What's this about?
Self-evolving language models have made real progress in *verifiable* domains — maths, code — where you can check the answer. J-Zero attacks the unverifiable case, co-evolving three roles from no data at all: a challenger that generates problems, a solver that attempts them, and a judge that scores. 35 upvotes.
Why is the unverifiable case so much harder?
Because in maths the environment tells you whether you are right, so a self-improvement loop has an anchor outside itself. Remove that and the only judge is another model — so the loop can drift anywhere, with all three components agreeing enthusiastically. Co-evolving the judge alongside the solver is the attempt to keep it honest, and it is also the obvious failure mode: nothing prevents challenger, solver and judge from converging on a shared delusion.
Why does this matter to you?
- It is the same structural problem as Warp's self-improving reviewer from last week, which worked because code review has a verifiable signal.
- "Self-evolution as a path toward superintelligence" is stated in the abstract, which is worth knowing about the framing before you read the results.
- Read against Zhipu's Full Self-Training announcement the same day — same idea, one as a paper, one as a product roadmap.