Truthfulness Steering on TruthfulQA
90.8T×I ScoreUNISTEER
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| UNISTEERTarget LLM=Qwen2.5-7B2026.05 | 90.8 | — | — | — | |
| ODESteerTarget LLM=Qwen2.5-7B2026.05 | 87.71 | — | — | — | |
| OriginalTarget LLM=Qwen2.5-7B2026.05 | 85.91 | — | — | — | |
| LoReFTTarget LLM=Qwen2.5-7B2026.05 | 83.96 | — | — | — | |
| RepETarget LLM=Qwen2.5-7B2026.05 | 81.99 | — | — | — | |
| CAATarget LLM=Qwen2.5-7B2026.05 | 81.53 | — | — | — | |
| Dense-optBase Model=LLaMA, Type=Gradient2026.05 | 79.28 | 0.0013 | 0.0009 | 0.016 | |
| Disco-QBase Model=LLaMA, Type=Static2026.05 | 78.66 | 0.0823 | 0.0745 | 0.2277 | |
| CAABase Model=LLaMA, Type=Static2026.05 | 75.72 | 0.0637 | 0.0497 | 0.2392 | |
| ITIBase Model=LLaMA, Type=Static2026.05 | 74.23 | 0.0332 | 0.0195 | 0.1983 | |
| Dense-optBase Model=Qwen, Type=Gradient2026.05 | 74.07 | 0.0024 | 0.0007 | 0.0285 | |
| LoReFTTarget LLM=Qwen2.5-1.5B2026.05 | 73.9 | — | — | — | |
| LLaMABase Model=LLaMA2026.05 | 73.66 | — | — | — | |
| ODESteerTarget LLM=Llama-3.2-1B2026.05 | 72.5 | — | — | — | |
| UNISTEERTarget LLM=Qwen2.5-1.5B2026.05 | 71.37 | — | — | — | |
| Disco-QBase Model=Qwen, Type=Static2026.05 | 70.78 | 0.0308 | 0.0271 | 0.1319 | |
| LoReFTTarget LLM=Llama-3.2-1B2026.05 | 70.08 | — | — | — | |
| CAABase Model=Qwen, Type=Static2026.05 | 69.14 | 0.1075 | 0.1022 | 0.2788 | |
| UNISTEERTarget LLM=Llama-3.2-1B2026.05 | 68.01 | — | — | — | |
| ITIBase Model=Qwen, Type=Static2026.05 | 67.78 | 0.0098 | 0.0068 | 0.0829 | |
| QwenBase Model=Qwen2026.05 | 67.49 | — | — | — | |
| CAATarget LLM=Llama-3.2-1B2026.05 | 65.27 | — | — | — | |
| OriginalTarget LLM=Llama-3.2-1B2026.05 | 63.95 | — | — | — | |
| ODESteerTarget LLM=Qwen2.5-1.5B2026.05 | 63.89 | — | — | — | |
| RepETarget LLM=Llama-3.2-1B2026.05 | 63.8 | — | — | — | |
| CAATarget LLM=Qwen2.5-1.5B2026.05 | 56.73 | — | — | — | |
| RepETarget LLM=Qwen2.5-1.5B2026.05 | 54.99 | — | — | — | |
| OriginalTarget LLM=Qwen2.5-1.5B2026.05 | 54.4 | — | — | — |