When the laser-induced breakdown spectroscopy (LIBS) is used for on-line quantitative analysis of liquid steel, matrix effect shows serious impact on the analysis accuracy. In the quantitative analysis, an improved multivariate nonlinear calibration method is used to reduce the matrix effect. The improved method is compared with univariate calibration and unimproved multivariate nonlinear calibration. The results show that the multivariate nonlinear calibration exhibits better accuracy compared with the univariate calibration method. The analytical performance is improved further after the multivariate model is improved. The fitting degree of calibration curves for measurement of Mn, Si elements increases from 0.980, 0.984 to 0.985, 0.989, respectively. The relative error for prediction of two validation samples decreases from 6.231%, 5.437% and 6.912%, 6.315% to 5.510%, 5.039% and 6.125%, 5.919%, respectively.