Exact recovery of injected triggers on Llama-3.1 70B
2,462Trigger CountRuleFit+XGB
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| RuleFit+XGBLLM=Llama 3.1 70B2025.05 | 2,462 | 14 | 14 | 14 | |
| RuleSHAP w/o Step 2LLM=Llama 3.1 70B2025.05 | 1,761 | 14 | 14 | 14 | |
| RuleSHAP w/o Step 3LLM=Llama 3.1 70B2025.05 | 1,715 | 57 | 60 | 60 | |
| RuleFitLLM=Llama 3.1 70B2025.05 | 1,422 | 21 | 25 | 26 | |
| RuleSHAPLLM=Llama 3.1 70B2025.05 | 1,360 | 57 | 60 | 60 | |
| Decision TreeLLM=Llama 3.1 70B2025.05 | 648 | 0 | 0 | 0 |