ResearchBenchmarksRule Extraction Functional Fidelity on LLM Injected Behaviors Non-judge output abstractionsFollow0.83MCCRuleSHAP0.74680.76840.790.8116May 16, 2025Evaluation ResultsMethodMethodLinksMCCRuleSHAPTop-k=10Top-k=102025.050.83RuleFitTop-k=10Top-k=102025.050.82RuleSHAPTop-k=3Top-k=32025.050.8RuleFitTop-k=3Top-k=32025.050.78RuleSHAPTop-k=1Top-k=12025.050.76RuleFitTop-k=1Top-k=12025.050.75