Concept Alignment on CAA (test)
75.18AICSelf Modulation
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| Self ModulationTarget LLM (mt)=Llama2, Source LLM (ms)=Llama2, Evaluation Protocol=Output Probabilities2025.01 | 75.18 | 91.2 | 89.61 | 73.27 | 60 | 70.08 | 88.71 | |
| L-Cross ModulationTarget LLM (mt)=Llama2, Source LLM (ms)=Llama3.1, Evaluation Protocol=Output Probabilities2025.01 | 74.95 | 91.41 | 89.75 | 73.62 | 59.97 | 70.2 | 88.48 | |
| L-Cross ModulationTarget LLM (mt)=Llama2, Source LLM (ms)=Qwen2, Evaluation Protocol=Output Probabilities2025.01 | 73.11 | 91.36 | 89.75 | 73.04 | 59.95 | 70.03 | 88.57 | |
| No ModulationTarget LLM (mt)=Llama2, Source LLM (ms)=None, Evaluation Protocol=Output Probabilities2025.01 | 30.06 | 63.8 | 81.41 | 74.64 | 33.86 | 69.18 | 74.24 | |
| Self ModulationTarget LLM (mt)=Qwen2, Source LLM (ms)=Qwen2, Evaluation Protocol=Output Probabilities2025.01 | 11.35 | 74.09 | 62.2 | 67.56 | 62.61 | 73.71 | 94.35 | |
| L-Cross ModulationTarget LLM (mt)=Qwen2, Source LLM (ms)=Llama3.1, Evaluation Protocol=Output Probabilities2025.01 | 11.29 | 75.15 | 61.95 | 67.66 | 62.6 | 74.16 | 93.99 | |
| L-Cross ModulationTarget LLM (mt)=Qwen2, Source LLM (ms)=Llama2, Evaluation Protocol=Output Probabilities2025.01 | 11.27 | 75.1 | 62.17 | 66.58 | 62.57 | 74 | 94.58 | |
| No ModulationTarget LLM (mt)=Qwen2, Source LLM (ms)=None, Evaluation Protocol=Output Probabilities2025.01 | 9.44 | 54.31 | 52.11 | 49.03 | 57.84 | 72.81 | 92.18 | |
| Self ModulationTarget LLM (mt)=Llama3.1, Source LLM (ms)=Llama3.1, Evaluation Protocol=GPT-Scoring2025.01 | 1.38 | 6.56 | 3.84 | 7 | 7.28 | 3.48 | 4.74 | |
| L-Cross ModulationTarget LLM (mt)=Llama3.1, Source LLM (ms)=Qwen2, Evaluation Protocol=GPT-Scoring2025.01 | 1.24 | 6.36 | 3.78 | 6.86 | 7.22 | 3.34 | 5.4 | |
| No ModulationTarget LLM (mt)=Llama3.1, Source LLM (ms)=None, Evaluation Protocol=GPT-Scoring2025.01 | 1.14 | 6.2 | 3.04 | 3.64 | 6.74 | 3.54 | 4.92 | |
| L-Cross ModulationTarget LLM (mt)=Llama3.1, Source LLM (ms)=Llama2, Evaluation Protocol=GPT-Scoring2025.01 | 1.14 | 6.74 | 4.18 | 6.29 | 7 | 3.58 | 4.62 |