Found first: a primary source the press has not covered yet.
A new framework called J-Zero co-evolves a task generator, a solver, and a judge from zero human-labeled data, gaining an average of 8.0 points over baselines on unverifiable domains and 4.2 points on verifiable ones. The paper, by Gyouk Chu, Myeongho Jeon, and Eunho Yang, reports that J-Zero continues improving through at least ten iterations while baselines degrade after two.
What the source says
J-Zero trains three components together: a Challenger that generates tasks, a Solver that answers them, and a Judge that evaluates responses. No human labels are used and no external reward signal is required. Rather than having the Judge score its own outputs to construct preference pairs, the framework derives preference ordering from how each response was produced: the Solver's answer is preferred over the Challenger's, and a decomposed-and-recombined answer is preferred over the Solver's one-shot answer. This sidesteps the circular reward hacking that causes other self-play judges to collapse. The Challenger and Solver co-evolve adversarially while the Judge adapts using these implicit preference signals.
Why it matters
Automated self-training has worked well in math and code because correctness can be checked programmatically. Creative writing and open-ended reasoning have resisted the same approach because no external verifier exists. The average +8.0-point gain on unverifiable domains is the sharpest evidence in the paper that this specific bottleneck can be broken without human annotation. The preference construction method is the mechanism to watch: by grounding pair ordering in the generation process rather than the Judge's own scores, it avoids the failure mode that limits comparable frameworks to two iterations. Affiliations are not listed in the abstract page; the authors' institutional home would need to be confirmed from the full paper.