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 2603.11784cs.LGstat.ML
收藏

带重放的语言生成:模型崩溃的学习理论视角

Language Generation with Replay: A Learning-Theoretic View of Model Collapse

University of Copenhagen(哥本哈根大学);Isara Labs(Isara实验室)

Giorgio Racca, Michal Valko, Amartya Sanyal

AI总结:

研究随着前沿大语言模型训练对数据需求增加,生成文本可能重入训练语料库致模型崩溃的问题。通过语言生成极限框架,引入重放对手,刻画重放何时限制生成。

摘要:

随着规模定律推动前沿大语言模型训练对数据需求不断增长,训练管道正接近可能耗尽大量公开在线文本的状态。同时,大语言模型的广泛使用增加了网络上机器生成内容的数量,这增加了生成文本重新进入未来训练语料库的可能性,从而增加了通常称为模型崩溃的性能下降风险。在实践中,模型开发者通过数据清理、加水印、合成数据策略或在某些情况下的 blissful ignorance 来解决这一问题。然而,生成模型中的模型崩溃问题尚未从学习理论的角度进行研究:我们通过语言生成极限框架的理论视角来研究它,引入一个重放对手,该对手用生成器自己过去的输出扩充示例流。我们的主要贡献是对重放从根本上限制生成的情况进行了细粒度的学习理论刻画:虽然重放对于最强的均匀生成概念是良性的,但它可证明会对较弱的非均匀生成和极限生成概念造成分离。有趣的是,我们的正面结果反映了实践中广泛使用的启发式方法,如数据清理、加水印和输出过滤,而我们的分离结果表明这些想法何时会失败。

英文摘要:

As scaling laws push the training of frontier large language models (LLMs) toward ever-growing data requirements, training pipelines are approaching a regime where much of the publicly available online text may be consumed. At the same time, widespread LLM usage increases the volume of machine-generated content on the web; together, these trends raise the likelihood of generated text re-entering future training corpora, increasing the associated risk of performance degradation often called model collapse. In practice, model developers address this concern through data cleaning, watermarking, synthetic-data policies, or, in some cases, blissful ignorance. However, the problem of model collapse in generative models has not been examined from a learning-theoretic perspective: we study it through the theoretical lens of the language generation in the limit framework, introducing a replay adversary that augments the example stream with the generator's own past outputs. Our main contribution is a fine-grained learning-theoretic characterization of when replay fundamentally limits generation: while replay is benign for the strongest notion of uniform generation, it provably creates separations for the weaker notions of non-uniform generation and generation in the limit. Interestingly, our positive results mirror heuristics widely used in practice, such as data cleaning, watermarking, and output filtering, while our separations show when these ideas can fail.

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。

↑