Stopping the computation once the answer is already clear rather than pushing every input through the entire model. Easy inputs do not need eighty layers. It saves a great deal on average while leaving hard cases fully processed.
Technical view
Attaches classifiers at intermediate layers and exits when confidence exceeds a threshold, requiring training of the intermediate heads and calibration of the threshold. Complicates batching, since different inputs exit at different depths, which limits gains in high-throughput serving. Related ideas include cascade routing between separate models of different sizes and adaptive computation allocated per token, which is easier to schedule.
Related
Part of AI Terminology Index, section: Making Models Smaller, Faster and Cheaper.