During inference, the model still follows the original decoding and post-processing pipeline, and the effect of the loss-function adjustment is mainly reflected in improved regression-learning ...