Evaluative Misalignment in LLM-Supported Policy Analysis: Two Heuristic Cases

Engineering notes

Engineering notes will be added by the aipentium editorial team.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为大语言模型、生成式AI、ChatGPT相关技术、计算机视觉、深度学习等高价值论文补充中文说明。

Original abstract

Large language models (LLMs) are increasingly used in policy analysis, where evaluation often relies on prompt optimization followed by output assessment. This ERF paper examines evaluative misalignment in two heuristic cases from our applied LLM work in policy analysis: qualitative coding and policy summarization. We identify two potential forms of misalignment in LLM evaluation processes: between human reasoning and model behavior, and among human evaluators. The cases show that plausible or high-scoring outputs may still diverge from the reasoning and contextual interpretation required in policy tasks. We suggest that LLM output evaluation in such settings should be understood as a process-centered, role-aligned activity. The paper offers a preliminary account of how evaluative misalignment emerges when output scores, model behavior, and human judgment are positioned within policy analysis workflows.

Links and sources

Comments