Adaptive Human-AI Collaboration: A Review of Multimodal Context Modeling, Uncertainty-Aware Intervention, and Longitudinal Co-Adaptation

RESUMO

Artificial intelligence is shifting from a static decision-support tool to an adaptive collaborator that must sense context, decide when and how to intervene, and improve through repeated interaction with humans individually and in groups. Yet meta-analytic evidence shows that human-AI combinations often fail to outperform the best of either partner alone, and the enabling literature remains fragmented across multimodal sensing, uncertainty quantification, reliance and delegation, facilitation, and teaming. This paper reports a review of adaptive human-AI collaboration literature through a closed-loop lens. Following iterative identification, staged selection against explicit criteria, structured extraction, taxonomy-driven synthesis, and snowballing, we analyze 50 reviewed works. We contribute i) a taxonomy of multimodal context modeling, from individual states to collective states such as group engagement and participation equality; ii) a taxonomy of uncertainty-aware intervention, covering uncertainty sources, estimation and calibration mechanisms, an intervention repertoire that ranges from explanation modulation and deferral to group facilitation, and intervention policies; and iii) a taxonomy of longitudinal co-adaptation and synergy-oriented evaluation. We integrate the three taxonomies into MCAL, a dual-timescale Multimodal Co-Adaptation Loop reference model, and instantiate it on a mixed human-robot workspace, walking every stage of the loop through one concrete setting to show what each taxonomy cell holds in practice.