"Olmo uniquely enabled this work because it is fully open source. Its value lies in being a research-first, fully transparent foundation model." — Yuan He, Former Research Associate at the University of Oxford

Consider this pair of sentences: “Diego Maradona has a sibling named Raúl Maradona” and “Raúl Maradona has a sibling named Diego Maradona.” To us, there’s no difference; they’re two ways of stating the same fact. But for LLMs, which name comes first can matter a lot.

A team led by Yuan He at Oxford found that language models often prefer one ordering over the other—and that the preference tracks what the model saw during pre-training. Because Olmo’s pre-training data is fully open, it made for the perfect window into the training process. The researchers could count how often names appeared in the actual training set, Dolma, and compare those counts to the model’s judgments.

“Our work introduces a framework for evaluating whether LLMs truly understand a fact, rather than simply memorizing a particular phrasing,” He says. “The key idea is to probe models with logically equivalent variations of a fact – different surface forms that should carry the same meaning – and then measure consistency.”

To keep things fair, they looked at symmetric relationships—cases that should hold in both directions. If Alice is Bob’s sibling, Bob is Alice’s sibling; if Town A borders Town B, Town B borders Town A. Starting from structured facts in a public dataset, the team turned many thousands of these pairs into plain-English prompts and checked whether the model treated each direction the same.

Here’s what they found in simple terms: When a sentence starts with a very common name or place and ends with a rarer one, the model is more likely to accept it as true. Flip the order – put the rarer name first and the common one second – and the model’s confidence drops. When both names are common, that gap largely disappears.

“We find that LLMs often fail to treat equivalent facts consistently, and we trace one root cause to the pre-training distribution,” He says. “This reveals gaps in semantic reasoning.”

Because Olmo’s training data is public, the researchers could link those preferences to real frequency counts in the corpus rather than guessing. They also showed working examples on Olmo models including Olmo-2-13B, making the data-to-behavior link concrete.

What seems like a small quirk has practical stakes. People often describe “hallucinations” as if they come from nowhere. This study shows a mechanism behind some of them. Uneven exposure during pre-training can subtly tilt a model toward one phrasing of a fact and away from its mirror image.

“Olmo uniquely enabled this work because it is fully open source,” He says. “Its value lies in being a research-first, fully transparent foundation model. It lowers the barrier to rigorous scientific investigation, such as interpretability and reliability studies that demand fine-grained access to model checkpoints and training data.”

Once you can measure that tilt, you can start to correct it—by rebalancing the data, adjusting prompts, or adding checks for cases where frequency gaps are likely to mislead. And because Olmo exposes both the models and the corpus, anyone can reproduce the analysis and test whether a fix really works. That’s the advantage of an open research stack over a closed, proprietary system.

“Olmo offers an open and credible platform where the community can investigate fundamental questions about reasoning, bias, fairness, and trustworthiness,” He says. “In a landscape dominated by commercially optimized systems, Olmo stands out by empowering deeper understanding rather than just application.”