A data scientist is developing a soft prompt to summarize legal documents. They use the formula hat(σ) = arg min_σ s(hat(y), hat(y)_σ) to find the optimal prompt hat(σ). In this process, hat(y) represents the model's desired summary (generated with full context) and hat(y)_σ is the summary generated using the soft prompt σ. The optimization successfully minimizes the dissimilarity s to near-zero on the training dataset. However, when tested on new, unseen legal documents, the soft prompt produces poor-quality summaries. Analyze the most likely reason for this discrepancy between high performance during training and poor performance in testing.