List your company

Multi-agent discussion attack on MMMU

2.3Delta AccuracyMCA
0.5812.1923.835.41Mar 22, 2026

Evaluation results

MethodLinks
MCASelection Strategy=Original strategy, Anomaly detector=RAGAs2026.032.33.64
MCASelection Strategy=Original strategy, Anomaly detector=G-Safeguard2026.033.13.85.1
MCASelection Strategy=Original strategy, Anomaly detector=SelfCheckGPT2026.034.153.5
MAD-SpearSelection Strategy=Original strategy, Anomaly detector=RAGAs2026.035.776.2
MAD-SpearSelection Strategy=Original strategy, Anomaly detector=G-Safeguard2026.037.17.27.1
MAD-SpearSelection Strategy=Original strategy, Anomaly detector=SelfCheckGPT2026.037.68.55.5
MCASelection Strategy=Variant I, Anomaly detector=RAGAs2026.039.213.54.1
MCASelection Strategy=Variant II, Anomaly detector=RAGAs2026.039.812.94.4
MCASelection Strategy=Variant II, Anomaly detector=G-Safeguard2026.0310.114.95.5
MCASelection Strategy=Variant I, Anomaly detector=G-Safeguard2026.0310.914.55.1
MCASelection Strategy=Variant II, Anomaly detector=SelfCheckGPT2026.0311.315.23.8
MCASelection Strategy=Variant I, Anomaly detector=SelfCheckGPT2026.0311.414.83.6
MCASelection Strategy=Variant III, Anomaly detector=RAGAs2026.0312.716.34
MCASelection Strategy=Variant IV, Anomaly detector=RAGAs2026.0312.717.84.3
MCASelection Strategy=Variant IV, Anomaly detector=G-Safeguard2026.031314.25.3
MCASelection Strategy=Variant IV, Anomaly detector=SelfCheckGPT2026.0313.5183.7
MCASelection Strategy=Variant V, Anomaly detector=RAGAs2026.0313.917.54
MCASelection Strategy=Variant VI, Anomaly detector=RAGAs2026.0314.218.24
MCASelection Strategy=Variant V, Anomaly detector=G-Safeguard2026.0314.718.55.2
MAD-SpearSelection Strategy=Variant II, Anomaly detector=RAGAs2026.0315.119.46.5
MCASelection Strategy=Variant VI, Anomaly detector=G-Safeguard2026.0315.319.85.1
MCASelection Strategy=Variant V, Anomaly detector=SelfCheckGPT2026.0315.419.43.6
MCASelection Strategy=Variant III, Anomaly detector=G-Safeguard2026.0315.519.85.2
MCASelection Strategy=Variant VI, Anomaly detector=SelfCheckGPT2026.0315.720.83.6
MCASelection Strategy=Variant III, Anomaly detector=SelfCheckGPT2026.0316.622.93.5
MAD-SpearSelection Strategy=Variant II, Anomaly detector=G-Safeguard2026.0316.921.37.4
MAD-SpearSelection Strategy=Variant IV, Anomaly detector=RAGAs2026.031722.66.5
MAD-SpearSelection Strategy=Variant II, Anomaly detector=SelfCheckGPT2026.0317.720.45.8
MAD-SpearSelection Strategy=Variant I, Anomaly detector=RAGAs2026.0317.818.46.3
MAD-SpearSelection Strategy=Variant I, Anomaly detector=SelfCheckGPT2026.031817.65.6
MAD-SpearSelection Strategy=Variant I, Anomaly detector=G-Safeguard2026.0318.319.57.1
MAD-SpearSelection Strategy=Variant IV, Anomaly detector=G-Safeguard2026.0318.4237.3
MAD-SpearSelection Strategy=Variant IV, Anomaly detector=SelfCheckGPT2026.0318.924.85.7
MAD-SpearSelection Strategy=Variant V, Anomaly detector=RAGAs2026.0319.325.86.2
MAD-SpearSelection Strategy=Variant III, Anomaly detector=RAGAs2026.0320.225.36.3
MAD-SpearSelection Strategy=Variant VI, Anomaly detector=RAGAs2026.0320.523.46.2
MAD-SpearSelection Strategy=Variant V, Anomaly detector=G-Safeguard2026.0321.427.57.2
MAD-SpearSelection Strategy=Variant VI, Anomaly detector=G-Safeguard2026.0321.526.87.2
MAD-SpearSelection Strategy=Variant III, Anomaly detector=G-Safeguard2026.0321.8267.2
MAD-SpearSelection Strategy=Variant VI, Anomaly detector=SelfCheckGPT2026.0321.828.65.6
MAD-SpearSelection Strategy=Variant III, Anomaly detector=SelfCheckGPT2026.0321.927.75.5
MAD-SpearSelection Strategy=Variant V, Anomaly detector=SelfCheckGPT2026.0322.529.35.5
Formulation-driven Attack StrategyBase Framework=MCA Amayuelas et al. (2024), Anomaly detector=RAGAs2026.0328.1323.9
Formulation-driven Attack StrategyBase Framework=MCA Amayuelas et al. (2024), Anomaly detector=G-Safeguard2026.0331.438.45
Formulation-driven Attack StrategyBase Framework=MCA Amayuelas et al. (2024), Anomaly detector=SelfCheckGPT2026.0332.639.53.4
Formulation-driven Attack StrategyBase Framework=MAD-Spear Cui & Du (2025), Anomaly detector=RAGAs2026.0341.744.66.1
Formulation-driven Attack StrategyBase Framework=MAD-Spear Cui & Du (2025), Anomaly detector=G-Safeguard2026.0342.249.47
Formulation-driven Attack StrategyBase Framework=MAD-Spear Cui & Du (2025), Anomaly detector=SelfCheckGPT2026.0345.351.15.4