Does teacher forcing use the model’s previous guess?

Full teacher forcing uses the previous reference target instead.

Is the correct next token shown as input?

It should not be visible to the prediction that is meant to learn it; correct shifting and masking are essential.

Can it apply beyond recurrent networks?

Yes. The principle concerns conditioning on reference history and is not limited to one sequence architecture.

AI Glossary / Natural language and speech / Intermediate

Teacher forcing

Definition

Teacher forcing is a training method for sequence models in which the correct previous target values are supplied as context for predicting later target values. For an autoregressive text model, this means conditioning each next-token prediction on the reference prefix rather than on tokens the model generated for earlier positions during that training example.

What it means

A sequence generator must learn how to continue from earlier outputs. During training, the reference sequence is available, so the model can be given the correct preceding tokens while its next prediction is compared with the correct next token.
The model still makes predictions and incurs a loss when they are poor. Supplying the reference history does not mean the current answer should be exposed as input to the same prediction.
At ordinary free-running generation time, the reference continuation is unavailable. The model conditions on the outputs it actually generated, which may differ from the prefixes it saw under teacher-forced training.

How it works

Take a target sequence and construct shifted inputs and targets. A beginning marker can provide the initial context; subsequent inputs contain the preceding reference tokens.
At each position, compute a distribution over possible next outputs and compare it with the reference target, often using cross-entropy. Update model parameters through the chosen training procedure.
For architectures that process several positions together, causal masking must prevent a prediction from accessing its own target or later tokens. Teacher forcing supplies earlier reference history, not unrestricted access to the future.
Evaluate generated sequences separately from teacher-forced loss. The two settings provide different evidence because the model can encounter its own earlier mistakes during generation. Methods that mix generated and reference histories modify the training policy rather than being identical to full teacher forcing.

Three practical examples

A translation decoder trains on reference target prefixes while predicting the next translated token.
A language-model experiment computes next-token loss using the actual preceding tokens in a held-out text.
A sequence-model researcher compares teacher-forced evaluation with free-running generation to investigate error accumulation.

Think of it this way

Think of practicing each next step after an instructor places you at the correct preceding step. You still have to choose the next move, but you are not always practicing recovery from your own earlier mistake. The analogy is limited because the model learns numerical parameters from many examples rather than consciously following an instructor.

Worked example

Suppose the fictional reference output is “open the window” followed by an end marker. The teacher-forced prediction tasks are: from the beginning marker, predict “open”; from “open,” predict “the”; from “open the,” predict “window”; and from “open the window,” predict the end marker.
Assume the model’s first prediction would prefer “close” over “open.” That error affects the loss for the first target. Under full teacher forcing, the second prediction still receives the reference token “open” as its preceding token, not the model’s preferred “close.”
During free-running generation, choosing “close” would instead make the later context begin with “close.” The subsequent outputs could therefore differ.
This example shows why a model can be trained against every correct next token while still needing separate evaluation of complete generated sequences. It also shows the importance of the one-position shift: giving “window” to the model as visible context while asking it to predict that same token would create target leakage.

Illustrative case study

In a fictional project, Silverline Translation monitors a decoder’s validation loss. The loss improves, but reviewers still find that an early mistaken output can lead a generated sentence away from the intended message.
The team keeps the teacher-forced loss as a useful training diagnostic and adds evaluation of complete translations produced without reference target prefixes. It examines where the generated histories diverge rather than assuming the two evaluations measure exactly the same behavior.
The imagined review does not imply that teacher forcing is inherently defective or that another training method automatically solves every generation error. It identifies the practical limits of one evaluation setting.

Common misconceptions and mistakes

Teacher forcing does not mean replacing the model’s loss with the correct answer. The model still predicts and is trained against a target.
It should not reveal the current or future target tokens to the prediction through an incorrect shift or mask.
It is not the same as ordinary inference with a known prompt; during training the reference continuation supplies successive histories.
Finally, low teacher-forced loss does not guarantee error-free generated sequences, especially when generation moves into contexts created by earlier model decisions.

Business use

Understanding teacher forcing helps teams interpret model-training reports and design meaningful evaluation. It clarifies why token-level loss and the quality of complete generated outputs can move differently.
Check the alignment of inputs and targets, the masking rules, and which positions contribute to the loss. These details determine whether the reported metric reflects a legitimate prediction task or accidental access to answers.
For products that generate sequences, evaluate the outputs users actually receive, including long continuations and cases where an early error matters. Training diagnostics are valuable, but they should be connected to the deployment behavior rather than treated as interchangeable evidence.

Check your understanding

The reference begins “send the file,” but the model would first output “delete.” Under full teacher forcing, what preceding token is supplied for predicting “the”?

Answer: “Send,” the reference token, while the mistaken first prediction still contributes to the training loss.

Related terms

Decoder
Cross-entropy
autoregressive model
sequence-to-sequence model
causal masking
target leakage
exposure bias
scheduled sampling

Learn next

Trace shifted inputs and targets for a short sequence, then compare teacher-forced loss with free-running output. Study how causal masking prevents future-target access and why alternative history-sampling policies change the training setup.

Sources

Source: Jurafsky and Martin, Speech and Language Processing, recurrent language-model training and encoder–decoder teacher forcing: https://web.stanford.edu/~jurafsky/slp3/14.pdf . Independently consulted sections 14.2 and 14.7.1 on reference-history conditioning, shifted next-token loss and training/inference differences.

Previous: TD(lambda) Next: Technical control