Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

UnWEIRDing LLM Entity Recommendations

Aayush Kumar, Sanket Mhatre

TL;DR

The paper investigates cultural biases in LLM-generated recommendations for real-world entities using the WEIRD framework. It evaluates pluralistic prompting strategies across nine models and 44 fine-grained entity types, with a WEIRD-score-based metric to quantify bias. The results show that prompting generally reduces WEIRD bias but effects vary by model and entity type, with some categories like Product remaining highly biased and geographic concentration around the United States persisting. This work highlights the limits of prompt-based mitigation and motivates development of more holistic bias-mitigation strategies to ensure cultural plurality in LLM outputs.

Abstract

Large Language Models have been widely been adopted by users for writing tasks such as sentence completions. While this can improve writing efficiency, prior research shows that LLM-generated suggestions may exhibit cultural biases which may be difficult for users to detect, especially in educational contexts for non-native English speakers. While such prior work has studied the biases in LLM moral value alignment, we aim to investigate cultural biases in LLM recommendations for real-world entities. To do so, we use the WEIRD (Western, Educated, Industrialized, Rich and Democratic) framework to evaluate recommendations by various LLMs across a dataset of fine-grained entities, and apply pluralistic prompt-based strategies to mitigate these biases. Our results indicate that while such prompting strategies do reduce such biases, this reduction is not consistent across different models, and recommendations for some types of entities are more biased than others.

UnWEIRDing LLM Entity Recommendations

TL;DR

The paper investigates cultural biases in LLM-generated recommendations for real-world entities using the WEIRD framework. It evaluates pluralistic prompting strategies across nine models and 44 fine-grained entity types, with a WEIRD-score-based metric to quantify bias. The results show that prompting generally reduces WEIRD bias but effects vary by model and entity type, with some categories like Product remaining highly biased and geographic concentration around the United States persisting. This work highlights the limits of prompt-based mitigation and motivates development of more holistic bias-mitigation strategies to ensure cultural plurality in LLM outputs.

Abstract

Large Language Models have been widely been adopted by users for writing tasks such as sentence completions. While this can improve writing efficiency, prior research shows that LLM-generated suggestions may exhibit cultural biases which may be difficult for users to detect, especially in educational contexts for non-native English speakers. While such prior work has studied the biases in LLM moral value alignment, we aim to investigate cultural biases in LLM recommendations for real-world entities. To do so, we use the WEIRD (Western, Educated, Industrialized, Rich and Democratic) framework to evaluate recommendations by various LLMs across a dataset of fine-grained entities, and apply pluralistic prompt-based strategies to mitigate these biases. Our results indicate that while such prompting strategies do reduce such biases, this reduction is not consistent across different models, and recommendations for some types of entities are more biased than others.

Paper Structure

This paper contains 21 sections, 2 equations, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. LLM-based Writing Completions
  4. Cultural Biases in LLMs
  5. Methodology
  6. Task Selection and Design
  7. Entity Type Framework
  8. Question Construction
  9. Prompting Techniques
  10. Models Evaluated and Response Generation
  11. WEIRD Score Calculation
  12. Response-Level WEIRD Score
  13. Results
  14. RQ1: Effect of Prompting on Recommendation Biases
  15. Overall Impact of Prompting Techniques
  16. ...and 6 more sections

Figures (4)

  • Figure 1: Comparison of average WEIRD percentage by prompting technique across all models.
  • Figure 2: Baseline WEIRD percentage by model with Base Diversity prompting.
  • Figure 3: Country-level representation showing top 30 countries. Blue bars indicate WEIRD countries, green bars indicate non-WEIRD countries.
  • Figure 4: WEIRD bias by entity category comparing baseline (without guidance, blue) and best-performing prompting technique (with guidance, green).