Measuring and Explaining Creativity in Large Language Models: A Mixed-Methods Investigation into the Effect of Cultural Persona Prompts on Business Solution Generation - Youssef Hariri - Doctoral student in Business of AI - Rennes School of Business and UpGrad

This repository contains the data, prompts, and results for the research paper: "Measuring and Explaining Creativity in Large Language Models: A Mixed-Methods Investigation into the Effect of Cultural Persona Prompts on Business Solution Generation (soon to be available online).

The files are organized into three categories:

1. Input Data: The datasets used as stimuli for the LLM.
2. Prompts: The evaluation prompts used to score the LLM's output, showing the optimization process.
3. Results: The raw data from the quantitative and qualitative analysis.

---

📜 The Paper

• ** business_problems_data.json Measuring and Explaining Creativity in LLMs - Hariri Youssef.pdf **

---

1. Input Data

• business_problems_data.json
  • A JSON file containing the set of standardized business problems that were presented to the LLM.
• cultural_personas_data.json
  • A JSON file defining the "cultural personas" (e.g., Innovator, Pragmatist, Traditionalist) used to prime the LLM.
• 125 real world examples.txt
  • A text file containing the raw 125 real-world business examples that were used to inspire or synthesize the final business_problems_data.json.

---

2. Prompts (Methodology)

This folder shows the iterative refinement of the prompt used to evaluate the creativity of the LLM's solutions.

• evaluation_prompt_original.md
  • The original, baseline (v0) prompt used for automated evaluation.
• evaluation_prompt_optimization_1.md
  • The first iteration (v1) of the refined evaluation prompt.
• evaluation_prompt_optimization_2.md
  • The second and final iteration (v2) of the evaluation prompt used to generate the scores in the paper.

---

---

3. Results & Analysis

• all_experiment_solutions.json

  • A JSON file containing the complete, raw text output of all solutions generated by the LLM for every problem and persona combination. Each entry links a problem_id, persona_id, and solution_id to the full text of the generated solution.

• cleaned_evaluation_scores.csv

  • This is the main quantitative results file (tracked with Git LFS). It contains the raw, structured scores for all generated solutions, mapping each solution_id to its evaluated creativity components: originality, practicality/feasibility, and innovation/novelty.

• ensemble_scores_FULL_DATA_6375.csv

  • The complete quantitative dataset. This CSV file contains all 6,375 scored solutions generated by the LLM, with scores for creativity, feasibility, novelty, etc., across all personas and problems.

• selected_solutions_for_qualitative_study.csv

  • A subset of the main dataset, this CSV lists the specific solutions that were selected for the in-depth qualitative (human) analysis.

• selected_solutions_justifications_qualitative_study.csv (The corrected filename)

  • The results of the qualitative analysis. This file contains the human-provided justifications, coding, and analysis for the solutions listed above.

---

How to Cite

@software{youssef_hariri_2025_17407392,
  author       = {Hariri, Youssef},
  title        = {Measuring and Explaining Creativity in Large Language Models: A Mixed-Methods Investigation into the Effect of Cultural Persona Prompts on Business Solution Generation},
  month        = oct,
  year         = 2025,
  publisher    = {Zenodo},
  version      = {v1.0.0},
  doi          = {10.5281/zenodo.17407392},
  url          = {https://doi.org/10.5281/zenodo.17407392}
}