Tahe oilfield in the Tarim Basin of China is a typical deep marine carbonate fracture-cavity type reservoir. Large paleokarst caves are the important storage space types in this reservoir. Drilling data reveals that most of the caves have varying degrees of collapsed breccia filling. A better understanding of cave collapse mechanism, modes, and distribution is crucial for highly efficient production. In this study, large karst caves in the Ordovician of Tahe oilfield are taken as research objects, and several types of typical cave conceptual models are established based on similarity theory. Physical tests and numerical simulation methods are used to reveal the collapse mechanism under different depths, shapes, filling types, and tectonic stresses of different periods. Collapse modes are established based on the revealed mechanism, morphology, and damage zone of simulated karst caves. Seismic prediction methods of different types of collapsing caves are explored according to their reflection characteristics. The results reveal that compaction and shearing damage are the main reasons for cave collapsing. The roof and surrounding rock of the cave sink in during the collapse process, with a shear-slip damage zone of 2–3 times larger than the cave diameter. The sand/mud-filled cave is shown to be more stable than the water-filled cave, and the unfilled cave is the least stable. Tectonic stresses acting in the early Hercynian period are the primary reason for the current collapsed caves. Two cave collapse modes are established, type Ⅰ being mainly affected by overlying gravity and type Ⅱ mainly affected by shear stress. There are local differences in seismic reflection characteristics between these two types. The prediction of the two types of collapsed karst caves can be realized by combining the seismic attribute of colored inversion, frequency division energy, GR inversion and coherent energy gradient, etc. The comprehensive prediction results of collapsed karst caves in Tahe oilfield show a good correspondence with oil well productivity.