用大语言模型绘制狭窄走廊
Mapping the Narrow Corridor with Large Language Models
浏览论文内容
中文总结 AI 辅助
研究探讨大语言模型能否将国家历史在二维空间路径的框架付诸实践，引入可重现流程，制作12国轨迹图集并比较四个模型，评估与专家指数的一致性及模型间一致性，还讨论了大语言模型注释器和人类专家偏差差异。
中文摘要 AI 辅助
阿西莫格鲁和罗宾逊（2024年诺贝尔经济学奖获得者）在《狭窄走廊》中追溯国家历史，将其视为二维（国家权力、社会权力）空间中的路径，但未绘制具体内容：量化每个时刻的两个轴是一项劳动密集且充满判断的任务，这使得该框架一直保持定性。我们探讨大语言模型（LLMs）能否将其付诸实践。我们引入了一个可重现的、与提供商无关的流程，通过思维链（先事件，后评分）并以上一时期的上下文锚定，逐期对一个国家进行评分，对照一个具有模式验证输出的明确评分标准。作为概念验证，我们制作了一个包含12个国家的轨迹图集（伊朗、法国、英国、美国、中国、智利、哥伦比亚、刚果民主共和国、黎巴嫩、赞比亚、索马里和印度），这些国家被选来代表书中四种利维坦类型中的每一种至少两个国家，并比较了四个模型。我们评估了与专家指数（V-Dem）的一致性得分以及模型间的一致性，将一致性视为并发一致性而非准确性，因为此类指数可能出现在模型的训练数据中，并从概念上讨论了大语言模型注释器和人类专家的偏差差异。代码、提示、运行结果以及动画轨迹的交互式图库已发布。
英文摘要
Acemoglu and Robinson (2024 Nobel laureates in Economics) trace national histories in The Narrow Corridor as paths through a two-dimensional (state power, society power) space but plot none: quantifying the two axes at each moment is the labor-intensive, judgment-laden task that has kept the framework qualitative. We ask whether large language models (LLMs) can operationalize it. We introduce a reproducible, provider-agnostic pipeline that scores a country period by period using chain-of-thought (events first, then score) and in-context anchoring on prior periods, against an explicit rubric with schema-validated output. As a proof of concept we produce a twelve-country trajectory atlas (Iran, France, the United Kingdom, the United States, China, Chile, Colombia, the Democratic Republic of the Congo, Lebanon, Zambia, Somalia, and India), chosen to place at least two countries in each of the book's four Leviathan types, and compare four models. We assess the scores for consistency with an expert index (V-Dem) and for inter-model agreement, reading agreement as concurrent consistency rather than accuracy, since such indices likely appear in the models' training data, and discuss conceptually how the biases of an LLM annotator and a human expert differ. Code, prompts, runs, and an interactive gallery of the animated trajectories are released.
