Pretraining Exposure Explains Popularity Judgments in Large Language Models

Why do LLMs think something is popular? Mostly because of how often they saw it during training.

📌 Overview

Large Language Models (LLMs) exhibit systematic preferences for well-known entities, often referred to as popularity bias. In this work, we show that these preferences are primarily driven by pretraining exposure—the frequency with which entities appear in the training corpus—rather than external signals such as real-world popularity.

Using fully observable pretraining data, we conduct the first large-scale analysis that directly links:

• Pretraining exposure
• Wikipedia popularity
• LLM-generated popularity judgments

🗂 Repository Structure

.
├── comparison/        # Pairwise comparison prompting & aggregation
│   ├── majority.py    # Majority voting over repeated pairwise judgments
│   ├── merge.py       # Merge pairwise outputs
│   ├── prompting.py   # Pairwise prompting pipeline
│   └── run.sh         # Run pairwise experiments
│
├── dataset/           # Full dataset with all signals
│   └── dataset.json   # Entities, metadata, exposure, pageviews, model outputs
│
├── directly/          # Direct popularity estimation pipeline
│   ├── merge.py       # Merge repeated outputs
│   ├── prompting.py   # Direct estimation prompting
│   └── run.sh         # Run direct estimation
│
└── experiments/       # Evaluation and analysis
    ├── correlation.py
    └── pairwise_accuracy.py

📊 Dataset

We release a fully integrated, analysis-ready dataset that consolidates all signals used in the paper into a single resource.

The dataset contains 2,000 entities spanning five semantic types: Person, Location, Organization, Art, and Product.

Each entity is enriched with multiple complementary signals:

• structured knowledge (Wikidata metadata and entity type)

• linguistic coverage (validated aliases)

• real-world popularity (Wikipedia pageviews)

• training-derived statistics (pretraining exposure counts)

• model-based estimates:

  • direct scalar popularity predictions
  • pairwise comparison outcomes

Unlike typical benchmarks, this dataset is self-contained: it includes both raw inputs and computed signals, allowing researchers to:

• reproduce results without rerunning expensive pipelines
• directly study relationships between signals
• extend the analysis to new models or ranking methods

🚀 Getting Started

Setup

git clone https://github.com/DataScienceUIBK/Pretraining-Exposure-Popularity.git
cd Pretraining-Exposure-Popularity
pip install -r requirements.txt

🤖 Models

We evaluate two fully open large language models released by the Allen Institute for AI:

• OLMo-3-7B-Instruct https://huggingface.co/allenai/Olmo-3-7B-Instruct

• OLMo-3.1-32B-Instruct https://huggingface.co/allenai/Olmo-3.1-32B-Instruct

These models are fully open and provide access to their training data, enabling direct measurement of pretraining exposure.

🧪 Running Experiments

🔹 Direct Estimation

cd directly
bash run.sh

🔹 Pairwise Comparison

cd comparison
bash run.sh

🔹 Analysis

cd experiments
python correlation.py
python pairwise_accuracy.py

🔁 Reproducibility

This repository is designed for full reproducibility of the reported results.

What is included

• Complete dataset with all signals (exposure, Wikipedia, LLM outputs)
• Prompting pipelines for direct and pairwise estimation
• Aggregation scripts (majority voting, merging)
• Evaluation scripts (correlation and accuracy)

Requirements

• Python 3.11+
• Dependencies listed in requirements.txt
• Access to GPT models (required for parts of the pipeline)
• GPU recommended for efficient inference

Notes

• Results may vary slightly due to LLM stochasticity
• Multiple runs and aggregation are used for stability
• All evaluated models are publicly available

📊 Key Findings

• Pretraining exposure strongly correlates with real-world popularity
• LLM judgments align more with exposure than Wikipedia signals
• Pairwise comparison yields the most reliable estimates
• Effects persist in long-tail entities

📄 Citation

📜 License

MIT License — see LICENSE.