challenger--solver--judge

Definition

A three-role training loop: a challenger proposes problems, a solver attempts them, and a judge scores the attempt and returns the reward signal. Splitting the roles lets difficulty adapt automatically, since the challenger can be pushed toward problems the solver currently fails. The weak point is the judge — whatever it fails to penalise becomes the ceiling on what the solver actually learns.

Explainers using this term

• Paper Walkthrough — J-Zero: Growing the Challenger, the Solver, and the Judge Together from Zero DataJ-Zero: Unified Challenger--Solver--Judge Co-Evolution from Zero Data