Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Comprehensive Analysis of Large Language Model Outputs: Similarity, Diversity, and Bias

Brandon Smith, Mohamed Reda Bouadjenek, Tahsin Alamgir Kheya, Phillip Dawson, Sunil Aryal

TL;DR

The paper conducts a large-scale empirical analysis of LLM outputs across 12 models using ~5{,}015 prompts and ~3 million texts to investigate inner- and inter-model similarities, variances in generation, authorship detection, language markers, and bias/ethics. It employs cosine and edit-distance-based similarity metrics, stylometric readability measures, classifier-based authorship detection (BERT/DeBERTa, XGBoost), PMI-driven language marker analysis, and embedding-based bias subspace methods with Word2Vec to quantify gender and racial biases. Key findings include: (i) same-model outputs are more similar to each other than to human texts; (ii) GPT-4 shows high lexical diversity and low self-similarity; (iii) models like Gemma-7B and Gemini-pro tend to be more balanced with respect to bias; and (iv) robust classification can distinguish human from LLM text and often identify the specific model, though close-model confusions (e.g., GPT-3.5 vs GPT-4) persist. These insights advance understanding of LLM behavior, diversity, and biases, informing future model design, evaluation, and ethical considerations along with methods for model attribution and bias mitigation.

Abstract

Large Language Models (LLMs) represent a major step toward artificial general intelligence, significantly advancing our ability to interact with technology. While LLMs perform well on Natural Language Processing tasks -- such as translation, generation, code writing, and summarization -- questions remain about their output similarity, variability, and ethical implications. For instance, how similar are texts generated by the same model? How does this compare across different models? And which models best uphold ethical standards? To investigate, we used 5{,}000 prompts spanning diverse tasks like generation, explanation, and rewriting. This resulted in approximately 3 million texts from 12 LLMs, including proprietary and open-source systems from OpenAI, Google, Microsoft, Meta, and Mistral. Key findings include: (1) outputs from the same LLM are more similar to each other than to human-written texts; (2) models like WizardLM-2-8x22b generate highly similar outputs, while GPT-4 produces more varied responses; (3) LLM writing styles differ significantly, with Llama 3 and Mistral showing higher similarity, and GPT-4 standing out for distinctiveness; (4) differences in vocabulary and tone underscore the linguistic uniqueness of LLM-generated content; (5) some LLMs demonstrate greater gender balance and reduced bias. These results offer new insights into the behavior and diversity of LLM outputs, helping guide future development and ethical evaluation.

A Comprehensive Analysis of Large Language Model Outputs: Similarity, Diversity, and Bias

TL;DR

The paper conducts a large-scale empirical analysis of LLM outputs across 12 models using ~5{,}015 prompts and ~3 million texts to investigate inner- and inter-model similarities, variances in generation, authorship detection, language markers, and bias/ethics. It employs cosine and edit-distance-based similarity metrics, stylometric readability measures, classifier-based authorship detection (BERT/DeBERTa, XGBoost), PMI-driven language marker analysis, and embedding-based bias subspace methods with Word2Vec to quantify gender and racial biases. Key findings include: (i) same-model outputs are more similar to each other than to human texts; (ii) GPT-4 shows high lexical diversity and low self-similarity; (iii) models like Gemma-7B and Gemini-pro tend to be more balanced with respect to bias; and (iv) robust classification can distinguish human from LLM text and often identify the specific model, though close-model confusions (e.g., GPT-3.5 vs GPT-4) persist. These insights advance understanding of LLM behavior, diversity, and biases, informing future model design, evaluation, and ethical considerations along with methods for model attribution and bias mitigation.

Abstract

Large Language Models (LLMs) represent a major step toward artificial general intelligence, significantly advancing our ability to interact with technology. While LLMs perform well on Natural Language Processing tasks -- such as translation, generation, code writing, and summarization -- questions remain about their output similarity, variability, and ethical implications. For instance, how similar are texts generated by the same model? How does this compare across different models? And which models best uphold ethical standards? To investigate, we used 5{,}000 prompts spanning diverse tasks like generation, explanation, and rewriting. This resulted in approximately 3 million texts from 12 LLMs, including proprietary and open-source systems from OpenAI, Google, Microsoft, Meta, and Mistral. Key findings include: (1) outputs from the same LLM are more similar to each other than to human-written texts; (2) models like WizardLM-2-8x22b generate highly similar outputs, while GPT-4 produces more varied responses; (3) LLM writing styles differ significantly, with Llama 3 and Mistral showing higher similarity, and GPT-4 standing out for distinctiveness; (4) differences in vocabulary and tone underscore the linguistic uniqueness of LLM-generated content; (5) some LLMs demonstrate greater gender balance and reduced bias. These results offer new insights into the behavior and diversity of LLM outputs, helping guide future development and ethical evaluation.
Paper Structure (18 sections, 3 equations, 10 figures, 4 tables)

This paper contains 18 sections, 3 equations, 10 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Dataset and LLMs Overview
  4. Large Language Models
  5. Data Description
  6. Empirical Evaluation
  7. RQ1: Comparative Analysis of LLM Texts
  8. Text Similarity Analysis
  9. Stylometric Analysis
  10. RQ2: Variance in Text Generation
  11. RQ3: Classification Performance
  12. RQ4: Language Markers Analysis
  13. RQ5: Bias and Ethics in LLMs
  14. Methodology
  15. Identifying the Bias Subspace
  16. ...and 3 more sections

Figures (10)

  • Figure 1: Distribution of first words in prompts used to create the dataset used in this analysis.
  • Figure 2: Comparison of inner-text similarity.
  • Figure 3: Comparaison of (average) inter-text similarity. The lower part (blue) displays the similarity calculated using word-level edit distance, while the upper part (orange) illustrates the similarity determined by cosine similarity.
  • Figure 4: Readability Statistics.
  • Figure 5: UMAP projection of high-dimensional text generated into a two-dimensional space for visualization.
  • ...and 5 more figures