[4R1-OS-8a] OS-8

[4R1-OS-8a-05] Prototyping Method for Controlling Opinion Change of Virtual Citizen Agents for LLM-based Discussion Simulator

〇Sora Matsumoto¹, Shun Shiramatsu¹, Takashi Iwata^2,3, Takeshi Mizumoto⁴ (1. Nagoya Institute of Technology, 2. HammerBird INC., 3. Keio Research Institute at SFC, 4. Hylable Inc.)

[[Online]]

Keywords:consensus building, large language model, controlling opinion change, discussion simulator

We attempt to develop a discussion simulator that enables users practice of discuss with virtual citizen agents using a large-scale language model toward promote citizen participation. For this purpose, we This will enable users to practice participation in discussions and improve their sense of self-efficacy when participating. So far, workshops have been held in June and December 2023 using this discussion simulator. However, the generated opinions failed to replicate natural opinion change. Therefore, in this paper, we developed a prototype of an opinion generation method that allows virtual citizens to be moderately influenced by other people's opinions while maintaining their own arguments appropriately and allowing them to change their opinions appropriately. Specifically, we quantified the degree of influence of other people's opinions, and developed a prototype method in which the degree of influence can be controlled by LLM. Furthermore, we verified whether our method can appropriately control the opinion change. The experimental result indicates that in-context learning using GPT-4 has a potential to control the tendency of opinion change and fine-tuning using GPT-3.5 is not suitable for this purpose.