ResearchBenchmarksDistributed Misuse Detection on DM-Bio 100x Trace-levelFollow62.2Trace-level ROCAUCMeerkat45.35249.72654.158.474Apr 13, 2026Evaluation ResultsMethodMethodLinksTrace-level ROCAUCDelta (Δ)MeerkatModel=Qwen3.5Model=Qwen3.52026.0462.20.036BufferModel=Qwen3.5Model=Qwen3.52026.0458.5—BayesianModel=GPT-5.4mModel=GPT-5.4m2026.0458.2—BayesianModel=Qwen3.5Model=Qwen3.52026.0457.2—BufferModel=GPT-5.4mModel=GPT-5.4m2026.0452.4—MeerkatModel=GPT-5.4mModel=GPT-5.4m2026.0451.1-0.071Naive AgentModel=GPT-5.4mModel=GPT-5.4m2026.0449.3—MonitorModel=Qwen3.5Model=Qwen3.52026.0446—