List your company

Agentic Oversight on Agent-SafetyBench

84.06Detection AccuracyFORMALJUDGE
50.197658.988867.7876.5712Feb 11, 2026

Evaluation results

MethodLinks
FORMALJUDGEJudge Model=Claude-3.5-Sonnet2026.0284.06
FORMALJUDGEJudge Model=Gemini-1.5-Pro2026.0280.54
FORMALJUDGEJudge Model=GPT-4o2026.0277.28
LLM + Struct.Judge Model=Gemini-1.5-Pro2026.0276.6
FORMALJUDGEJudge Model=Qwen-7B2026.0274.5
FORMALJUDGEJudge Model=Qwen-72B2026.0273.26
LLM + Few-ShotJudge Model=Claude-3.5-Sonnet2026.0272.73
LLMJudge Model=Claude-3.5-Sonnet2026.0272.22
LLM + PythonJudge Model=Qwen-7B2026.0272
FORMALJUDGEJudge Model=Qwen-32B2026.0271.79
FORMALJUDGEJudge Model=Qwen-14B2026.0271.77
LLM + CoTJudge Model=Claude-3.5-Sonnet2026.0269.7
LLM + PythonJudge Model=GPT-4o2026.0268.7
LLM + Struct.Judge Model=Claude-3.5-Sonnet2026.0268.69
LLM + PythonJudge Model=Qwen-32B2026.0268.18
LLM + PythonJudge Model=Gemini-1.5-Pro2026.0267.38
LLM + CoTJudge Model=Gemini-1.5-Pro2026.0267.33
LLM + PythonJudge Model=Qwen-14B2026.0266.67
LLM + Struct.Judge Model=Qwen-14B2026.0266.16
LLM + Few-ShotJudge Model=Gemini-1.5-Pro2026.0265.63
LLM + PythonJudge Model=Qwen-72B2026.0263
LLM + Struct.Judge Model=GPT-4o2026.0262.3
LLM + Struct.Judge Model=Qwen-7B2026.0262
LLMJudge Model=Qwen-7B2026.0261.98
LLM + Few-ShotJudge Model=Qwen-7B2026.0261.98
LLM + Struct.Judge Model=Qwen-32B2026.0261.62
LLM + CoTJudge Model=GPT-4o2026.0261.5
LLMJudge Model=Qwen-14B2026.0260.1
LLM + Struct.Judge Model=Qwen-72B2026.0259.5
LLM + Few-ShotJudge Model=GPT-4o2026.0259.1
LLM + PythonJudge Model=Claude-3.5-Sonnet2026.0258.59
LLMJudge Model=GPT-4o2026.0258
LLMJudge Model=Gemini-1.5-Pro2026.0258
LLMJudge Model=Qwen-72B2026.0256.57
LLM + Few-ShotJudge Model=Qwen-14B2026.0255.56
LLM + CoTJudge Model=Qwen-7B2026.0255
LLM + CoTJudge Model=Qwen-14B2026.0253.54
LLMJudge Model=Qwen-32B2026.0252.53
LLM + CoTJudge Model=Qwen-32B2026.0252.53
LLM + CoTJudge Model=Qwen-72B2026.0252.5
LLM + Few-ShotJudge Model=Qwen-32B2026.0252.02
LLM + Few-ShotJudge Model=Qwen-72B2026.0251.5