💬 NLP

Pretraining Exposure Explains Popularity Judgments in Large Language Models

By leveraging the fully observable Dolma corpus and OLMo models, this study demonstrates that large language models' popularity judgments are primarily driven by pretraining exposure statistics rather than external real-world popularity signals, with exposure serving as a more accurate predictor of model bias than Wikipedia pageviews.

Original authors: Jamshid Mozafari, Bhawna Piryani, Adam Jatowt

Published 2026-05-13

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Jamshid Mozafari, Bhawna Piryani, Adam Jatowt

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a giant, digital brain (a Large Language Model, or LLM) that has read almost everything on the internet. You ask this brain, "Who is more famous: a specific local bakery or a world-famous pop star?" The brain almost always says the pop star. We've long assumed this is because the brain "knows" the pop star is popular in the real world.

But this paper asks a different question: Is the brain actually judging real-world fame, or is it just remembering which names it saw the most often while it was studying?

To find the answer, the researchers used a special, fully open-source brain called OLMo. Unlike most AI brains, we can see exactly what books and websites OLMo read during its "schooling" (pretraining). This is like having the entire library of a student's homework, rather than just guessing what they read based on their test scores.

Here is the breakdown of their experiment and findings, using some simple analogies:

1. The Setup: Counting the "Glances"

The researchers treated the AI's training data like a massive library. They used a special tool (Infini-Gram) to count exactly how many times 2,000 different things (people, places, brands, songs) appeared in the books the AI read.

The "Exposure" Score: This is just a count of how many times the AI "glanced" at a name. If the AI saw "Taylor Swift" 1 million times and "Local Baker Bob" 5 times, Taylor Swift has a high exposure score.

2. The Three Competitors

The researchers wanted to see which "score" best predicted how the AI would answer popularity questions. They compared three things:

The Real-World Score (Wikipedia): They looked at how many people actually visited the Wikipedia pages for these entities. This is the "ground truth" of real-world fame.
The "Direct" Guess: They asked the AI, "On a scale of 0 to 1000, how famous is Taylor Swift?"
The "Head-to-Head" Matchup: They asked the AI, "Who is more famous: Taylor Swift or Local Baker Bob?" (This is like a tournament bracket).

3. The Big Discovery: It's About What Was Read, Not What Is Known

The results were surprising and clear:

The AI's "Memory" matches the Library: The AI's idea of what is famous matched its "Exposure Score" (how often it saw the name) much better than it matched the real-world Wikipedia numbers.
The "Head-to-Head" is the Best Mirror: When the AI was asked to compare two things directly (like a tournament), its answers lined up almost perfectly with how often it had seen those names in its training books. It was even a better match than the real-world Wikipedia data.
The "Long Tail" Problem: For very obscure things (like "Local Baker Bob"), Wikipedia data is often messy or non-existent (nobody visits those pages). However, the AI still had a clear opinion on who was more famous. Why? Because even if the AI only saw "Local Baker Bob" 5 times and "Another Baker" 2 times, it knew the first one was "more famous" simply because it saw the name slightly more often. The AI wasn't using real-world knowledge; it was just counting its homework.

4. Bigger Brains = Stronger Memories

The researchers tested a smaller brain (7 billion parameters) and a bigger brain (32 billion parameters).

The bigger brain was even more obsessed with its training data. Its answers aligned even more closely with the "Exposure Score" than the smaller brain did.
This suggests that as AI gets smarter and bigger, it doesn't necessarily become more "aware" of the real world; it just becomes better at remembering exactly what it was fed during training.

The Bottom Line

The paper concludes that when an AI tells you who is famous, it isn't necessarily giving you a report on real-world society. It is giving you a report on its own reading list.

Think of it like a student who has only read one specific magazine. If you ask that student, "Who is the most famous person in the world?" they won't tell you the actual most famous person; they will tell you the person who appeared on the cover of that magazine the most times. The AI's "popularity bias" is simply a reflection of the frequency of words in its training data, not a true understanding of the world.

