Zhu et al. (2024) — ARXIV-2412.14689
Key takeaways
- The premise, which is not hypothetical. As AI output proliferates, future models will be trained on a blend of synthetic and human text whether or not anyone chooses that. So the question is not whether to use synthetic data but what happens when you do.
- The measurement. Pre-training across different proportions of synthetic data yields a negative correlation between the proportion and performance. Statistical analysis of the synthetic data locates why: distributional shift, and over-concentration of n-gram features. Generated text is narrower than what it imitates, and training on it narrows the next model further.
- The remedy inverts the usual move. Rather than generating better synthetic data, edit human data at the token level to produce semi-synthetic data. Because the result stays anchored to a real distribution, the authors prove the test error is bounded above — collapse is prevented by construction rather than by filtering harder.
- Validated on pre-training from scratch, continual pre-training and supervised fine-tuning.
Standing in the anthology
The counterweight to a practice the record holds. SOTA-124 and the recipes in LIT-119 and LIT-131 lean on rephrasing — Kimi K3 rephrases its knowledge and mathematics corpora with diverse prompting and verifies fidelity against the source. That is generation conditioned on human text, which is nearer to this paper’s semi-synthetic prescription than to the pure synthesis it warns about; but the difference is one of degree, and this note is what makes the degree visible.
DeepSeek-V4 (LIT-139) cites it for the other end of the pipeline: filtering batched auto-generated and templated content out of web data, explicitly “to mitigate the risk of model collapse”. So the record now has both uses — a reason to filter the crawl, and a constraint on how to generate.
Filed via the reference pass in #40.