The linearity correction corrects the integrated counts in the science images for the non-linear response of the detector. The correction is applied pixel-by-pixel, group-by-group, integration-by-integration within a science exposure. The correction is represented by an nth-order polynomial for each pixel in the detector, with n+1 arrays of coefficients read from the linearity reference file. The values from the linearity reference file DQ array are propagated into the PIXELDQ array of the input science exposure using a bitwise OR operation.
This link to the JWST pipeline provides detailed information on this step.