Large language models answer questions about a list of facts based on where each fact sits in the list, not on the name mentioned in the question.
Researchers tested models including Qwen, Gemma, and Llama by feeding them short fact lists, such as "Alice eats an apple. Bob eats a pear," and tracking how each model's internal state shifted as questions moved from the first fact to the second. That shift turned out to be consistent across many lists, so the team isolated it as a single internal pattern they call an ordinal vector. Injecting that vector into a question about the first fact in a brand-new list made the model answer with the second fact instead, regardless of which names were involved. The pattern showed up in a shared region of each model's middle-to-late layers, held across sizes from 1.5 billion to 32 billion parameters, and was already in place early in training.
That's a useful clue for anyone debugging why models garble who-said-what in long documents or chat transcripts: the model isn't filing facts under names, it's filing them under slots, and names get looked up after the fact. It also means the first thing mentioned in a context window carries a built-in recall advantage, not just a side effect of attention or recency.
The paper even notes this slot-based recall looks "surprisingly similar" to human memory, where first-mentioned items also stick best. A neat parallel, but worth remembering this is lab-bench interpretability work on toy sentences - real documents are messier, and nobody has shipped a product fix based on an ordinal vector yet.