When confronted with the same type of emergency and under equivalent top-down directives, local governments at the same administrative level exhibit markedly divergent approaches in both the substance and methodology of crisis learning. Drawing on 89 cases of major and particularly serious production safety accidents, this study employs grounded theory and social network analysis to explore the types of crisis learning among local governments and the mechanisms behind their formation. The research identifies three distinct types of crisis learning—Cognitive-Based, Action-Integrated, and Comprehensively-Embedded—each comprising complex combinations of cognitive, behavioral, and institutional learning elements. The selection of learning types is shaped by five key factors: incident characteristics, top-down directives, social pressure, internal motivation, and deterrent sanctions. Further analysis reveals that Cognitive-Based Learning emerges from a mechanism of "top-down directive-attribute-driven"; Action-Integrated Learning arises through "top-down directive-internal and external drivers"; and Comprehensively-Embedded Learning is driven by a "top-down directive-self-motivation" mechanism. To enhance local governments’capacity for crisis learning, tailored improvement strategies should be applied based on the characteristics of each learning type, leveraging top-down efforts while fostering local initiative.