[5Yin-A] Poster 8

[5Yin-A-55] Evaluating Inference Accuracy Improvements in Causal Structure Estimation via Few-Shot Learning

〇Takuya Yasuda¹, Eijiro Mochida¹, Yuri Murayama¹, Kiyoshi Izumi¹ (1. The University of Tokyo Graduate School)

Keywords:

LLM,Causal Graph,Statistical Causal Inference

This study evaluates the capability of large language models (LLMs) to infer causal structures and estimate causal effect strengths in economic and financial causal inference tasks. Specifically, we investigate how few-shot learning (FSL)—where pairs of causal graphs and corresponding causal strengths are provided as in-context exemplars—affects inference performance. Using causal structure data extracted from existing empirical studies in the financial and economic domains, we assess both causal structure inference and causal strength estimation within a unified evaluation framework. Structural accuracy of inferred causal graphs is measured using Structural Hamming Distance (SHD) and F1 score, while errors in causal strength estimation are quantified by Mean Absolute Error (MAE) and Root Mean Square Error (RMSE). In addition, we examine the robustness of LLM-based causal inference by introducing irrelevant nodes into the candidate variable set and analyzing performance under varying noise levels and few-shot configurations. The results clarify the effectiveness and limitations of LLMs for causal structure and causal strength inference, and provide foundational insights for specifying causal structural assumptions in statistical causal inference and related practical applications.