List your company

Agentic Oversight on VitaBench

82.13Detection AccuracyFORMALJUDGE
55.81862.64969.4876.311Feb 11, 2026

Evaluation results

MethodLinks
FORMALJUDGEJudge Model=Gemini-1.5-Pro2026.0282.13
FORMALJUDGEJudge Model=GPT-4o2026.0278.53
FORMALJUDGEJudge Model=Qwen-32B2026.0278.53
LLM + PythonJudge Model=Gemini-1.5-Pro2026.0276.23
LLM + Struct.Judge Model=Gemini-1.5-Pro2026.0275.04
FORMALJUDGEJudge Model=Claude-3.5-Sonnet2026.0274.93
LLM + Struct.Judge Model=GPT-4o2026.0274.72
FORMALJUDGEJudge Model=Qwen-14B2026.0274.52
LLMJudge Model=Gemini-1.5-Pro2026.0274.36
LLM + PythonJudge Model=Qwen-32B2026.0273.85
LLM + Few-ShotJudge Model=Gemini-1.5-Pro2026.0273.54
LLM + PythonJudge Model=GPT-4o2026.0273.14
LLMJudge Model=GPT-4o2026.0272.23
LLM + Few-ShotJudge Model=GPT-4o2026.0272.21
LLM + CoTJudge Model=Gemini-1.5-Pro2026.0272.15
FORMALJUDGEJudge Model=Qwen-7B2026.0270.97
LLM + Struct.Judge Model=Qwen-32B2026.0270.82
LLM + PythonJudge Model=Claude-3.5-Sonnet2026.0269.84
LLM + CoTJudge Model=Claude-3.5-Sonnet2026.0268.42
LLM + Few-ShotJudge Model=Qwen-32B2026.0267.93
LLMJudge Model=Qwen-14B2026.0267.82
LLM + PythonJudge Model=Qwen-14B2026.0267.51
LLM + CoTJudge Model=Qwen-32B2026.0267.02
LLM + CoTJudge Model=Qwen-14B2026.0266.13
LLMJudge Model=Qwen-32B2026.0266.05
LLM + PythonJudge Model=Qwen-7B2026.0265.92
LLM + CoTJudge Model=GPT-4o2026.0265.84
LLM + Struct.Judge Model=Qwen-7B2026.0264.52
LLM + Struct.Judge Model=Qwen-14B2026.0264.45
LLM + Few-ShotJudge Model=Qwen-14B2026.0264.43
FORMALJUDGEJudge Model=Qwen-72B2026.0264.24
LLM + CoTJudge Model=Qwen-7B2026.0262.76
LLMJudge Model=Claude-3.5-Sonnet2026.0262.61
LLM + PythonJudge Model=Qwen-72B2026.0262.14
LLM + Struct.Judge Model=Claude-3.5-Sonnet2026.0262.03
LLMJudge Model=Qwen-7B2026.0260.94
LLM + Few-ShotJudge Model=Qwen-7B2026.0260.91
LLMJudge Model=Qwen-72B2026.0259.31
LLM + Struct.Judge Model=Qwen-72B2026.0259.28
LLM + Few-ShotJudge Model=Claude-3.5-Sonnet2026.0258.52
LLM + CoTJudge Model=Qwen-72B2026.0258.04
LLM + Few-ShotJudge Model=Qwen-72B2026.0256.83