J-Zero: Unified Challenger--Solver--Judge Co-Evolution from Zero Data
J-Zero trains a challenger, solver, and judge together without starting from labeled data.
The paper targets self-evolution in both verifiable and harder-to-check domains. Its judge is updated from preference pairs whose order is known from the response process, rather than from the judge’s own scores. The authors report gains of 4.2 points in verifiable domains and 8.0 points in unverifiable ones, with improvement continuing through at least ten iterations. HF Daily Papers' note
The paper targets self-evolution in both verifiable and harder-to-check domains. Its judge is updated from preference pairs whose order is known from the response process, rather than from the judge’s own scores. The authors report gains of 4.2 points in verifiable domains and 8.0 points in unverifiable ones, with improvement continuing through at least ten iterations. HF Daily Papers' note
score 5