Same Prompt, Different Outcomes: Evaluating the Reproducibility of Data Analysis by LLMs

TL;DR

The reproducibility of data analysis conducted by Large Language Models can vary, even given the same task, data, and settings, and this suggests if an LLM is being used to conduct data analysis, then it should be run multiple times independently and the distribution of results considered.

Abstract

We systematically evaluate the reproducibility of data analysis conducted by Large Language Models (LLMs). We evaluate two prompting strategies, six models, and four temperature settings, with ten independent executions per configuration, yielding 480 total attempts. We assess the completion, concordance, validity, and consistency of each attempt and find considerable variation in the analytical results even for consistent configurations. This suggests, as with human data analysis, the data analysis conducted by LLMs can vary, even given the same task, data, and settings. Our results mean that if an LLM is being used to conduct data analysis, then it should be run multiple times independently and the distribution of results considered.

Figures (8)

• Figure 1: Evaluation metrics across pipeline steps, models, temperatures, and prompting strategies. Each tile shows the rate for one model-step combination. Rows are grouped by prompting strategy, columns by temperature. Color intensity indicates the metric value from 0 (red) to 1 (green). GPT-5-mini does not support custom temperature and is only evaluated at its default temperature of 1.
• Figure 2: Comparison of LLM-estimated reappointment rates to those from human analysis at the department-year level. Each point is one department-year observation from one execution. The dashed 45-degree line indicates the estimates are the same. GPT-5-mini is only evaluated at its default temperature of 1.
• Figure 3: Output from each individual execution, by configuration. Each dot is one independent execution. The number in brackets shows how many executions produced valid regression output out of the total attempts. GPT-5-mini configurations are collapsed across temperature settings because the model does not support custom temperature. Configurations are grouped by model, then prompt strategy, then temperature. The dashed vertical line shows the human-generated value.
• Figure A1: Annual appointment counts by reappointment status in New Brunswick (2013--2024). Dark bars show new appointments while light bars show reappointments. There is year-to-year variation in both total appointments and the number of reappointments.
• Figure A2: Completion