A new benchmark finds that even top AI agents routinely ignore who they are talking to and act outside a user's role.
Researchers built ReFract, a 150-entry benchmark drawn from anonymized industrial support conversations. Each entry requires an agent to act differently depending on the user's role, such as a technician versus a supervisor. The benchmark uses Text World Models that simulate equipment environments, scoring agents on whether they take actions a role is actually authorized to take, not just whether the response sounds reasonable. The best models solved at most 69% of the tasks, and more than half of all attempted trajectories included at least one perspective-violating action.
In coding, a bad AI suggestion gets caught in code review. In industrial maintenance, an agent that lets the wrong role trigger the wrong action can damage equipment or hurt someone. The paper argues that perspective awareness, knowing not just what to do but for whom, is a mostly unaddressed gap in agent evaluation that focuses on task success alone.
Most agent benchmarks ask whether the answer was correct. This one asks whether the agent should have given it at all, a more useful question once these systems start acting on real machines instead of text boxes.