Traditional approaches include introducing intermediate rewards, reward shaping, or hindsight relabeling (such as HER), but these methods either rely on manual design or are difficult to adapt to the highly heterogeneous scenarios in MLE where code execution feedback is complex.