Synthetic data tends to have lower perplexity than human authored data because LLMs tend to write similarly. I hypothesize that this is both why (a) often you need more synthetic data to learn something than if you had human-authored data and (b) synthetic data causes collapse—because the relative size of gradients in certain directions that aren't related to LLM-style is larger.
If you are inspired by this idea, you can reach out to the authors for collaboration or cite it:
@misc{holtzman-small-gradients-as-2026,
author = {Holtzman, Ari},
title = {Small Gradients as a the Potential Mechanism for Synthetic Data Issues},
year = {2026},
url = {https://hypogenic.ai/ideahub/idea/JEVfjBdf894UQGbmk4Qq}
}Please sign in to comment on this idea.
No comments yet. Be the first to share your thoughts!