Teacher forcing during SFT: which description nails it?

**Answer B is correct.** Teacher forcing is the standard way transformer language models are trained on supervised data. At every position in the sequence, the model is conditioned on the gold previous tokens (the tokens from the training example), not on tokens it generated itself. This design choice has a huge practical payoff. Because each position is conditioned on gold context, the predictions at different positions are independent given the input. That means the model can score the whole sequence in a single parallel forward pass through the transformer. Every position predicts the next gold token simultaneously, and cross-entropy is summed across all positions. If you did not use teacher forcing, you would have to generate token by token, plugging each generated token into the next step's input. That would be hundreds of times slower per training example and would also chase a moving target, since the input distribution at each step depends on the model's own (initially terrible) outputs. The three wrong options describe different things entirely. Option A is scheduled sampling, option C is knowledge distillation, option D is fiction.

**Option B.** Teacher forcing means that at every position in the training sequence, the input fed to the model is the gold previous token from the training example, not a token the model generated. This is the standard training recipe for autoregressive transformers. The property that matters is parallelism. With gold context at every position, the model can produce next-token predictions for the entire sequence in a single forward pass. The transformer's causal attention mask ensures each position only sees prior tokens, so the prediction at position t depends on positions 0 through t-1, all of which are gold. Cross-entropy is computed at every position and summed, giving the full loss in one shot. Without teacher forcing, you would need to generate token by token, feeding each generated token back as input for the next step. That serial process is hundreds of times slower and also chases a moving target during early training, when the model's outputs are noisy. The distractors describe related but distinct concepts. Scheduled sampling (option A) mixes gold and sampled context to bridge train-test mismatch. Knowledge distillation (option C) is about what targets the student matches, not what previous tokens condition each step. Option D invents a mechanism that does not exist.