From 11a856510b61e7a179b6c03d8f1001150159841a Mon Sep 17 00:00:00 2001 From: mmorge <maxime.morge@univ-lyon1.fr> Date: Fri, 9 May 2025 08:38:10 +0200 Subject: [PATCH] LLM4AAMAS: Add li23etal --- README.md | 21 +++++++++++++++++++++ 1 file changed, 21 insertions(+) diff --git a/README.md b/README.md index 5d24540..00ace28 100644 --- a/README.md +++ b/README.md @@ -394,6 +394,27 @@ produce the most successful plan and scale best to large numbers of agents. *2024 IEEE International Conference on Robotics and Automation (ICRA)*, pp. 4311-4317. +This study investigates the performance of LLM-based agents in a cooperative +multi-agent text game involving Theory of Mind (ToM) inference tasks, and +compares them with Multi-Agent Reinforcement Learning (MARL). The study +introduces a prompt-engineering approach designed to mitigate common failures +that hinder coordination among these agents. A key challenge identified is that +LLMs tend to overlook relevant information included early in the prompt, +especially when it is distant from the specific planning query. To address this, +the authors incorporate an explicit belief state into the prompt to re-emphasize +task-relevant details, thereby improving coherence and decision-making. +Furthermore, the lack of explicit belief representation often leads to +hallucinations, where agents generate inconsistent or incorrect outputs. The +proposed method improves collaboration by enabling agents to form and maintain +more accurate beliefs. Finally, the study notes that while LLM-based agents can +rapidly propagate misinformation—particularly when they struggle to track +information flow during ToM inference. + +- **[Theory of Mind for Multi-Agent Collaboration via Large Language Models](https://doi.org/10.18653/v1/2023.emnlp-main.13)** + Huao Li, Yu Chong, Simon Stepputtis, Joseph Campbell, Dana Hughes, Charles Lewis, Katia Sycara, Kalika Bali (2023) + In *Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP)*, Singapore, pp. 180–192. + Association for Computational Linguistics. + ### Social Simulation LLMs can simulate realistic perceptions, reasoning, and decision-making, react -- GitLab