Self-evolving language models improve by training on tasks they generate themselves, but that creates a recurring question: which updated model should the next round inherit? A new arXiv paper, Selection Is Not Learning: Choosing Model Futures, argues that this selection step deserves its own analysis, separate from the learning process.
The paper points to existing loops such as R-Zero, Agent0, and Skill Self-Play, noting that they make the inheritance decision through the task itself. That approach, the authors suggest, may conflate two different problems: how to learn from generated tasks, and how to choose which model version to carry forward.
By framing selection as a distinct issue, the work opens a new angle for improving self-evolving systems. Rather than assuming that better learning automatically yields better choices, researchers may need to design explicit selection mechanisms that consider the future trajectory of the model, not just the current round's performance.
As a single-source digest, this summary reflects only the abstract's stated motivation. The full paper likely offers deeper analysis, but the core claim is already clear: selection is not learning, and treating it as such could change how self-evolving models are built.