Prompt-level GCG attack on JailbreakBench (held-out prompts)
69ASRGCG-GPR
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| GCG-GPRModel=Qwen2.5-3B, Gradient source=GPR (predicted)2026.05 | 69 | 0.3 | 0.24 | |
| GCGModel=Qwen2.5-3B, Gradient source=Exact gradient2026.05 | 60 | 0.2 | 0.1 | |
| GCGModel=Llama-2-7B, Gradient source=Exact gradient2026.05 | 21 | 0.1 | 0.02 | |
| GCG-GPRModel=Llama-2-7B, Gradient source=GPR (predicted)2026.05 | 21 | 0.2 | 0.04 | |
| GCGModel=Qwen3-4B, Gradient source=Exact gradient2026.05 | 0 | 0.1 | 0 | |
| GCG-GPRModel=Qwen3-4B, Gradient source=GPR (predicted)2026.05 | 0 | 0.3 | 0 |