AI Safety in Multi-Agent LLM Systems – Zhijing Jin | IASEAI 2025 скачать в хорошем качестве

AI Safety in Multi-Agent LLM Systems – Zhijing Jin | IASEAI 2025

What happens when large language model agents live in a community—do they cooperate, or act selfishly? In this IASEAI ’25 session, AI Safety in Multi-Agent LLM Systems, Zhijing Jin (Postdoctoral Researcher at the Max Planck Institutes and incoming Assistant Professor at the University of Toronto) introduces GovSim, a benchmarking platform designed to test whether LLM agents fall into the Tragedy of the Commons. Findings show that even the best-performing model (GPT-4o) survives less than 54% of the time, raising urgent questions about AI safety in multi-agent systems. Jin explores the implications of these results and the need for stronger safeguards as AI agents increasingly interact and make collective decisions. About IASEAI: https://www.iaseai.org Zhijing Jin: https://zhijing-jin.com/ #ZhijingJin #AISafety #MultiAgentAI #LLMAgents #IASEAI