J-Zero Enables Self-Evolving LLMs in Unverifiable Domains.

The pursuit of superintelligence through self-evolving language models is gaining traction, largely due to its potential to reduce reliance on costly human supervision. While progress has been notable in domains where task verification is straightforward, self-evolution in "unverifiable domains" (where objective correctness is hard to determine) remains a significant challenge. This paper introduces J-Zero (Judge co-adaptation from Zero data), a comprehensive framework designed for self-improvement across both verifiable and unverifiable domains. J-Zero operates on a co-evolutionary principle involving three components: a Challenger, a Solver, and a Judge. The Challenger generates increasingly difficult tasks, while the Solver learns to produce higher-quality responses in an adversarial loop. Simultaneously, the Judge adapts by learning from preference pairs where the ordering is intrinsically known (e.g., a Solver's improved answer over a Challenger's initial one), rather than relying on external scores. This novel approach allows J-Zero to continuously improve over many iterations, outperforming baselines by substantial margins in both verifiable and, critically, unverifiable domains, where other methods often degrade.