Reasoning on GSM8K (ASR and CACC)
83.32CACCSFT+RL
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| SFT+RLBackbone=Llama3.1-8B, Poison Ratio=0.052026.04 | 83.32 | 85.22 | |
| MirageBDBackbone=Llama3.1-8B, Poison Ratio=0.052026.04 | 81.58 | 92.65 | |
| BenignBackbone=Llama3.2-3B, Poison Ratio=0.052026.04 | 81.35 | — | |
| DecepChainBackbone=Llama3.1-8B, Poison Ratio=0.052026.04 | 80.74 | 80.69 | |
| MirageBDBackbone=Llama3.2-3B, Poison Ratio=0.052026.04 | 79.83 | 93.69 | |
| DecepChainBackbone=Llama3.2-3B, Poison Ratio=0.052026.04 | 78.47 | 71.27 | |
| BadChainBackbone=Llama3.2-3B, Poison Ratio=0.052026.04 | 77.94 | 0.32 | |
| SFT+RLBackbone=Qwen2.5-7B, Poison Ratio=0.052026.04 | 77.03 | 90.44 | |
| SFT+RLBackbone=Llama3.2-3B, Poison Ratio=0.052026.04 | 76.35 | 86.68 | |
| MirageBDBackbone=Qwen2.5-7B, Poison Ratio=0.052026.04 | 73.92 | 93.23 | |
| MirageBDBackbone=Qwen2.5-3B, Poison Ratio=0.052026.04 | 73.77 | 88.33 | |
| BenignBackbone=Qwen2.5-7B, Poison Ratio=0.052026.04 | 73.16 | — | |
| BadChainBackbone=Qwen2.5-3B, Poison Ratio=0.052026.04 | 72.48 | 0.36 | |
| BadChainBackbone=Qwen2.5-7B, Poison Ratio=0.052026.04 | 72.26 | 0 | |
| DecepChainBackbone=Qwen2.5-7B, Poison Ratio=0.052026.04 | 71.8 | 77.3 | |
| DecepChainBackbone=Qwen2.5-3B, Poison Ratio=0.052026.04 | 71.04 | 57.68 | |
| BenignBackbone=Llama3.1-8B, Poison Ratio=0.052026.04 | 70.51 | — | |
| SFT+RLBackbone=Qwen2.5-3B, Poison Ratio=0.052026.04 | 69.83 | 61.37 | |
| BadChainBackbone=Qwen2.5-1.5B, Poison Ratio=0.052026.04 | 67.78 | 0.5 | |
| DecepChainBackbone=Qwen2.5-1.5B, Poison Ratio=0.052026.04 | 66.64 | 2.24 | |
| MirageBDBackbone=Qwen2.5-1.5B, Poison Ratio=0.052026.04 | 66.64 | 93.69 | |
| BenignBackbone=Qwen2.5-3B, Poison Ratio=0.052026.04 | 66.57 | — | |
| BenignBackbone=Qwen2.5-1.5B, Poison Ratio=0.052026.04 | 66.11 | — | |
| SFT+RLBackbone=Qwen2.5-1.5B, Poison Ratio=0.052026.04 | 55.04 | 66.77 | |
| BadChainBackbone=Llama3.1-8B, Poison Ratio=0.052026.04 | 54.13 | 0 |