Paper's CrossFit Cuts Search-Agent 'Co-Cheating' to 3.7%

A new arXiv preprint from Meijia Chen, Alaa Khamis and colleagues gives a name to a failure mode they argue is quietly inflating self-evolving search agents: 'co-cheating,' where the proposer and solver settle on the same wrong answer and keep rewarding each other for it. Their proposed training fix, CrossFit, cuts false-agreement mass from 8.8% to 3.7% on Qwen3.5-9B and from 6.1% to 3.0% on Qwen3.5-4B.

In the authors' own description, 'the proposer and solver increasingly agree on shared errors, so internal reward improves without a matching gain in external correctness.' A standard baseline, Multi-Sample Verification, barely moves the needle: on the 9B model it only drops false-agreement from 8.8% to 7.2%.

CrossFit partitions the source corpus so each half's questions are scored by a solver trained on the other half. The authors report average benchmark gains of 8.4 points on 9B and 8.8 points on 4B over coupled self-evolution, and 7.8 and 8.7 points over Search-R1. A further variant they call Source-Excluded Feedback drives residual false-agreement down to 0.1% on 9B and 0.4% on 4B.

It lands the same day as a Hugging Face paper shipping 50,228 agent error-diagnosis pairs, part of a run of agent-robustness work our tracker counted at 416 stories across the last 90 days.