By flipping the coding of the independent variables for worst tasks, we're making the assumption that worst utility is just the mirror opposite of best utility. This allows us to combine as a single estimation, expanding a single best-worst CBC task into two separate and independent choice tasks (choice of best, choice of worst).
This is a simplification, as there are often subtle differences in how respondents evaluate things as worst compared to how they evaluate things as best. Several papers have been presented at our conference that investigate more complex models that allow for differences in utility and scale between best and worst choices.
That said, the simple approach we use as default in our software works well for practical purposes. And, if you are interested in investigating the possible differences between best utilities and worst utilities for your CBC experiment, you can ask the software to filter then estimation only on the best tasks or only on the worst tasks.