Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Scalable Entity-Based Framework for Auditing Bias in LLMs

Akram Elbouanani, Aboubacar Tuo, Adrian Popescu

TL;DR

The paper introduces a scalable, entity-based bias auditing framework for LLMs that leverages synthetic data to probe how model outputs vary with entity identity under controlled templates. It validates synthetic data against real benchmarks using Pearson correlations, enabling a large-scale audit of ~1.9 billion data points across multiple languages, models, and prompting configurations. The analysis uncovers systematic biases: Western and wealthy entities are favored, non-Western entities are penalized, and biases intensify with model scale while instruction tuning mitigates but does not reverse them. The framework enables rigorous auditing prior to deployment in high-stakes contexts and highlights the need for careful dataset design and cross-language considerations to reduce entrenched priors.

Abstract

Existing approaches to bias evaluation in large language models (LLMs) trade ecological validity for statistical control, relying on artificial prompts that poorly reflect real-world use, or on naturalistic tasks that lack scale and rigor. We introduce a scalable bias-auditing framework using named entities as probes to measure structural disparities in model behavior. We show that synthetic data reliably reproduces bias patterns observed in natural text, enabling large-scale analysis. Using this approach, we conduct the largest bias audit to date, comprising 1.9 billion data points across multiple entity types, tasks, languages, models, and prompting strategies. Our results reveal systematic biases: models penalize right-wing politicians, favor left-wing politicians, prefer Western and wealthy nations over the Global South, favor Western companies, and penalize firms in the defense and pharmaceutical sectors. While instruction tuning reduces bias, increasing model scale amplifies it, and prompting in Chinese or Russian does not attenuate Western-aligned preferences. These results indicate that LLMs should undergo rigorous auditing before deployment in high-stakes applications.

A Scalable Entity-Based Framework for Auditing Bias in LLMs

TL;DR

The paper introduces a scalable, entity-based bias auditing framework for LLMs that leverages synthetic data to probe how model outputs vary with entity identity under controlled templates. It validates synthetic data against real benchmarks using Pearson correlations, enabling a large-scale audit of ~1.9 billion data points across multiple languages, models, and prompting configurations. The analysis uncovers systematic biases: Western and wealthy entities are favored, non-Western entities are penalized, and biases intensify with model scale while instruction tuning mitigates but does not reverse them. The framework enables rigorous auditing prior to deployment in high-stakes contexts and highlights the need for careful dataset design and cross-language considerations to reduce entrenched priors.

Abstract

Existing approaches to bias evaluation in large language models (LLMs) trade ecological validity for statistical control, relying on artificial prompts that poorly reflect real-world use, or on naturalistic tasks that lack scale and rigor. We introduce a scalable bias-auditing framework using named entities as probes to measure structural disparities in model behavior. We show that synthetic data reliably reproduces bias patterns observed in natural text, enabling large-scale analysis. Using this approach, we conduct the largest bias audit to date, comprising 1.9 billion data points across multiple entity types, tasks, languages, models, and prompting strategies. Our results reveal systematic biases: models penalize right-wing politicians, favor left-wing politicians, prefer Western and wealthy nations over the Global South, favor Western companies, and penalize firms in the defense and pharmaceutical sectors. While instruction tuning reduces bias, increasing model scale amplifies it, and prompting in Chinese or Russian does not attenuate Western-aligned preferences. These results indicate that LLMs should undergo rigorous auditing before deployment in high-stakes applications.
Paper Structure (76 sections, 8 equations, 11 figures, 16 tables)

This paper contains 76 sections, 8 equations, 11 figures, 16 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Bias Measurement
  5. Bias Auditing Framework
  6. Entity Sets per Domain.
  7. Downstream Tasks.
  8. Textual Templates.
  9. Models and Languages.
  10. Inference Prompts.
  11. Synthetic vs. Real Dataset Alignment
  12. Results
  13. Validation of Synthetic Proxy
  14. General Bias Trends
  15. Politicians.
  16. ...and 61 more sections

Figures (11)

  • Figure 1: Entity-level audit results aggregated across tasks/languages/models/prompts that reflect LLM outputs under sentence templates with light supporting evidence. They should not be interpreted as ground-truth properties or normative judgments about entities. We use cross-task aggregation for compact visualization. Finer-grained results (§ \ref{['sec_results']}) represent the primary evidence of the study.
  • Figure 2: Bias Profiles for (a) politicians by alignment, (b) countries by GDP quantile, (c) companies by domain.
  • Figure 3: Bias Distribution by Language, Model Family, and Prompt Setting (columns) and entity types (Politicians, Countries, and Companies). Ellipses represent the 95% confidence intervals for bias scores.
  • Figure 4: Prompt template used for generating synthetic evaluation samples for the Credibility Assessment task.
  • Figure 5: Bias distribution across entity categories for supplementary models. The panels illustrate the deviation of bias scores for (a) political orientations ranging from far-left to far-right for politicians, (b) country GDP quantiles (Q1 (lowest) to Q5 (highest)) for countries, (c) industrial sectors of companies, and (d) geographical regions for companies.
  • ...and 6 more figures