Learning is recognized as central to collaborative adaptive management (CAM), yet few longitudinal studies examine how learning occurs in CAM or apply the science of learning to interpret this process. We present an analysis of decision-making processes within the collaborative adaptive rangeland management (CARM) experiment, in which 11 stakeholders use a structured CAM process to make decisions about livestock grazing and vegetation management for beef, vegetation, and wildlife objectives. We analyzed four years of meeting transcripts, stakeholder communications, and biophysical monitoring data to ask what facilitated and challenged stakeholder decision making, how challenges affected stakeholder learning, and whether CARM met theorized criteria for effective CAM. Despite thorough monitoring and natural resource agency commitment to implementing collaborative decisions, CARM participants encountered multiple decision-making challenges born of ecological and social complexity. CARM was effective in achieving several of its management objectives, including reduced ecological uncertainty, knowledge coproduction, and multiple-loop social learning. CARM revealed limitations of the idealized CAM cycle and challenged conceptions of adaptive management that separate reduction of scientific uncertainty from participatory and management dimensions. We present a revised, empirically grounded CAM framework that depicts CAM as a spiral rather than a circle, where feedback loops between monitoring data and management decisions are never fully closed. Instead, complexities including time-lags, trade-offs, path-dependency, and tensions among stakeholders' differing types of knowledge and social worlds both constrain decision making and foster learning by creating disorienting dilemmas that challenge participants' pre-existing mental models and relationships. Based on these findings, we share recommendations for accelerating learning in CAM processes.