Behaviour-driven adaptive feedback in LLM-guided 3D spatial tasks: effects on trust, reflection, and task performance

Understanding how intelligent systems can adaptively support human spatial cognition represents a critical challenge in human-computer interaction. We investigate a behaviour-driven adaptive feedback mechanism that leverages Large Language Models (LLMs) to provide contextually aware guidance during 3D spatial tasks. Our approach monitors real-time user behaviours – including hesitation patterns, repeated errors, and strategy changes – to trigger personalised natural language feedback aligned with users’ cognitive states. In a within-subjects experiment (N = 30), participants completed three Soma cube assembly tasks of equivalent difficulty (3-star rating), one per feedback condition: no feedback, request-based feedback, and adaptive feedback. We measured user trust, perceived feedback appropriateness, reflective engagement, and cognitive load using validated scales, alongside behavioural indicators. Adaptive feedback significantly enhanced user trust (r = .370, p = .004), perceived appropriateness (r = .260, p = .046), and reflective engagement (r = .310, p = .017) compared to request-based feedback. Mediation analysis revealed that trust fully mediated the relationship between feedback strategy and feedback appropriateness (98.2% of total effect). Behavioural analysis showed reduced backtracking in feedback-enabled conditions. These findings demonstrate how LLMs can serve as intelligent intermediaries that enhance human cognitive processes while maintaining user autonomy, offering practical guidance for designing adaptive support systems in educational and training environments.