A new preprint on arXiv, titled CoEvo-RSI: When Harnesses Evolve, Which Models Should Follow?, explores a paradigm called model–harness co-evolution. The authors describe this as a prominent approach for recursive self-improving (RSI) agents, where the agent's execution scaffold (the harness) and its intrinsic capabilities evolve together rather than separately.
The abstract states that the work studies a "persistent yet often overlooked" aspect of this setup, though the specific detail is cut off in the available text. The core contribution appears to be a framework that treats the harness and the model as coupled components in the improvement loop, asking which models should follow when harnesses evolve.
As a preprint, the paper has not yet undergone peer review. The abstract provides only a high-level summary, so the full methodology and findings remain to be examined. Still, the idea of co-evolving both the tooling and the model itself could be relevant to researchers working on self-improving AI systems.