The refpix step corrects for bias drift. The reference-pixel reference file contains frequency-dependent weights that are used to compute (in Fourier space) the filtered reference pixels and reference output for the reference-pixel correction scheme that is applied to NIRSpec data when exposures use the IRS2 readout pattern. Only the NIRSpec IRS2 readout format requires a reference file; no other instruments or exposure modes require a reference file for this step.
For each sector, the correction is applied as follows: data * alpha[i] + reference_output * beta[i]. Alpha and beta are 2-D arrays of values read from the reference file. The first axis is the sector number (but only for the normal pixel data and reference pixels, not the reference output). The second axis has length 2048 * 712, corresponding to the time-ordered arrangement of the data. Data is the science data for the current integration. The shape is expected to be (ngroups, ny, 3200), where ngroups is the number of groups, and ny is the pixel height of the image. The width 3200 of the image includes the “normal” pixel data, plus the embedded reference pixels, and the reference output. Reference_output is the length of the reference output section.
This link to the JWST pipeline provides detailed information on this step.