Instruction Following on HarmBench Clean
99Clean RateBase (8B Instruct)
Evaluation Results
| Method | Links | |
|---|---|---|
| Base (8B Instruct)Backbone=Llama 32026.05 | 99 | |
| LPA (ours)Backbone=Llama 32026.05 | 99 | |
| L3-LATBackbone=Llama 32026.05 | 92 | |
| L3-CATBackbone=Llama 32026.05 | 92 | |
| LPA-overfit (ours)Backbone=Llama 32026.05 | 88 | |
| LPA (ours)Backbone=Llama 22026.05 | 79 | |
| Base (7B chat-hf)Backbone=Llama 22026.05 | 78 | |
| L2-CATBackbone=Llama 22026.05 | 67 | |
| LPA-overfit (ours)Backbone=Llama 22026.05 | 60 | |
| L2-LATBackbone=Llama 22026.05 | 13 |