Safety Evaluation on HarmBench first 50 behaviors (fixed subset)
0ASRRe-mask only (ablation)
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Re-mask only (ablation)Model=Dream-7B-Instruct, Prefix length (Lg)=642026.03 | 0 | 0 | 1 | |
| TrajHijack (core: re-mask + prefix)Model=LLaDA-8B-Instruct, Prefix length (Lg)=5122026.03 | 52 | 86 | 3.4 | |
| TrajHijack (core: re-mask + prefix)Model=LLaDA-8B-Instruct, Prefix length (Lg)=2562026.03 | 78 | 88 | 4.1 | |
| TrajHijack (core: re-mask + prefix)Model=LLaDA-8B-Instruct, Prefix length (Lg)=1282026.03 | 84 | 90 | 4.3 | |
| TrajHijack (core: re-mask + prefix)Model=Dream-7B-Instruct, Prefix length (Lg)=5122026.03 | 84 | 92 | 4.4 | |
| TrajHijack (core: re-mask + prefix)Model=Dream-7B-Instruct, Prefix length (Lg)=642026.03 | 90 | 98 | 4.6 | |
| TrajHijack (core: re-mask + prefix)Model=Dream-7B-Instruct, Prefix length (Lg)=2562026.03 | 90 | 98 | 4.6 | |
| TrajHijack (core: re-mask + prefix)Model=LLaDA-8B-Instruct, Prefix length (Lg)=642026.03 | 94 | 96 | 4.6 |