Source
Zhu et al. (2024), LIT-205, and Finke et al. (2025), LIT-485 — read as NOTE-234. Two groups, two corpora, two instruments, no citation in either direction.
What it explains
| practice | what it says to do | what this says is going on |
|---|---|---|
| SOTA-172 | build synthetic pretraining data by editing human text at the token level, not by generating from scratch | editing keeps the corpus anchored to a distribution that was never passed through a generator, so there is nothing to concentrate |
| SOTA-295 | parameterize the generating prompt above the word list, and constrain how each sample opens | if you must generate, the entropy has to be supplied from outside the model, because the model will not supply it itself |
The two practices disagree about the remedy and agree about the disease. That agreement is what this document is for: it was stated in prose in SOTA-295’s conditions and had no page of its own.
The account
Sampling from a model is sampling from a distribution the model learned, and that distribution is not the distribution it was trained on. It is sharper. Whatever the training data contained, the model puts more mass on the continuations it is most confident about, and every sample is a draw from the sharpened version. Generate a corpus this way and the corpus inherits the sharpening; train on that corpus and the next model sharpens what was already sharp.
The statistic where this becomes visible is n-gram frequency, and both sources land on it independently.
LIT-205 pretrains across varying proportions of synthetic data and finds the proportion correlates negatively with performance, then attributes it statistically to distributional shift and over-concentration of n-gram features. The remedy follows from the diagnosis rather than from intuition: edit human text at the token level, and the test error is bounded above by construction because the result never leaves the neighbourhood of a real distribution.
LIT-485 measures a corpus rather than a model trained on one, and finds
the same concentration with a battery of instruments. 59.38% of
TinyStories contains “once upon a time”; the next three 4-grams are 28.24%,
16.52% and 11.55%. Every sample contains one of the hundred commonest
part-of-speech sequences — a POS-template rate of 100. Compression ratio
and Self-BLEU both say the same thing at p < 0.001.
What makes the second source more than a repetition is what it rules out. TinyStories was not careless. It had already identified this failure and built a mechanism against it: sample a lexicon per story so the prompt differs every time. That mechanism was lexical, and lexical variation turns out to be the wrong level — the openings stayed near-identical because the shape of the request never changed. So the narrowing is not merely “the prompt was the same”. A prompt distribution designed to be diverse still produced a corpus whose top 4-gram covers three fifths of it.
The trace reaches into a model trained on it: in a single-layer model trained
on TinyStories, [bos, 'once'] is the strongest bigram outlier in the
weights.
A second narrowing, and why this document does not claim it is the same one
The record holds another account of a text distribution narrowing. THEORY-051 describes Change.org’s AI drafting tool raising mean pairwise similarity between petitions by 23%, with the text features that had predicted success ceasing to predict it. Same destination — a narrower distribution of text — reached by a different route: there a human accepts, edits and posts what a model suggested, where here the model emits the tokens directly.
It is tempting to call these one phenomenon, and this record declines to, for reasons it can state:
- The metrics are not the same measurement. N-gram over-concentration and mean pairwise embedding similarity move independently in general — paraphrase preserves one and destroys the other — and nobody has run both instruments on either corpus.
- The human-in-the-loop path has a selection step this one lacks. Writers accept, reject and rewrite. That could damp the narrowing or amplify it, and LIT-487 cannot tell: its treatment is access to the tool, not use of it.
- The harms are different in kind. Here the damage is to a model trained on the corpus. There the damage is that a signal between people stopped carrying information. Two things with the same statistics need not have the same remedy.
What would settle it is a measurement, not a reading: run both instruments across both boundaries — the AI-access boundary on Change.org and the synthetic/human boundary in TinyStories and SimpleStories. If n-gram concentration and embedding homogenization track each other, that is one phenomenon with two metrics. If they dissociate, these are two things that share a word. #234 is where that was argued out, and the conclusion was to file the two accounts separately and point them at each other rather than to assert a connection neither source claims. DP-009: two is not several, and it is certainly not a mechanism.
Why Proposed
Neither source isolates the model’s contribution. A generated corpus is narrow, and the narrowness has at least two possible owners: the model’s sharpened output distribution, and the prompt distribution that was fed to it. LIT-485 rules out the lexical part of the second by construction — TinyStories varied the lexicon and stayed formulaic — but nothing measures the concentration of the seed distribution itself as a floor. Until somebody runs one seed distribution through two generators, “the model concentrates” is the plausible reading rather than the measured one.
The mechanism is stated rather than derived, in both. LIT-205 attributes the effect to n-gram over-concentration by statistical analysis of outcomes; it does not show the sharpening step happening. LIT-485 measures the result and does not model its cause at all.
Two corpora, both narrow, both English, both short-form. TinyStories is a deliberately restricted distribution of children’s stories. Whether the same concentration appears in generated code, generated mathematics, or generated long-form text is untested here — and generated mathematics is exactly where the field currently generates most heavily.
What it does not say
It does not say generated data is useless. SOTA-295 exists because a generated corpus with the entropy supplied from outside is a good corpus, and it beats its predecessor on every diversity instrument with no significant difference in judged simplicity. The claim is about where the entropy has to come from, not about whether generation works.
It does not say the two remedies are in competition. They apply to different jobs — SOTA-172 where the goal is a pretraining corpus standing in for real text, SOTA-295 where generating from scratch is the point. Reading either as general advice about synthetic data gets both wrong.
It does not predict a rate. Nothing here says how much narrowing per generation, or after how many rounds it matters. The negative correlation in LIT-205 is measured at that paper’s scale, and the recipes in this record that rephrase at frontier scale are not covered by it.