Current modeled ability.
Difficulty admitted over training
Eligible difficulty ≤ competence 0.20
Examples eligible under the pace.
Challenge adjusted by current skill.
Penalty when easy examples disappear.
Ordering is an optimization choice.
Pacing controls exposure
A curriculum defines when examples become eligible. Too slow wastes data; too fast behaves like uniform sampling.
Difficulty needs a signal
Human rules, model loss, confidence, length, noise, or task structure can rank examples, but every proxy can encode bias.
Revisit earlier material
Hard-first transitions can destabilize learning, while removing easy examples entirely can cause forgetting. Mix replay with expansion.