OneBench
Model Confidence Under Answer-Preserving Attacks: An Informativeness-Manipulability Frontier | OneBench: AI Insights