An AI development team is creating a dataset to improve a language model's ability to solve complex physics word problems. They are considering two approaches for structuring their data:

• Approach A: Each data point consists of the word problem (input) and only the final numerical answer (output).
• Approach B: Each data point consists of the word problem (input) and a detailed, step-by-step derivation of the solution, including the formulas used, intermediate calculations, and the final answer (output).

Which approach is more effective for teaching the model to reason through new, unseen problems? Justify your answer by explaining the underlying learning mechanism for the model in each case.