Accurate sky subtraction is essential for precise flux calibration in high-resolution spectroscopy. For the ESPRESSO spectrograph at the Very Large Telescope (VLT), the official ESO pipeline provides two sky-subtraction strategies: "pixel-by-pixel" subtraction and sky modeling via the "smoothed" method. While pixel-by-pixel subtraction affects the final signal-to-noise ratio in particular for faint targets, the smoothed method leads to a systematic underestimation of the observed flux. This "zero-level problem" manifests as a clear negative offset in saturated absorption regions of quasar spectra, driven by flux error distributions and their subsequent propagation during exposure combination.
In this poster, I will present an alternative approach: the running median method. By constructing a robust sky continuum model, this method is designed to overcome the limitations of both official sky-subtraction strategies. Applying this technique to a sample of 23 high-redshift, ultra-high-resolution quasar spectra, I will evaluate its accuracy, performance, and limitations. From this analysis, I will draw general conclusions to inform the development of future data reduction pipelines for high-resolution spectrographs at the ELTs.