An autoregressive model can receive a clean prefix at every training position and still face a different input distribution during generation. Training commonly scores the next reference token while conditioning on earlier reference tokens. At inference time, the prefix contains the model’s own outputs instead.

This mismatch is called exposure bias. It matters because an early generation error does more than make one token incorrect. That token becomes part of the context for later predictions, placing the model in a prefix state that may have been rare or absent during training.

Token likelihood is conditioned on a prefix

For a target sequence y_1, ..., y_T, autoregressive factorization writes its probability as:

p(y_1, ..., y_T | x)
  = product_t p(y_t | y_<t, x)

Under ordinary teacher-forced maximum-likelihood training, y_<t is the reference prefix. Each position is optimized using the sequence that appears in the training example, regardless of what the model itself would have emitted at earlier positions.

Generation uses a different recurrence:

y_hat_t ~ p(. | y_hat_<t, x)

Now each predicted token becomes input to the next prediction. If y_hat_4 differs from the reference token, position 5 is conditioned on a prefix that training did not use for that example.

The objective is still coherent: maximum likelihood estimates conditional token distributions on observed data. The mismatch appears when those conditionals are composed autoregressively under prefixes produced by the model itself.

A local error can change later inputs

Consider a structured output with a required delimiter:

reference:  name : value , status : active
generated:  name = value , status : active

The substitution of = for : is a local error, but subsequent predictions now see name = rather than name :. A model that has only been optimized on valid reference prefixes receives no direct training signal for every possible malformed prefix it may create.

The effect is not limited to syntax. A generated entity, number, or assertion can alter later token probabilities because later attention operates over that generated content. Errors can therefore propagate through the same conditioning mechanism that gives autoregressive models their sequential coherence.

This does not imply that every early error causes a cascade. A model may recover if nearby context strongly constrains the next tokens. Exposure bias describes the train-inference distribution mismatch, not a guarantee of failure after any incorrect token.

Teacher forcing makes position-wise training convenient

Reference-prefix conditioning has a major computational property in transformer training: target positions can be processed in parallel under a causal mask. The model receives the known target sequence, and losses for many positions can be computed in one forward pass.

Replacing reference prefixes with sampled model prefixes changes that structure. Generated token t depends on earlier generated tokens, so obtaining realistic sampled prefixes introduces sequential dependence or requires an approximation.

This is one reason exposure-bias interventions are not merely a data-format change. They can alter the training objective, the prefix distribution, and the computation needed to produce those prefixes.

Scheduled sampling changes the prefix source

Scheduled sampling was proposed as a way to expose a sequence model to some of its own predictions during training. Instead of using a reference previous token at every position, training sometimes feeds a model-generated token.

A simplified choice looks like:

input_t =
    reference_(t-1)    with probability p
    generated_(t-1)    with probability 1 - p

The probability can change over training. This reduces the gap between fully reference-conditioned training and autoregressive inference, but it also changes the objective. The model is now optimized on prefixes assembled from different sources rather than solely on prefixes drawn from the reference data.

For transformers, the dependency is more involved than replacing one recurrent input. Each position can attend to the full causal prefix, so implementations of scheduled sampling may require multiple decoding passes or another mechanism for constructing mixed prefixes.

Scheduled sampling should therefore be treated as one specific intervention, not as a general correction that preserves all properties of maximum-likelihood training.

Sequence-level evaluation reveals effects hidden by token loss

Teacher-forced validation loss evaluates next-token probabilities under reference prefixes. That is useful for measuring the fitted conditional distribution, but it does not reproduce the state distribution encountered during free-running generation.

A generation system can therefore benefit from evaluation that actually generates sequences. The relevant check depends on the output contract: exact structural validity, execution success, task accuracy, constraint satisfaction, or another property tied to the application.

The distinction is especially useful when comparing two models with similar teacher-forced loss. Similar loss does not imply that both models recover equally after generating an off-reference prefix. Free-running evaluation observes the consequences of the model’s own prior outputs.

Controlled prefix perturbations can add another view. Supplying a plausible but incorrect earlier token and measuring later behavior tests sensitivity to states outside the pristine reference path. Such tests should use perturbations that match realistic generation errors rather than arbitrary noise.

Decoding policy also changes the visited prefixes

Exposure bias originates in the training and inference prefix mismatch, but the inference distribution is not fixed independently of decoding. Greedy decoding, beam search, temperature sampling, and token filters can lead the same model into different prefix regions.

A low-probability token admitted by sampling may create a prefix that greedy decoding never visits. Conversely, deterministic decoding can repeatedly select a locally preferred continuation that leads to a poor sequence-level outcome.

As a result, robustness to generated prefixes should be evaluated with the decoding policy intended for deployment. A prefix distribution collected under one policy is not automatically representative of another.

The useful boundary is distributional, not terminological

Exposure bias is most actionable when stated as a concrete distribution shift: the model is optimized mainly on reference prefixes but deployed on prefixes induced by its own predictions and decoding policy.

That framing keeps the engineering question precise. The issue is not simply that generation is autoregressive. It is that later predictions depend on states whose frequency can differ between training and inference. Changes to training data, objectives, prefix construction, or decoding should be judged by how they alter that state distribution and by the sequence-level behavior that follows.

Artificial Intelligence LLM Sequence Models Training Inference