Presentation Information
[4Yin-B-02]Structural Biases in Industrial Synergy Estimation by LLMsQuantitative Verification Based on Input-Output Tables
〇Yohei Nishitsuji1, Shogo Masaya2 (1. Sumitomo Corporation, 2. INPEX Corporation)
Keywords:
LLM,Industrial Synergy,Input-Output Tables,Network Analysis
The advent of autonomous agent-to-agent commerce (A2A economy) has the potential to fundamentally transform conventional industrial structures. However, the ability of large language models (LLMs), which underpin agent decision-making, to accurately capture real-world economic structures remains unverified. In this study, we use vertical and complementary synergy indicators derived from input-output tables as ground truth and evaluate zero-shot estimation accuracy and structural biases across four models: Claude 4.5 Sonnet, GPT-5.2, Gemini 3 Pro, and Grok 4 Expert. Results show that Grok 4 Expert successfully reproduces sparse network structures and industrial centrality close to those observed in the real economy, while Gemini 3 Pro demonstrates high accuracy in estimating link strengths. This paper discusses the importance of developing specialized models that account for the characteristics of LLMs in the design of agents for the A2A economy.