Show Your Work: Mechanistic Transparency in LLM-Powered Human-Robot Interaction

Large Language Models(LLMs) introduce powerful new capabilities for social robots, yet their black-box nature creates a barrier to trust. Transparency is already established as important for humanrobot trust, but how to convey LLM intentions and reasoning in real-time, embodied interaction remains poorly understood. We developed a task-level mechanistic transparency system for an LLM-powered Pepper robot that displays its internal reasoning process dynamically on the robot’s tablet during interaction. In a mixed-design study, participants engaged with Pepper across four trust-relevant tasks in either a Transparency-ON condition or Transparency-OFF condition. Transparency produced significantly greater trust growth than opacity, and a substantial increase in perceived reliability, indicating that transparency remains a key design element for trust calibration in LLM-driven human-robot interaction.