Technical Summary: Pretraining Exposure Explains Popularity Judgments in Large Language Models

Problem Statement
Large Language Models (LLMs) exhibit systematic preferences for well-known entities, a phenomenon widely attributed to "popularity bias." However, the origin of these preferences remains ambiguous: it is unclear whether they reflect genuine real-world popularity or merely the statistical frequency of entities within the pretraining corpus. Previous research has struggled to resolve this because the training data for most released LLMs is inaccessible, forcing reliance on indirect evidence or downstream behaviors that conflate exposure with other confounding factors.

Methodology
This study provides the first direct, large-scale analysis of popularity bias by leveraging the fully open OLMo models and their complete pretraining corpus, Dolma. The methodology proceeds through the following steps:

Data Indexing and Exposure Computation: The authors indexed the entire OLMo pretraining corpus (comprising Pretraining, Midtraining, and Long Context datasets, totaling 7.4 trillion tokens) using the Infini-Gram toolkit.
Entity Sampling and Validation: From a Wikidata dump (dated January 24, 2026), 5,000 entities were sampled for each of five types: Person, Location, Organization, Art, and Product. To ensure accurate frequency counting, all candidate aliases were validated using GPT-4.1 to filter out ambiguous references, retaining only those that refer exclusively to the target entity.
Exposure Scoring: For each of the 2,000 selected entities (400 per type, split evenly between 200 low-frequency "sparse" and 200 high-frequency "popular" entities), an exposure score was computed based on the frequency of the entity and its validated aliases in the indexed corpus.
Popularity Signal Comparison: The study compared three distinct signals against the pretraining exposure scores:
- Wikipedia Pageviews: An external proxy for real-world popularity, aggregated from July 1, 2015, to December 31, 2024 (matching the corpus cutoff).
- Direct Scalar Estimation: The OLMo model (7B and 32B variants) was prompted to assign a popularity score (0–1000) to individual entities.
- Pairwise Comparison: The model was prompted to determine which of two entities was more popular. These pairwise preferences were converted into listwise scores using the Bradley–Terry model.

Key Results
The analysis, based on Spearman rank correlations ( $\rho$ ), yielded the following findings:

Exposure as a Proxy for Real-World Popularity: Pretraining exposure strongly correlates with Wikipedia pageviews across all entity types (average $\rho \approx 0.755$ ), validating exposure as a meaningful proxy for real-world salience during the training period.
LLM Judgments Align More with Exposure than Wikipedia: LLM-generated popularity signals align more closely with pretraining exposure than with Wikipedia pageviews. This is particularly evident in the Pairwise Comparison method, which achieved the highest correlation with exposure (average $\rho \approx 0.795$ for the 32B model), outperforming both the Direct scalar estimation and the Wikipedia baseline.
The Long-Tail Regime: For "sparse" (low-frequency) entities, Wikipedia pageviews become unreliable, showing near-zero or negative correlations with exposure ( $\rho \approx 0.10$ ). In contrast, LLM popularity judgments (especially via pairwise comparison) maintain a positive correlation with exposure ( $\rho \approx 0.31$ for the 32B model), suggesting that for rare entities, model priors are governed almost entirely by pretraining frequency.
Model Scale Effects: Increasing model scale from 7B to 32B systematically improves the alignment between LLM judgments and pretraining exposure. The 32B model demonstrates sharper, more consistent exposure-based priors, particularly in direct scalar estimation tasks.

Significance and Claims
The paper claims to offer the first direct, large-scale evidence that popularity priors in LLMs are primarily shaped by pretraining statistics rather than external popularity signals. The authors argue that "data exposure" plays a central role in driving popularity bias. Specifically, the findings demonstrate that:

LLMs do not merely reflect external popularity metrics; their internal ranking preferences are deeply rooted in the statistical distribution of their training data.
Pairwise elicitation is a more reliable method for uncovering these exposure-based priors than direct scalar estimation.
The "popularity bias" observed in LLMs is not necessarily a failure of reasoning but a direct reflection of the frequency patterns inherent in the pretraining corpus.

The authors conclude that these biases are measurable properties of training data and suggest that future work should investigate mitigation strategies to improve fairness and diversity in model outputs.