[4S1-GS-2] Machine learning:

[4S1-GS-2-01] Analysis of LLM contextual understanding in a specific domain

〇Riku Ogata¹, Masahiro Okano¹, Junichi Okubo¹, Junichiro Fujii¹ (1. Yachiyo Engineering Co., Ltd.)

Keywords:mechanistic interpretability, context, specific domain

Large Language Models (LLMs) have low accuracy for long-tail knowledge, and solutions include Retrieval Augmented Generation (RAG) and fine tuning. On the other hand, there are reports that it is not practical due to factors such as differences in the evaluation task. In practical applications, more complex tasks can be solved and it is necessary to understand the context. However, it is not clear how LLMs internally processes texts and captures the context in specific domains such as civil engineering, which is a long-tail domain. We believe that clarifying how LLMs capture context in specific domains will contribute to improving for long-tail knowledge. In this study, we first identified uncertainty in response generation in both the general and specific domains, and then clarified the features of LLM internal processing processes in a specific domain by analyzing the variation of entropy in the intermediate representation of each layer of the LLM.