Limits of Large Language Models in Debating Humans

This research investigates the capabilities of Large Language Models (LLMs) in debate-based opinion consensus formation. It tests LLMs in environments involving humans only, agents and humans, and agents only. The study found that while LLMs can blend in and facilitate human productivity, they are less convincing in debate, with their behavior deviating from humans’. The research concludes that LLMs need further evolution to be viable debaters.

Publication date: 6 Feb 2024
Project Page: https://arxiv.org/abs/2402.06049
Paper: https://arxiv.org/pdf/2402.06049