Abstract: Coal spontaneous combustion is a global challenge that poses a serious threat to the safe extraction, storage, and transportation of coal resources. With the continuously increasing demand for coal, mining activities have progressively advanced into deeper seams, where coal masses are subjected to long-term and complex in situ stress conditions. Such loading alters both the physical structure and chemical reactivity of coal, profoundly affecting key processes including oxidative heat release, mass transfer, and heat transfer, and resulting in prominent multi-scale complexity in the mechanism of coal spontaneous combustion under stress. To clarify the characteristics at different scales, recent research advances in the spontaneous combustion of stressed loose coal are systematically reviewed from microscopic, mesoscopic and macroscopic perspectives. At the microscale, it emphasizes the mechanism of stress’s influence on intrinsic kinetic parameters such as coal molecular active groups and apparent activation energy. It expounds on the intrinsic mechanism of the mechanochemical effect regulating reaction energy barriers, free radical generation, and functional group evolution from both quantum and molecular levels. At the mesoscale, it analyzes the evolution law of the coal body’s pore-fracture network under stress and its coupled effect on oxygen gas transport and heat accumulation. At the macroscale, it summarizes research on spontaneous combustion experiments and numerical simulations of coal piles or goafs based on porous media theory, revealing the macroscopic effects of environmental parameters. Comprehensive application is made of the research methodology of “experimental testing—theoretical modeling—numerical simulation”, and comments are provided on the cross-scale correlation from microscopic reaction mechanisms to macroscopic disaster evolution, as well as the challenges involved in model construction. Finally, addressing bottleneck problems such as unclear multi-physics coupling mechanisms and missing cross-scale models, it looks ahead to future research directions.