Evaluating Generalization Capabilities of LLM-Based Agents in Mixed-Motive Scenarios Using Concordia

shuqing shi (Kings College London) · Yali Du (King‘s College London) · Yang Liu (CUHK) · José Hernández-Orallo (Universitat Politècnica de València) · Mengmeng Wang (Zhejiang University of Technology) · Akash Kundu (Heritage Institute of Technology) · Zilong Zheng (Beijing Institute for General Artificial Intelligence) · Yuxuan Chen (Tsinghua University, Tsinghua University) · Xue Feng (BIGAI) · Chandler Smith (Oxford University) · Marwa Abdulhai (University of California, Berkeley) · Manfred Díaz (Mila, Quebec) · Marko Tesic (University of Cambridge) · Rakshit Trivedi (Massachusetts Institute of Technology) · Sasha Vezhnevets (DeepMind) · Lewis Hammond (University of Oxford / Cooperative AI Foundation) · Jesse Clifton (Center on Long-Term Risk) · Minsuk Chang (Google Deepmind) · Edgar Duenez-Guzman (Google DeepMind) · John Agapiou (Google DeepMind) · Jayd Matyas (DeepMind) · Danny Karmon (Google DeepMind) · Beining Zhang (University of Southampton ) · Jim Dilkes (University of Southampton) · Hieu Minh Nguyen (Apart Research) · Emanuel Tewolde (Carnegie Mellon University) · Jebish Purbey (Tribhuvan University) · Ram Mohan Rao Kadiyala · Siddhant Gupta (Indian Institute of Technology, Roorkee) · Aliaksei Korshuk (Coframe) · Buyantuev Alexander (Higher School of Economics) · Ilya Makarov (AIRI & ISP RAS) · Gang Zhao (Shanghai Research Institute for Intelligent Autonomous Systems, Tongji University) · Rolando Fernandez (University of Texas at Austin) · Zhihan Wang (University of Texas at Austin) · Caroline Wang (The University of Texas at Austin | Google DeepMind) · Jiaxun Cui (Meta) · Lingyun Xiao (University of Texas at Austin) · Di Shi (University of Texas at Austin) · Yoonchang Sung (Nanyang Technological University) · Muhammad Arrasy Rahman (The University of Texas at Austin) · Peter Stone (The University of Texas at Austin, Sony AI) · Yipeng Kang (National Key Laboratory of General Artificial Intelligence) · Hyeonggeun Yun (Companoid Labs) · Ananya Ananya (Stanford University) · Taehun Cha (Korea University) · Zhiqiang Wu (Tongji University) · Elizaveta Tennant (University College London) · Olivia Macmillan-Scott (UCL) · Marta Segura (University College London, University of London) · Diana Riazi (Department of Computer Science, University College London, University of London) · Fuyang Cui (University of Toronto) · Sriram Ganapathi (University of Waterloo) · Toryn Klassen (University of Toronto) · Nico Schiavone (University of Toronto) · Mogtaba Alim (University of Toronto) · Sheila McIlraith (University of Toronto and Vector Institute) · Manuel Rios (Universidad de los Andes) · Oswaldo Peña (Universidad Nacional de Colombia) · Carlos Rojas (Grupo Bancolombia) · Manuela Chacon-Chamorro (Universidad de los Andes) · Rubén Manrique (Universidad de Los Andes) · Luis Felipe Giraldo (Universidad de Los Andes) · Nicanor Quijano (Universidad de Los Andes) · Yiding Wang (Peking University) · Fangwei Zhong (Beijing Normal University) · Wenming Tu (Shanghai Jiaotong University) · Zhaowei Zhang (Peking University) · Ziang Chen (Tsinghua University, Tsinghua University) · Zixia Jia (BigAI) · Chichen Lin · Weijian Fan (Communication University of China) · Chenao Liu (Communication University of China) · Sneheel Sarangi (New York University Abu Dhabi) · Ziyan Wang (Facebook) · Avinaash Anand Kulandaivel · Wu Ruiyang (Communication University of China) · Chetan Talele · 陆孙嘉 (Communication University of China) · Gema Parreno (--) · Shamika Dhuri (Carnegie Mellon University) · Bain McHale (CMU, Carnegie Mellon University) · Tim Baarslag (Centrum Wiskunde & Informatica / Eindhoven University of Technology) · Dylan Hadfield-Menell (MIT) · Natasha Jaques (University of Washington, Google DeepMind) · Joel Leibo (DeepMind)

collective action problemsconcordiaempirical results evaluation methodsgeneral cooperative intelligencelarge language model llm agentsmixed-motive environmentsmulti-agent simulationmutual gainnegotiationnorm enforcementpersuasionrobust generalization social interaction zero-shot

Large Language Model (LLM) agents have demonstrated impressive capabilities for social interaction and are increasingly being deployed in situations where they might engage with both human and artificial agents. These interactions represent a critical frontier for LLM-based agents, yet existing evaluation methods fail to measure how well these capabilities generalize to novel social situations. In this paper, we introduce a method for evaluating the ability of LLM-based agents to cooperate in zero-shot, mixed-motive environments using Concordia, a natural language multi-agent simulation environment. Our method measures general cooperative intelligence by testing an agent's ability to identify and exploit opportunities for mutual gain across diverse partners and contexts. We present empirical results from the NeurIPS 2024 Concordia Contest, where agents were evaluated on their ability to achieve mutual gains across a suite of diverse scenarios ranging from negotiation to collective action problems. Our findings reveal significant gaps between current agent capabilities and the robust generalization required for reliable cooperation, particularly in scenarios demanding persuasion and norm enforcement.