RevengeBench probes whether models can reverse-engineer code-space policies from behavior alone

arXiv·medium signal

A new arXiv benchmark (2606.26094) frames policy understanding as an inverse problem: can a model infer the hidden code/rules governing an agent purely from observing its outward actions across behavioral experiments? This matters for agent auditing and interpretability, where operators often must reconstruct what a black-box agent is actually optimizing for. The framing treats agent behavior the way experimental science treats organisms — inferring mechanism from action.