Pedro Dias asks if more content makes AI models recall a brand

A pupil who can recite a textbook passage will still fail a question whose answer sits inside that passage.

That is the frame Pedro Dias, an independent consultant at Visively, uses in Search Engine Journal to question a quiet claim behind many "get your brand into model memory" plans: that enough content, published often enough, teaches a commercial model to recommend your product. Dias works through four research papers and makes the opposite case. Each paper measures a gain on one task, and the gain holds only for that task; a related task sitting next to it can see a different result or none at all.

The pupil who could recite the lesson

The first paper, Physics of Language Models: Part 3.1, trained small models from scratch on fictional biographies, then fine-tuned them on questions about some people and tested questions about the others. The model had seen every test person's biography but had never been trained on questions about those people. It could complete the passages and still fail the questions.

In the same experiment, varying the wording, shuffling sentence order and repeating each person's full name throughout the biography moved accuracy on factual questions in a 124-million-parameter GPT-2 model from 9.7 percent to 96.2 percent. Dias notes those were synthetic facts in a small training experiment, not commercial brand recommendations, and that the researchers both chose the training material and did the training themselves. A publisher puts pages online; a model provider decides whether those pages enter training, which copies to keep and how to use them.

How repetition changes what a model can produce

A second paper, Quantifying Memorization Across Neural Language Models, found that duplication in a training corpus does raise verbatim recall: give the model the start of a repeated passage and it is more likely to continue it. A third, Don't Forget! Decomposing the Training Dynamics of Memorization in Language Models, used saved Pythia model snapshots to simulate training updates and found that learning from other data worked against predicting frequently repeated memorised passages more strongly than rare ones, especially in a one-billion-parameter model. Being able to finish the passage still leaves open whether the model can use the facts to answer a different kind of question.

The group whose biographies were unchanged

In the biography experiment, one group's biographies were left as they were while a second group's biographies received the varied versions. The later question-answer training used examples about only the varied group. Accuracy on questions about the unchanged group rose from 4.4 percent to 86.8 percent. Training on material about other people had helped the model answer questions about this group.

Dias points at this as the finding that makes a neat brand-content explanation incomplete. The quality of writing about every brand may matter, as may the model's ability to use that kind of information at all, developed from training on other examples. Your brand could become easier to recall after a model change without anyone having rewritten your brand material.

When retrieval beats memory

A fourth paper, Fine-Tuning or Retrieval?, scored a Mistral-7B model at 50.4 percent on a new-fact multiple-choice test after training on the original documents and 58.8 percent with paraphrased versions. Giving the model relevant passages at answer time scored 87.5 percent. Dias also cites Source Preference in the Wild, which held titles and snippets fixed, changed the source shown with each one, and watched the models' choices change. Your product can pass a recall test while a competing result gets picked for a reason the test never examined.

Source

It Knew Every Word by Pedro Dias, Search Engine Journal.