Scholarlink

<p>Power-law distributions govern essential features of many real-world systems. We test GPT-4o’s ability to generate synthetic data with power-law/Zipf-like scaling across three scenarios: city populations, webpage visits, and company data-while varying prompt styles (Natural, Mixed, Controlled). Natural prompts yield exponents within empirical heavy-tail ranges (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\alpha \!\approx \!1.65\)</EquationSource> </InlineEquation> cities, <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\alpha \!\approx \!1.68\)</EquationSource> </InlineEquation> webpage visits, <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(\alpha \!\approx \!1.63\)</EquationSource> </InlineEquation> company data). Controlled prompts substantially increase <InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(\alpha\)</EquationSource> </InlineEquation> (<InlineEquation ID="IEq5"> <EquationSource Format="TEX">\(\approx \!3.0\)</EquationSource> </InlineEquation>, <InlineEquation ID="IEq6"> <EquationSource Format="TEX">\(\approx \!4.6\)</EquationSource> </InlineEquation>, <InlineEquation ID="IEq7"> <EquationSource Format="TEX">\(\approx \!3.0\)</EquationSource> </InlineEquation>), producing steeper (lighter-tailed) samples. Pareto tail fit is strong for cities and companies in the Controlled condition, but poor for web traffic. Mixed prompts are often intermediate. Scaling is strongly prompt-dependent. Synthetic datasets generated by LLMs should therefore be treated as provisional and verified with tail-aware diagnostics before use in research or teaching.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Prompt-driven biases in generative pre-trained transformer-generated data: a statistical examination of Zipf and power-law patterns

  • 期刊论文

  • 发表日期：

    2026年4月11日

查看全文
  • Andrej Novak,
  • Goran Oblakovic,
  • Mato Njavro,
  • Martin Ehler

摘要

Power-law distributions govern essential features of many real-world systems. We test GPT-4o’s ability to generate synthetic data with power-law/Zipf-like scaling across three scenarios: city populations, webpage visits, and company data-while varying prompt styles (Natural, Mixed, Controlled). Natural prompts yield exponents within empirical heavy-tail ranges ( \(\alpha \!\approx \!1.65\) cities, \(\alpha \!\approx \!1.68\) webpage visits, \(\alpha \!\approx \!1.63\) company data). Controlled prompts substantially increase \(\alpha\) ( \(\approx \!3.0\) , \(\approx \!4.6\) , \(\approx \!3.0\) ), producing steeper (lighter-tailed) samples. Pareto tail fit is strong for cities and companies in the Controlled condition, but poor for web traffic. Mixed prompts are often intermediate. Scaling is strongly prompt-dependent. Synthetic datasets generated by LLMs should therefore be treated as provisional and verified with tail-aware diagnostics before use in research or teaching.