Exposure Bias

The problem where a model sees ground truth during training but its own output at inference, so errors compound.

Exposure bias is a classic problem in sequence generation models. Autoregressive models (models that predict the next element step by step) are usually trained with teacher forcing, feeding in the real history at every step; but at inference, the only history available is the model's own previous output. Because the model never saw its own mistakes during training, once it drifts off the training distribution, its errors compound further and further. Bengio and colleagues' scheduled sampling (2015, gradually switching training to use the model's own predictions as input) and Ranzato and colleagues' sequence-level training (2016) were early representative fixes. In embodied AI it shows up in two common forms: a behavior-cloning policy drifting further and further from the demonstrated trajectory once it deviates (often called compounding error or covariate shift), and an autoregressive video world model's image quality collapsing after generating dozens of frames in a row — methods like Self Forcing try to ease this by having the model continue generating from its own generated frames during training too.

ExampleGiven a video world model one current frame and asked to autoregressively predict the next 100 frames: during training it saw the real previous frame at every step, but at inference it sees its own slightly flawed generated frame, so objects tend to deform and the image tends to blur more and more the further out it goes.

Also called

Train-Test Mismatch

Related

Teacher Forcing · Autoregressive Decoding · Compounding Error · Distribution Shift · Self Forcing · Diffusion Forcing

Sources

Sequence Level Training with Recurrent Neural Networks (Ranzato et al., ICLR 2016)
Scheduled Sampling for Sequence Prediction with Recurrent Neural Networks (Bengio et al., 2015)
Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion (arXiv 2506.08009)