Generating code with execution feedback is difficult because errors often require multiple corrections, and fixing them in a structured way is not simple. Training models to learn from execution ...