Exposure Bias Problem
You'll learn: Why autoregressive models struggle when their training and inference conditions don't match.
The Core Mismatch
During training, we use teacher forcing: the model always sees the correct ground-truth tokens from the training data as context when predicting the next token. It's like learning to write an essay with perfect reference sentences always available.
But at inference time, the model uses autoregressive sampling: it generates text token-by-token using its own previous predictions as context—not the ground truth. If it makes a mistake early on, that error becomes part of the context for all future predictions.
The Problem in Practice
Imagine learning to navigate by always being given the correct previous directions. Then suddenly, you must navigate using your own possibly-wrong turns as reference points. Small errors compound: one wrong prediction leads to increasingly off-track subsequent predictions because the model never learned to recover from its own mistakes during training.
Example scenario:
- Training context: "The cat sat on the [correct: mat]" → predict next word
- Inference reality: "The cat sat on the [model predicted: car]" → now must predict next word given this error
The model was trained on perfect sequences but must generate from imperfect ones. This distribution shift between training inputs (always correct) and test inputs (sometimes wrong) is exposure bias.
Consequences
Models may generate fluent text initially but drift into repetition, incoherence, or off-topic content as errors accumulate. The model hasn't been "exposed" to recovering from its own mistakes during training.
Key Takeaway: Exposure bias arises because teacher forcing during training creates a mismatch with autoregressive inference, where the model must condition on its own potentially incorrect predictions rather than ground-truth tokens.