Definition
Plain language
A signal that favours lessons the learner is currently making headway on — not too easy, not impossible.
As stated in the literature
Curriculum reward based on alignment between a candidate example's gradient and the learner's recent optimizer-scaled weight trajectory, used in place of rewarding raw difficulty.
Also called: learning-progress
Why it matters: Without it, automatic curricula drift toward whatever is hardest, which wastes training on examples the model cannot learn anything from.
For example, a tutor who notices you are finally getting long division keeps feeding you long-division problems, rather than drilling you on addition you already know or calculus you cannot touch yet.
Heard on the show
“This particular learning-progress signal makes adaptive program generation more useful, than continuing to sample from a fixed program distribution.”Episode 280 — Two Random Networks Teach Each Other To Predict Real Data