跳转至

递归语言模型训练中的脆弱性图谱

A Fragility Spectrum for Recursive Language-Model Training

arXiv:2609.11149 [cs.CL]
Subjects: Computation and Language (cs.CL)
Authors: Yangze Liu, Zhongyi Han
Submitted: 10 September 2026
Primary Subject: Computation and Language (cs.CL)
arXiv ID: 2609.11149 | DOI: 10.48550/arXiv.2609.11149

文章背景与核心概要

随着互联网上合成数据的爆发式增长，大语言模型 (Large Language Model, LLM) 的生成内容正不可避免地被回流用于新一轮模型的训练。这种“用自己生成的数据训练自己”的递归污染，往往会导致模型输出多样性严重衰退甚至发生“模型崩溃 (Model Collapse)”。本项研究深入剖析了 13 个开源模型在共享语料库上递归训练五代的表现，揭示出模型抗崩溃能力的“脆弱性图谱”——某些模型经历五代依然完好，另一些则迅速退化为无意义的重复词句。更关键的是，研究表明这种脆弱性由模型自身内在特性决定且与参数规模非单纯正相关，同时提出了仅需自我迭代两三代即可廉价预测该风险的诊断方法与有效的推理期缓解策略。

📋 内容摘要

📋 Summary

当语言模型经历数代更迭、持续使用自身生成的文本进行递归训练时，其输出的多样性通常会急剧衰减并走向崩溃。然而，不同模型面对这一递归退化过程时，展现出的耐受力却有着天壤之别。

When language models are recursively trained on their own generated text over multiple generations, output diversity typically collapses. However, different models exhibit drastically different resilience to this process.

本研究通过追踪 13 个公开发布的模型检查点 (Checkpoints) 在连续五代共享通用训练语料库中的表现，系统探究了这种脆弱性图谱 (Fragility Spectrum)。核心发现如下：

This paper investigates a fragility spectrum by observing 13 publicly released model checkpoints sharing a common training corpus over five generations. The findings reveal that:

  • 输出多样性差异悬殊：历经五代递归训练后，不同模型生成的独立 4-gram (Unique 4-gram) 比例分布在 \(0.187\) 至 \(0.940\) 之间，差距高达约五倍；这意味着部分模型几乎未受污染影响，而另一些模型则已退化为高度机械重复的词句片段。
  • 脆弱性是模型的固有属性：模型的抗崩溃脆弱性主要取决于具体的模型检查点本身。无论是改变共享数据池的混合比例、混入人类撰写的真实文本，还是更换随机数种子，各模型的脆弱性排名相关性依然极高 (Spearman 相关系数维持在 \(0.91\text{--}0.98\) 之间)。
  • 参数规模并非唯一决定因素：单纯的模型参数量并不能决定抗崩溃能力；即便在同一模型家族内部，不同参数规模的模型在抗崩溃性上也并不呈现单调变化。
  • 快速预测与缓解策略：仅需让目标模型在自身输出上独立迭代两到三代，即可低成本地推断出其在整个模型生态中的脆弱性。此外，在文本生成阶段收紧 top-p 采样过滤（截断低概率的长尾 Token），可在三代内几乎彻底遏制崩溃趋势，使脆弱性图谱上处于不同区间的各模型重新趋于稳定。
  • Output Diversity Varies Widely: After five generations, unique 4-gram outcomes range from \(0.187\) to \(0.940\) (a roughly five-fold spread), meaning some models remain virtually unaffected while others devolve into repetitive fragments.
  • Fragility is an Intrinsic Property: Model vulnerability is a persistent characteristic of the specific checkpoint itself. It is largely unaffected by changes in data composition, human text mixing, or random seeds (Spearman correlations remain high at \(0.91\text{--}0.98\)).
  • Scale is Not the Sole Predictor: Parameter size alone does not dictate fragility; model families do not scale monotonically in their resistance to collapse.
  • Fast Detection & Mitigation: A model's ecosystem fragility can be cheaply predicted by letting it iterate on its own output for just two or three generations. Furthermore, interventions like tightening top-p generation-time filtering can nearly halt collapse entirely and stabilize diverse checkpoints across the spectrum.

📑 论文摘要

📑 Abstract

AI 模型生成的合成文本正在源源不断地回流到训练语料库中，大量研究证据表明，反复在这类合成数据上进行训练会导致模型输出的多样性崩溃。先前的研究主要聚焦于现象本身，例如探讨哪些递归协议和数据配比会诱发崩溃。然而，在面对完全相同的递归训练过程时，不同的模型却表现出了截然不同的反应与耐受力。

Model-generated text is finding its way back into training corpora, and there is plenty of evidence that training on such data over and over collapses output diversity. Prior work has studied the phenomenon itself: which protocols and which data mixtures cause collapse. But different models behave very differently under the same process.

我们固定了一套递归数据污染协议，让 13 个公开发布的模型检查点构成一个生态系统，使其在五代更迭中持续共享同一个训练语料池。历经五代训练后，各检查点的独立 4-gram 结果在 \(0.187\) 到 \(0.940\) 之间大幅波动，差距约达五倍：有些模型几乎完好无损，而另一些模型则退化成了机械重复的语言碎片。无论是改变共享数据池中的成分构成，还是混入人类真实文本，各模型抗崩溃能力的排序保持着极高的相关性 (Spearman 秩相关系数为 \(0.91\text{--}0.97\) )，更改随机数种子后相关性同样维持在 \(0.93\text{--}0.98\) 。

We fix one recursive contamination protocol and let 13 publicly released checkpoints form an ecosystem that shares a common corpus for five generations. The unique 4-gram outcome after five generations ranges from \(0.187\) to \(0.940\) across checkpoints, a roughly five-fold spread: some models are barely touched, others degenerate into repetitive fragments. Changing the composition of the shared pool or mixing in human text keeps the Spearman correlation of the ordering at \(0.91\text{--}0.97\), and changing the random seed keeps it at \(0.93\text{--}0.98\).

因此，一个模型在递归训练下是否容易发生崩溃，实际上是该模型检查点本身固有的一项内在属性，而此前这一特性在很大程度上被学界所忽视。单纯的参数规模无法解释这种差异，因为在同一模型家族的三个不同尺寸梯队中，耐受力与尺寸并未呈现单调对应关系，且我们测试的所有静态评估指标也都无法对其进行有效预测。行之有效且计算开销极低的方法是：仅需让单个模型在自身生成的输出上独立迭代训练两到三代，便足以推断出其在更大规模生态环境中的脆弱性水平。此外，崩溃的恶化速度对干预手段也有明显的响应。在生成阶段收紧 top-p 采样以剔除低概率的长尾 Token ，几乎能在三代之内阻止崩溃的发生，并使横跨整个脆弱性图谱的六个检查点全部稳定下来；而单纯从数据端进行清洗过滤虽能减缓崩溃速度，却无法彻底阻止崩溃。

Whether a model collapses easily under recursive training is, then, a property of the checkpoint itself, and one that has gone largely unexamined. Parameter scale alone does not explain it, since a three-size ladder within one family is not monotonic in size, and none of the static indicators we tested predicts it either. What does work is cheap: let a model iterate on its own output for two or three generations, and its fragility in the larger ecosystem can be inferred from that alone. Collapse speed also responds to intervention. Tightening top-p, which cuts the low-probability tail at generation time, nearly stops collapse within three generations and stabilizes six checkpoints spanning the whole spectrum together, while data-side filtering slows collapse without stopping it.

🔗 论文资源与相关链接