Abstract:316L stainless steel can produce kinematic hardening effect under cyclic loading, resulting in the accumulation of ratcheting strain and thus greatly reducing the fatigue life of the material. The generation of kinematic hardening can be considered as the movement of the yield surface in the stress space during the loading process, which leads to the asymmetry of the yield strength in tension and compression. Based on the results of cyclic loading tests, many kinematic hardening rules have been proposed. Some rules are suitable for the simulation of uniaxial cycle loading case. However, for the multiaxial non-proportional variable amplitude histories, the calculation results of existing models have large overestimation compared with test results. In this work, the evolution of back stress of several classic kinematic hardening models are studied. And the moving direction of back stress is discussed as well. The stress-controlled ratcheting experiments of 316L stainless steel under uniaxial and multiaxial loading paths are conducted to verify the influences of mean stress, stress amplitude and loading history. The strain-controlled cyclic loading is also conducted to verify the stress relaxation. It is demonstrated that the axial ratcheting is obvious under symmetrical shear loading path, and the ratcheting strain increases with the increasing stress amplitude and mean stress. The influences on the direction of back stress component increment, which are induced by the uniaxial and multiaxial parameters in Chen-Jiao and Jiang-Sehitoglu kinematic hardening rules, are discussed. Based on the experimental results, Chen-Jiao’s kinematic hardening rule is improved by replacing the multiaxial parameter with a surface saturated ratio, and a new parameter is introduced for correcting the plastic modulus coefficients. The calculated results show that the improved rule predicts a similar mean stress with Chen-Jiao’s rule, and coincides with experimental data much better than Chen-Jiao’s kinematic hardening rule under multiaxial loading case, which also proves the improvement is correct and valid.