A research team is training a language model to solve multi-step physics problems. The model is trained by generating a complete solution, and the training system automatically checks only if the final numerical answer is correct. If the final answer is correct, the entire generated solution is given a positive reward; if it's incorrect, the entire solution is given a negative reward. Despite extensive training, the model frequently produces solutions that contain logical errors in the intermediate steps, even if it sometimes stumbles upon the correct final answer. Evaluate the team's training methodology. What is the fundamental flaw in this approach for teaching complex reasoning, and why does it lead to unreliable performance? Propose a more effective supervision strategy to address this flaw.