Prompt-level GCG Attack on HarmBench (held-out prompts)
50.9ASRGCG-GPR
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| GCG-GPRModel=Qwen2.5-3B, Gradient source=GPR (predicted)2026.05 | 50.9 | 0.3 | 0.18 | |
| GCGModel=Qwen2.5-3B, Gradient source=Exact gradient2026.05 | 49.4 | 0.2 | 0.08 | |
| GCG-GPRModel=Llama-2-7B, Gradient source=GPR (predicted)2026.05 | 19.4 | 0.2 | 0.04 | |
| GCGModel=Llama-2-7B, Gradient source=Exact gradient2026.05 | 17.2 | 0.1 | 0.02 | |
| GCGModel=Qwen3-4B, Gradient source=Exact gradient2026.05 | 5 | 0.1 | 0.01 | |
| GCG-GPRModel=Qwen3-4B, Gradient source=GPR (predicted)2026.05 | 4.4 | 0.3 | 0.01 |