Which practice makes perfect? The role of variability and agency in learning adaptive problem solving

Adaptive problem solving-the ability to flexibly devise solutions to novel challenges-is a hallmark of human intelligence. Yet it remains unclear how best to cultivate this skill through training. Some researchers argue that variable practice promotes learning of generalisable rules, whereas others emphasize repetition and mastery of a single problem space before progressing to another. To address this question, 200 adults solved computerised physical reasoning problems under one of five practice regimens: active high-variable training (AHVT; many different problems), active low-variability training (ALVT; small set repeated), passive high-variable training (PHVT; watching complete trial-and-error attempts on many different problems), passive low-variable training (PLVT; watching complete trial-and-error attempts on a small set repeatedly), or no practice (control). All participants completed a common set of novel problems before and after practice, and we evaluated their performance and strategies. Our findings supported a strong effect of variability: high-variability training led to significantly greater improvements in problem-solving success and efficiency (fewer attempts), together with a more streamlined strategy search characterised by leaner tool use and lower switching between tools than low-variability training. By contrast, agency had minimal impact-participants who actively solved problems did not significantly outperform those who learned by observing, given equal exposure to the same problems, including full attempts, errors, feedback, and eventual successes. There were no reliable interactive effects of variability and agency. Findings suggest that practice variability, more than the physical act of performing each attempt, is key to mastering a set of skills that generalize to new problems. We discuss how variable experiences encourage abstract problem-solving strategies and "learning-to-learn," why observation can sometimes match active doing, and implications for cognitive theory, education, and AI.