arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2510.16657stat.MLcs.LG
收藏

通过合成数据验证逃避模型崩溃:短期改进与长期收敛

Escaping Model Collapse via Synthetic Data Verification: Near-term Improvements and Long-term Convergence

Department of Statistics, University of Chicago(统计学系,芝加哥大学);Department of Computer Science, University of Chicago(计算机科学系,芝加哥大学);Shanghai University of Finance and Economics(上海财经大学);UChicago(芝加哥大学)

Bingji Yi, Qiyuan Liu, Yuwei Cheng, Haifeng Xu

AI总结:

本文提出通过引入外部验证者防止模型崩溃,通过理论分析和实验验证,展示验证者引导的重新训练在短期提升模型性能,但长期趋向于验证者的知识中心。

摘要:

合成数据已被越来越多地用于训练前沿生成模型。然而,最近的研究指出,反复在生成的合成数据上重新训练生成模型可能会持续恶化模型性能,这种现象通常被称为模型崩溃。在本文中,我们研究了修改合成重新训练过程以避免模型崩溃的方法,并可能帮助从崩溃转向改进。我们的关键发现是,通过引入外部合成数据验证者(无论是人类还是更好的模型)的信息,合成重新训练不会导致模型崩溃。具体来说,我们将在基本的线性回归设置中进行理论分析,显示验证者引导的重新训练可以带来短期改进,但最终会使参数估计趋向于验证者的“知识中心”。我们的理论进一步预测,除非验证者完全可靠,这些早期收益将趋于平缓,甚至可能逆转。确实,我们在线性回归、在MNIST上训练的变分自编码器(VAEs)以及在XSUM任务上微调SmolLM2-135M的实验中验证了这些理论见解。

英文摘要:

Synthetic data has been increasingly used to train frontier generative models. However, recent studies raise key concerns that iteratively retraining a generative model on its self-generated synthetic data may keep deteriorating model performance, a phenomenon often coined model collapse. In this paper, we investigate ways to modify the synthetic retraining process to avoid model collapse, and even possibly help reverse the trend from collapse to improvement. Our key finding is that by injecting information through an external synthetic data verifier, whether a human or a better model, synthetic retraining will not cause model collapse. Specifically, we situate our theoretical analysis in the fundamental linear regression setting, showing that verifier-guided retraining can yield near-term improvements, but ultimately drives the parameter estimate to the verifier's "knowledge center" in the long run. Our theory further predicts that, unless the verifier is perfectly reliable, these early gains will plateau and may even reverse. Indeed, our experiments across linear regression, Variational Autoencoders (VAEs) trained on MNIST, and fining-tuning SmolLM2-135M on the XSUM task confirm these theoretical insights.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