Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

On the Credibility of Evaluating LLMs using Survey Questions

Jindřich Libovický

TL;DR

This paper critiques the credibility of evaluating LLMs by matching survey-question responses to human data, showing that prompting, decoding, and cross-question dependencies can drastically alter conclusions. It introduces Self-Correlation Distance to quantify structural alignment beyond per-question accuracy and compares direct vs Chain-of-Thought prompts under greedy and sampling-based decoding across World Value Survey data in multiple languages and countries. The results reveal that CoT prompting with sampling generally improves surface alignment but may degrade structural alignment, and that independence-based metrics can overestimate true similarity. The work advocates a robust, multi-metric evaluation framework—including self-correlation analyses and large sample sizes—for more reliable cross-cultural value alignment assessments of LLMs, with practical implications for benchmarking and policy-related AI evaluation.

Abstract

Recent studies evaluate the value orientation of large language models (LLMs) using adapted social surveys, typically by prompting models with survey questions and comparing their responses to average human responses. This paper identifies limitations in this methodology that, depending on the exact setup, can lead to both underestimating and overestimating the similarity of value orientation. Using the World Value Survey in three languages across five countries, we demonstrate that prompting methods (direct vs. chain-of-thought) and decoding strategies (greedy vs. sampling) significantly affect results. To assess the interaction between answers, we introduce a novel metric, self-correlation distance. This metric measures whether LLMs maintain consistent relationships between answers across different questions, as humans do. This indicates that even a high average agreement with human data, when considering LLM responses independently, does not guarantee structural alignment in responses. Additionally, we reveal a weak correlation between two common evaluation metrics, mean-squared distance and KL divergence, which assume that survey answers are independent of each other. For future research, we recommend CoT prompting, sampling-based decoding with dozens of samples, and robust analysis using multiple metrics, including self-correlation distance.

On the Credibility of Evaluating LLMs using Survey Questions

TL;DR

This paper critiques the credibility of evaluating LLMs by matching survey-question responses to human data, showing that prompting, decoding, and cross-question dependencies can drastically alter conclusions. It introduces Self-Correlation Distance to quantify structural alignment beyond per-question accuracy and compares direct vs Chain-of-Thought prompts under greedy and sampling-based decoding across World Value Survey data in multiple languages and countries. The results reveal that CoT prompting with sampling generally improves surface alignment but may degrade structural alignment, and that independence-based metrics can overestimate true similarity. The work advocates a robust, multi-metric evaluation framework—including self-correlation analyses and large sample sizes—for more reliable cross-cultural value alignment assessments of LLMs, with practical implications for benchmarking and policy-related AI evaluation.

Abstract

Recent studies evaluate the value orientation of large language models (LLMs) using adapted social surveys, typically by prompting models with survey questions and comparing their responses to average human responses. This paper identifies limitations in this methodology that, depending on the exact setup, can lead to both underestimating and overestimating the similarity of value orientation. Using the World Value Survey in three languages across five countries, we demonstrate that prompting methods (direct vs. chain-of-thought) and decoding strategies (greedy vs. sampling) significantly affect results. To assess the interaction between answers, we introduce a novel metric, self-correlation distance. This metric measures whether LLMs maintain consistent relationships between answers across different questions, as humans do. This indicates that even a high average agreement with human data, when considering LLM responses independently, does not guarantee structural alignment in responses. Additionally, we reveal a weak correlation between two common evaluation metrics, mean-squared distance and KL divergence, which assume that survey answers are independent of each other. For future research, we recommend CoT prompting, sampling-based decoding with dozens of samples, and robust analysis using multiple metrics, including self-correlation distance.
Paper Structure (21 sections, 2 figures, 7 tables)

This paper contains 21 sections, 2 figures, 7 tables.

Table of Contents

  1. Introduction
  2. LLMs and Surveys
  3. Experiments
  4. Model Prompting
  5. Questionnaire design.
  6. Scores vs. Chain-of-Thought.
  7. Greedy vs. Sampling.
  8. Evaluation
  9. Models.
  10. Mean Square Difference.
  11. KL Divergence.
  12. Self-Correlation Distance.
  13. Results
  14. Comparing MSD, KLD, and self-correlation distance.
  15. Cross-lingual and cross-cultural patterns.
  16. ...and 6 more sections

Figures (2)

  • Figure 1: Correlation patterns between human answers in the USA (under the diagonal) and between answers of the LLaMA 3 and Mistral 2 models in English (above the diagonal).
  • Figure 2: Mean-squared difference and KL-divergence of LLaMA 3 answers when compared to the USA data of the World Value Survey. It compares the greedy decoding and sampling from the model.