Prompt-level Targeted Bit-flip Attack on WebShop
100CDAFlip-S
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Flip-SModel=AgentLM-7B2026.03 | 100 | 79.6 | |
| Flip-AgentModel=AgentLM-7B2026.03 | 100 | 99.2 | |
| Flip-SModel=AgentLM-13B2026.03 | 100 | 61.1 | |
| Flip-AgentModel=Qwen3-4B-Instruct-25072026.03 | 100 | 98 | |
| TBTModel=Llama-3.1-8B-Instruct2026.03 | 97.9 | 16.7 | |
| Flip-AgentModel=Llama-3.2-3B-Instruct2026.03 | 95 | 98.1 | |
| TBTModel=AgentLM-7B2026.03 | 95 | 21.5 | |
| Flip-AgentModel=AgentLM-13B2026.03 | 95 | 94.4 | |
| Flip-AgentModel=DeepSeek-R1-Distill-Llama-8B2026.03 | 95 | 92.6 | |
| Flip-AgentModel=Llama-3.1-8B-Instruct2026.03 | 90 | 97.3 | |
| Flip-SModel=DeepSeek-R1-Distill-Llama-8B2026.03 | 90 | 72.4 | |
| TBTModel=DeepSeek-R1-Distill-Llama-8B2026.03 | 85.4 | 11.1 | |
| TBTModel=AgentLM-13B2026.03 | 85 | 0 | |
| TrojViTModel=AgentLM-7B2026.03 | 80 | 75 | |
| TrojViTModel=AgentLM-13B2026.03 | 80 | 61.2 | |
| Flip-SModel=Qwen3-4B-Instruct-25072026.03 | 80 | 87.5 | |
| TrojViTModel=DeepSeek-R1-Distill-Llama-8B2026.03 | 80 | 63.2 | |
| TBTModel=Llama-3.2-3B-Instruct2026.03 | 70 | 55.6 | |
| TrojViTModel=Qwen3-4B-Instruct-25072026.03 | 70 | 79.4 | |
| TrojViTModel=Llama-3.2-3B-Instruct2026.03 | 65 | 83.5 | |
| Flip-SModel=Llama-3.2-3B-Instruct2026.03 | 60 | 88.9 | |
| TrojViTModel=Llama-3.1-8B-Instruct2026.03 | 60 | 72.8 | |
| TBTModel=Qwen3-4B-Instruct-25072026.03 | 50 | 22.2 | |
| Flip-SModel=Llama-3.1-8B-Instruct2026.03 | 20 | 81.5 |