“The STaR algorithm improves iterative fine-tuning by adding a "rationalization" process where the model generates Chain-of-Thought explanations backward from the ground truth answer for problems it failed to solve.”