LLM Alignment on Harmlessness
87.85WRAlignX
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| AlignXBackbone=DeepSeek-7B2026.02 | 87.85 | 33.15 | 80.65 | 45.12 | |
| AMBSBackbone=DeepSeek-7B, Throughput (Tk)=225, Memory Usage (Me)=20, Compute Cost (GP)=42025.09 | 86.2 | 16.2 | 83.8 | 51.2 | |
| AlignXBackbone=Mistral-7B2026.02 | 86.05 | 33.9 | 78.9 | 43.68 | |
| AlignXBackbone=Gemma-7B2026.02 | 83.18 | 34.6 | 76.05 | 41.54 | |
| TrinityXBackbone=LLaMA-2-7B2026.02 | 81.5 | 23.1 | 80.17 | 46.19 | |
| AlignXBackbone=LLaMA-2-7B2026.02 | 80.2 | 23.25 | 76.85 | 44.6 | |
| AMBSBackbone=Mistral-7B, Throughput (Tk)=212, Memory Usage (Me)=20, Compute Cost (GP)=62025.09 | 64.4 | 15.5 | 25.5 | 24.8 | |
| H3Fusion2026.02 | 59.86 | 33 | 32.03 | 19.63 | |
| AMBSBackbone=Gemma-7B, Throughput (Tk)=218, Memory Usage (Me)=20, Compute Cost (GP)=52025.09 | 51.4 | 12.3 | 19.2 | 19.4 | |
| AMBSBackbone=LLaMA-2-7B, Throughput (Tk)=213, Memory Usage (Me)=23, Compute Cost (GP)=72025.09 | 49.3 | 5.3 | 33.5 | 25.8 | |
| SteeringBackbone=Mistral-7B, Throughput (Tk)=211, Memory Usage (Me)=23, Compute Cost (GP)=72025.09 | 41.1 | 18.6 | 23.9 | 15.5 | |
| AlignerBackbone=Mistral-7B, Throughput (Tk)=178, Memory Usage (Me)=27, Compute Cost (GP)=82025.09 | 40.6 | 16.8 | 23.1 | 15.6 | |
| DExpertsBackbone=Mistral-7B, Throughput (Tk)=171, Memory Usage (Me)=28, Compute Cost (GP)=102025.09 | 39.5 | 17.9 | 23.4 | 15 | |
| PPLMBackbone=Mistral-7B, Throughput (Tk)=147, Memory Usage (Me)=30, Compute Cost (GP)=112025.09 | 37.6 | 20.4 | 22.1 | 13.1 | |
| SteeringBackbone=Gemma-7B, Throughput (Tk)=217, Memory Usage (Me)=22, Compute Cost (GP)=72025.09 | 36.6 | 16.5 | 17.7 | 12.6 | |
| AlignerBackbone=Gemma-7B, Throughput (Tk)=184, Memory Usage (Me)=26, Compute Cost (GP)=72025.09 | 36.1 | 14.8 | 17.4 | 12.9 | |
| DExpertsBackbone=Gemma-7B, Throughput (Tk)=177, Memory Usage (Me)=27, Compute Cost (GP)=92025.09 | 35 | 15.8 | 17.1 | 12.1 | |
| SteeringBackbone=LLaMA-2-7B, Throughput (Tk)=205, Memory Usage (Me)=24, Compute Cost (GP)=82025.09 | 33.8 | 20.6 | 21.8 | 11.7 | |
| PPLMBackbone=Gemma-7B, Throughput (Tk)=153, Memory Usage (Me)=29, Compute Cost (GP)=102025.09 | 33.3 | 18.3 | 16.4 | 10.5 | |
| DExpertsBackbone=LLaMA-2-7B, Throughput (Tk)=166, Memory Usage (Me)=29, Compute Cost (GP)=112025.09 | 32.9 | 19.7 | 21.1 | 11.4 | |
| SteeringBackbone=DeepSeek-7B, Throughput (Tk)=223, Memory Usage (Me)=21, Compute Cost (GP)=72025.09 | 32.4 | 19.5 | 42.7 | 18.5 | |
| AlignerBackbone=DeepSeek-7B, Throughput (Tk)=190, Memory Usage (Me)=25, Compute Cost (GP)=72025.09 | 31.9 | 18 | 42.1 | 18.7 | |
| DExpertsBackbone=DeepSeek-7B, Throughput (Tk)=183, Memory Usage (Me)=26, Compute Cost (GP)=92025.09 | 30.9 | 18.9 | 41.3 | 17.8 | |
| PPLMBackbone=LLaMA-2-7B, Throughput (Tk)=141, Memory Usage (Me)=31, Compute Cost (GP)=122025.09 | 30.7 | 22.5 | 19.6 | 9.3 | |
| PPLMBackbone=DeepSeek-7B, Throughput (Tk)=159, Memory Usage (Me)=28, Compute Cost (GP)=102025.09 | 29.7 | 21.2 | 40.5 | 16.3 | |
| Aligner2026.02 | 25.4 | 7.2 | — | 6.06 | |
| AlignerBackbone=LLaMA-2-7B, Throughput (Tk)=172, Memory Usage (Me)=28, Compute Cost (GP)=92025.09 | 25.4 | 7.2 | 20.9 | 13 |