[4G3-GS-6] Language media processing:

[4G3-GS-6-02] Causal-Chain Search Augmented Language Models

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

Keywords:AI, NLP, LLM-based Agent, Economic Causal Chain

This research proposes a method for a large language model-based agent that leverages Economic Causal-Chain search as a tool to perform causal reasoning based on causal facts. According to the previous work, causal reasoning by generative language models merely replicates patterns found in training data and does not reflect a genuine understanding of causal relationships. To address this limitation, we designed an enhanced generative approach that leverages external information sources, specifically economic causal-chain, to enable fact-based, multi-step causal reasoning. To evaluate the effectiveness of the proposed method, we applied and compared it with prompt-based techniques, specifically Chain of Thought and Tree of Thoughts methods. The experimental results demonstrated significant improvements in the diversity and factuality of generated texts. These findings indicate that utilizing causal chains improves the causal reasoning capabilities of language models. This contribution highlights the utility of artificial intelligence technologies in finance, economics, and other domains where advanced causal inference is required.