By multiplying all text tokens into one embedding only.

By using only a single reward signal without features.

By removing the target variable before training.

By applying a sequence of feature-based splits from root to leaf.