Job interview anxiety is a prevalent challenge among university students and can undermine both performance and confidence in high-stakes evaluative situations. Social robots have shown promise in reducing anxiety through emotional support, yet how such systems should balance psychological safety with effective instructional guidance remains an open question. In this work, we present a three-phase iterative design study of a robotic interview coach grounded in Person-Centered Therapy (PCT) and instructional scaffolding theory. Across three weekly sessions (N=8), we systematically explored how different interaction strategies shape users' emotional experience, cognitive load, and perceived utility. Phase I demonstrated that a PCT-based robot substantially increased perceived psychological safety but introduced a Safety-Guidance Gap, in which users felt supported yet insufficiently coached. Phase II revealed a Scaffolding Paradox: immediate feedback improved clarity but disrupted conversational flow and increased cognitive load, whereas delayed feedback preserved realism but lacked actionable specificity. To resolve this tension, Phase III introduced an Agency-Driven Interaction Mode that allowed users to opt in to feedback dynamically. Qualitative findings indicated that user control acted as an anxiety buffer, restoring trust, reducing overload, and reframing the interaction as collaborative rather than evaluative. Quantitative measures further showed significant reductions in interview-related social and communication anxiety, while maintaining high perceived warmth and therapeutic alliance. We synthesize these findings into an Adaptive Scaffolding Ecosystem framework, highlighting user agency as a key mechanism for balancing emotional support and instructional guidance in social robot coaching systems.

翻译：面试焦虑是大学生普遍面临的挑战，在高风险评估情境中可能损害其表现与自信。社交机器人已显示出通过情感支持减轻焦虑的潜力，但此类系统应如何平衡心理安全感与有效教学指导仍是一个开放性问题。本研究基于以人为中心疗法（PCT）与教学脚手架理论，提出了一个三阶段迭代设计的机器人面试教练系统。通过为期三周的八人参与实验，我们系统探究了不同交互策略如何影响用户的情感体验、认知负荷与感知效用。第一阶段表明：基于PCT的机器人显著提升了感知心理安全感，但引发了"安全-指导缺口"现象——用户感到被支持却未获得充分指导。第二阶段揭示了"脚手架悖论"：即时反馈能提升清晰度却破坏对话流畅性并增加认知负荷，而延迟反馈虽保持真实感却缺乏可操作的针对性。为化解此矛盾，第三阶段引入了"能动性驱动交互模式"，允许用户动态选择是否接收反馈。定性研究发现：用户控制权发挥了焦虑缓冲作用，能重建信任、减轻认知超载，并将互动重构为协作性而非评估性体验。定量测量进一步显示，面试相关的社交与沟通焦虑显著降低，同时维持了较高的感知亲和力与治疗联盟。我们整合这些发现构建了"自适应脚手架生态系统"框架，强调用户能动性作为社交机器人教练系统中平衡情感支持与教学指导的关键机制。