Scholarlink

<p>Surveys are a cornerstone of research metrics and innovation studies, providing key indicators for research evaluation, policy design, and comparative analysis. Yet they increasingly face declining response rates, survey fatigue, and limited explanatory depth. Traditional surveys capture what firms do but provide restricted insight into why specific strategies are chosen. This study explores whether large language models (LLMs) can generate synthetic innovation survey responses that both replicate empirical patterns and offer contextually grounded explanations. Using GPT-4o, we simulate firm-level responses to the Korea Innovation Survey (KIS) by assigning firm characteristics such as size, industry, R&amp;D intensity, human capital, and market orientation. We evaluate whether LLM-generated responses reproduce established innovation patterns, including firm size effects and Pavitt’s sectoral taxonomy, and provide coherent explanations consistent with organizational decision-making theories. Results show that LLMs replicate systematic patterns across innovation dimensions. Response rates rise with firm size, aligning with resource-based theories. LLM responses reflect Pavitt classifications: supplier-dominated firms stress process innovation; science-based firms emphasize patents and university ties; scale-intensive firms leverage internal capabilities; and specialized suppliers draw on customer feedback. Generated explanations invoke theoretical concepts such as absorptive capacity and appropriability without prompting. We position LLMs not as replacements for human respondents but as complementary tools for rapid scenario exploration, hypothesis generation, and preliminary pattern identification. Yet limitations remain: LLMs mirror training data rather than empirical reality, and their explanations are generated text, not causal reasoning. Appropriate applications include exploratory analysis and pilot testing, but not empirical validation, policy evaluation, or causal inference.</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

LLMs as complementary tools for innovation surveys research: pattern replication and contextual relevance

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月9日

查看全文
  • Jae Hyung Park,
  • Jae-Suk Yang

摘要

Surveys are a cornerstone of research metrics and innovation studies, providing key indicators for research evaluation, policy design, and comparative analysis. Yet they increasingly face declining response rates, survey fatigue, and limited explanatory depth. Traditional surveys capture what firms do but provide restricted insight into why specific strategies are chosen. This study explores whether large language models (LLMs) can generate synthetic innovation survey responses that both replicate empirical patterns and offer contextually grounded explanations. Using GPT-4o, we simulate firm-level responses to the Korea Innovation Survey (KIS) by assigning firm characteristics such as size, industry, R&D intensity, human capital, and market orientation. We evaluate whether LLM-generated responses reproduce established innovation patterns, including firm size effects and Pavitt’s sectoral taxonomy, and provide coherent explanations consistent with organizational decision-making theories. Results show that LLMs replicate systematic patterns across innovation dimensions. Response rates rise with firm size, aligning with resource-based theories. LLM responses reflect Pavitt classifications: supplier-dominated firms stress process innovation; science-based firms emphasize patents and university ties; scale-intensive firms leverage internal capabilities; and specialized suppliers draw on customer feedback. Generated explanations invoke theoretical concepts such as absorptive capacity and appropriability without prompting. We position LLMs not as replacements for human respondents but as complementary tools for rapid scenario exploration, hypothesis generation, and preliminary pattern identification. Yet limitations remain: LLMs mirror training data rather than empirical reality, and their explanations are generated text, not causal reasoning. Appropriate applications include exploratory analysis and pilot testing, but not empirical validation, policy evaluation, or causal inference.