Training error is low, but validation error remains high as model complexity grows. Which direction should be evaluated?

1. Greater complexity as the only remedy
2. Moving validation labels into input features
3. Stronger regularization, simpler models, or more representative training data
4. Reporting only training error