Judge2025.02 | 78.2 | |
| SAE-KLBackbone=Qwen2.5-7B, Evaluation Protocol=Judge2025.02 | 78 | |
| SAEBackbone=Qwen2.5-7B, Evaluation Protocol=Judge2025.02 | 77.8 | |
| TranscoderBackbone=Qwen2.5-7B, Evaluation Protocol=Judge2025.02 | 77.8 | |
| Skip TranscoderBackbone=Llama-3.1-8B, Evaluation Protocol=Self-Sim2025.02 | 76.8 | |
| OriginBackbone=Llama-3.1-8B, Evaluation Protocol=Self-Sim2025.02 | 76.5 | |
| SAEBackbone=Llama-3.1-8B, Evaluation Protocol=Self-Sim2025.02 | 76.3 | |
| SAE-KLBackbone=Llama-3.1-8B, Evaluation Protocol=Self-Sim2025.02 | 76.3 | |
| TranscoderBackbone=Llama-3.1-8B, Evaluation Protocol=Self-Sim2025.02 | 76.3 | |
| TranscoderBackbone=Mistral-7B-v0.3, Evaluation Protocol=Linear Probe2025.02 | 75.8 | |
| SAE-Multi-TopKBackbone=Llama-3.1-8B, Evaluation Protocol=Self-Sim2025.02 | 75.5 | |
| TranscoderBackbone=Qwen2.5-7B, Evaluation Protocol=Linear Probe2025.02 | 75.5 | |
| TranscoderBackbone=Mistral-7B-v0.3, Evaluation Protocol=Self-Sim2025.02 | 69.3 | |
| Skip TranscoderBackbone=Mistral-7B-v0.3, Evaluation Protocol=Self-Sim2025.02 | 69.3 | |
| SAEBackbone=Mistral-7B-v0.3, Evaluation Protocol=Self-Sim2025.02 | 68.8 | |
| SAE-KLBackbone=Mistral-7B-v0.3, Evaluation Protocol=Self-Sim2025.02 | 68.8 | |
| SAE-Multi-TopKBackbone=Mistral-7B-v0.3, Evaluation Protocol=Self-Sim2025.02 | 68.8 | |
| OriginBackbone=Mistral-7B-v0.3, Evaluation Protocol=Self-Sim2025.02 | 67.5 | |
| TELLMEBackbone=Mistral-7B-v0.3, Evaluation Protocol=Judge2025.02 | 63 | |
| SAEBackbone=Mistral-7B-v0.3, Evaluation Protocol=Judge2025.02 | 57.8 | |
| SAE-KLBackbone=Mistral-7B-v0.3, Evaluation Protocol=Judge2025.02 | 57.5 | |
| SAE-Multi-TopKBackbone=Mistral-7B-v0.3, Evaluation Protocol=Judge2025.02 | 57 | |
| OriginBackbone=Mistral-7B-v0.3, Evaluation Protocol=Judge2025.02 | 56.3 | |
| Skip TranscoderBackbone=Mistral-7B-v0.3, Evaluation Protocol=Judge2025.02 | 54.8 | |
| TranscoderBackbone=Mistral-7B-v0.3, Evaluation Protocol=Judge2025.02 | 54.5 |