Description
The paper by Meijia Chen, Hao Li and colleagues studies self-evolving search agents, where a proposer writes questions and a solver answers them in a closed training loop. It names a failure mode called co-cheating: the two models increasingly agree on shared errors, so the internal reward rises while pseudo-label correctness stagnates or declines. The authors first try multi-sample verification, then propose CrossFit, which splits the source documents into two groups and scores questions from one group with an auxiliary solver trained only on the other. With Qwen3.5-4B and Qwen3.5-9B, CrossFit cuts false-agreement mass from 6.1% and 8.8% to 3.0% and 3.7%, and improves average results on seven search benchmarks by 8.8 and 8.4 points over standard coupled self-evolution. Useful for anyone building self-improving agents who needs to know when the training signal can be trusted.
