[论文解读] Same Prompt, Different Outcomes: Evaluating the Reproducibility of Data Analysis by LLMs

Jiaxin Cui, Rohan Alexander|arXiv (Cornell University)|Feb 15, 2026

Topic Modeling被引用 0

一句话总结

本文系统评估在不同模型、提示和温度下，由大语言模型生成的数据分析的可重复性，即使在相同配置下也存在显著差异，并建议进行多次独立运行。

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.

研究动机与目标

• Motivate the need to study reproducibility in LLM-generated data analysis and its implications for scientific findings.
• Assess how prompting strategy (single-step vs multi-step), model, and temperature affect reproducibility.
• Quantify completion, concordance, validity, and consistency across configurations using a five-step data pipeline.
• Provide guidance on running multiple executions and considering result distributions in LLM-based analyses.

提出的方法

• Evaluate six models from three providers (Anthropic, OpenAI, Google) across two prompting strategies (single-step, multi-step) and four temperatures (0.0, 0.3, 0.7, 1.0, with GPT-5-mini at default 1.0).
• Run ten independent executions per configuration, totaling 480 attempts, on a five-step data-analysis pipeline applied to New Brunswick appointment data.
• Assess four metrics—completion, concordance, validity, and consistency—for each attempt; analyze outputs including code execution, alignment with human analyses, data types, and regression results.
• Use a five-step pipeline: consolidate CSVs, identify reappointments, aggregate to organization-year summaries, perform OLS regression, and generate visualizations.
• Document that multi-step prompting suffers from error propagation; single-step prompts yield higher completion and more consistent pipelines under many configurations.
• Employ R with tidyverse and tinytable for analysis, tables, and figures.

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-mi

实验结果

研究问题

• RQ1 How reproducible are LLM-generated data analyses when the same task, data, and settings are used across multiple executions?
• RQ2 Do single-step prompts yield more reliable outputs than multi-step prompts due to error propagation, and how do models and temperatures influence this?
• RQ3 To what extent do LLM-generated analyses align with human analyses (concordance) and meet validity criteria across pipeline steps?
• RQ4 How do data-preparation decisions within LLM-generated pipelines affect downstream estimates such as regression slopes and t-statistics?

主要发现

• Single-step prompting generally achieves higher completion rates than multi-step prompting due to reduced error propagation.
• Generated code is structurally valid but data-preparation choices (sorting, missingness, inclusion) differ from human analyses and drive downstream variability.
• There are distinct clusters of estimates across executions, with slopes and t-statistics varying in sign and magnitude even when most configurations yield non-significant t-stats.
• Most configurations produce t-statistics near zero, but some single-step configurations yield potentially significant results; however, across many runs, variability undermines single definitive conclusions.
• Consistency across executions is imperfect; identical configurations can lead to different outputs, underscoring the need for multiple independent runs.
• The study emphasizes evaluating distributions of results and, where possible, using ensembles or multi-provider comparisons to account for variability.

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 temperat