One Injected Passage, Learned Then Lost by Training's End

One Injected Passage Moved 32 Models Measurably. By the End of Training It Was Statistically Gone.

A seed-controlled counterfactual finds a large geometric footprint with almost no functional consequence — and a caution for anyone auditing models by weight deltas.

Aug 20, 2026Source: arXiv

Zachary Speck and Asa Shepard posted "Learned, Then Lost: A Measured Single-Example Counterfactual in Pre-training" to arXiv at 17:46:08 UTC on 19 August. It asks a question that usually gets argued from intuition: does one document, seen once, stick?

The experiment

32 GPT-2-scale models at 124M parameters trained on OpenWebText — 24 experiments, four conditions across eight seeds. At training step 200 of 9,536 a 194-token passage was injected into exactly one batch row, under three conditions (fluent prose with a real subject; fluent prose with a fabricated subject; random characters) plus an uninjected control.

What happened

Fifty steps after injection the model held a measurable advantage on that passage: 0.039 to 0.044 nats of cross-entropy, significant at p < 10^-4 in eight of eight seeds. By the final training step it was gone — p = 0.25, 0.71 and 0.54, with the interpolation loss barrier at p = 0.509 and held-out cross-entropy at p = 0.310. The authors' own summary: "The passage is learned from one exposure and then decays."

The finding underneath the finding

The injection displaced the weights by 44.1% of the Euclidean distance between independently seeded models, while producing a loss barrier of only 3.0% of the seed-to-seed barrier, with 92% settlement by the midpoint. A large geometric footprint, almost no functional consequence. That gap is a direct caution to anyone auditing models by measuring weight deltas: weight-space distance is a poor proxy for what a model has learned.

What the common telling gets wrong

Two opposite readings are both wrong. This is not "single documents get memorised" — the effect vanished. And it is not "training data doesn't matter" or "memorisation is a myth," which is how a null result gets flattened as it is aggregated. What the paper shows is narrower and more useful than either. It is also unavoidably true that GPT-2 at 124M is four orders of magnitude below frontier scale, and this is a preprint with no peer review. Nothing here licenses a claim about whether a frontier model retains a single document.

Why It Matters

Single-example memorisation sits underneath live copyright litigation, privacy law and unlearning research, and almost all of that argument runs on intuition. This is a cheap, seed-controlled experiment that puts a p-value on the question — and answers no, at this scale.