Definition
Plain language
Teaching a model by having it copy recorded examples of experts doing the task correctly.
As stated in the literature
A training paradigm that fits a policy to expert demonstrations via supervised next-action prediction; effective but data-hungry and brittle to distribution shift, contrasted with interaction-driven reinforcement learning.
Why it matters: It matters because copying experts is a fast way to teach competence, but it leaves models brittle when they hit situations the demonstrations never covered.
For example, you might train a driving model by feeding it thousands of recordings of human drivers and having it predict the next move a human would make.
Heard on the show
“RL learns from the model's own dead ends — and dead ends are precisely the data that imitation learning structurally cannot contain.”Episode 163 — Why Training Only on Perfect Solutions Cripples a Model's Reasoning