A development team is training an AI assistant to solve complex, multi-step word problems in mathematics. Their current training method works as follows: the AI generates a complete solution, and if the final numerical answer is correct, the entire solution is marked as 'good.' If the final answer is incorrect, the entire solution is marked as 'bad.' Despite extensive training, the AI struggles to generalize its problem-solving skills to new types of problems.