J-Zero Grows Challenger, Solver and Judge From Zero Data

Self-improvement loops have a standard cause of death: the judge. Fix a reward model or an LLM judge in place, let the policy improve against it, and within a couple of rounds the policy learns the judge instead of the task. J-Zero, from Eunho Yang's group at KAIST, attacks that directly by refusing to freeze anyone: a Challenger generates progressively harder tasks, a Solver answers them, and a Judge learns to evaluate, all three co-evolving from zero external data. Paper: arXiv 2608.26582.

The headline result isn't the score bump, though it's respectable: +4.2 points on verifiable tasks, +8.0 on unverifiable ones. It's the durability. J-Zero keeps improving through ten iterations where baseline self-play setups deteriorate after two. And the bigger gain landing on unverifiable domains matters most: that's exactly where compute-driven RL usually can't go, because there's no ground truth to check answers against. A judge that grows with the solver is the substitute for ground truth.

File this with the self-improvement-with-receipts thread. The evolving-judge idea is also what every "agent skills that optimize themselves" system quietly depends on; if the three-role version holds up outside KAIST's benchmarks, it's a template.