Self evolving search agents avoid co cheating with cross fitted verification
False Frontiers: Diagnosing and Mitigating Co-Cheating in Self-Evolving Search Agents
Computation and LanguageArtificial IntelligenceMachine Learning
Summary
Self-evolving search agents create their own training tasks and try to solve them, but sometimes they trick themselves by agreeing on wrong answers, a problem called co-cheating. The authors found this leads to better internal scores but not better real-world correctness. They improved the process with a new method called CrossFit, which separates data sources to check answers better and rewards only reliable question proposals. Their method reduced errors and improved performance on multiple search benchmarks, showing a way to make self-training search AI more reliable.
What this means in practice
- •For search engine developers: Improve search agents’ training by reducing false agreement in generated question-answer pairs, enhancing overall accuracy during self-evolution.
- •For natural language processing engineers: Incorporate cross-fitting techniques to verify data sources and reduce error reinforcement in AI models that generate and solve their own tasks.
Authors
Meijia Chen, Hao Li, Zheng Lu, Hongshan Lin, Junbai Tian, Yichen Liu, Zijun Tian, Yufan Zou, Shuhan Sun, Hanxin Chen, Zeyu Zhang, Weizhi Du, Yueting Li, Tianyu Shi, Alaa Khamis
Abstract
Self-evolving search agents build their own training curricula by jointly optimizing a proposer that generates questions and a solver that answers them. This closed loop introduces a failure mode we call co-cheating: the proposer and solver increasingly agree on shared errors, so internal reward improves without a matching gain in external correctness. A post-hoc audit against source evidence shows co-cheating growing more severe over successive rounds of self-evolution, with pseudo-label correctness stagnating or declining even as the in-loop training signal improves. The most direct mitigation is to verify proposals before training: we introduce multi-sample verification (MSV), which queries the same model three times with the source and three times without it to decide task admission and replace unreliable pseudo-labels. MSV partially reduces false agreement but leaves substantial residual co-cheating and costs six extra labeler generations per candidate. These limitations motivate CrossFit, our main method: it partitions the proposer's source documents into groups A and B; questions generated from A are scored by an auxiliary solver trained only on B, and vice versa. The cross-fitted agreement determines proposer reward, so a same-source pseudo-label cannot be reproduced through the feedback solver, while the original solver's update rule is unchanged. Rerunning the loop with Qwen3.5-4B and Qwen3.5-9B, MSV reduces false-agreement mass from 6.1% to 5.7% and from 8.8% to 7.2%, whereas CrossFit reduces it to 3.0% and 3.7%. Replaying identical proposals with source-excluded feedback further reduces false agreement to 0.4% and 0.1%, isolating feedback ancestry from curriculum changes. Across seven downstream search benchmarks, CrossFit improves average performance over standard coupled self-evolution by 8.8 and 8.4 points and over Search-R1 by 8.7 and 7.8 points at 4B and 9B.