Effective Rank Analysis on Ds AdvBench-style harmful-behavior requests v4 (held-out prompts)
12Rank EpsilonLlama-3.1-8B-Instruct
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Llama-3.1-8B-Instructd=4096, M_variant=Mtemplate, epsilon=0.05, n=200, Aggregation=mean across hidden layers2026.05 | 12 | 0.0029 | 0.06 | |
| Qwen-2.5-7B-Instructd=3584, M_variant=Mtemplate, epsilon=0.05, n=200, Aggregation=mean across hidden layers2026.05 | 15.7 | 0.0044 | 0.078 | |
| Gemma-2-9B-itd=3584, M_variant=Mtemplate, epsilon=0.05, n=200, Aggregation=mean across hidden layers2026.05 | 17.4 | 0.0048 | 0.096 |