Various types of fiber-optic temperature sensors have been developed on the basis of modal interference in multimode fibers, which include not only glass fibers but also polymer optical fibers (POFs). Herein, we investigate the spectral patterns of the modal interference in multi-core POFs (originally developed for imaging) and observe their unique temperature dependencies with no clear frequency shift or critical wavelength. We then show that, by machine learning, the modal interference in the multi-core POFs can be potentially used for highly accurate temperature sensing with an error of ∼0.3 °C.