2026年第54期(总第1198期)
演讲主题：When LLM Agents Negotiate: Private Information and Dynamic Bargaining in Supply Chains
主讲人：梁辰 美国康涅狄格大学商学院运营与信息管理系副教授
主持人：李建斌 供应链管理与系统工程系教授、副院长
活动时间：2026年07月07日（周二）09:00-10:30
活动地址：管院大楼107教室
主讲人简介：
Chen Liang is an Associate Professor in the Operations and Information Management Department and the Dean's Ackerman Scholar at the University of Connecticut School of Business. She earned her PhD from Arizona State University. Her research focuses on digital platforms, future of work, human–AI interaction, and the economics of AI. Her work has been published in leading journals including Management Science, Information Systems Research, and MIS Quarterly, as well as premier conferences such as IJCAI, EMNLP, and CHI. She has received several prestigious honors, including the INFORMS Information Systems Society (ISS) Gordon B. Davis Young Scholar Award, the Association for Information Systems (AIS) Early Career Award, an NSF Dissertation Grant, and multiple Best Paper awards.
活动简介：
As LLM agents move from decision support to autonomous procurement, firms need to know whether they negotiate efficiently, divide value predictably, and avoid irrational deals. This paper studies these questions through 9,300 LLM-to-LLM negotiations in a canonical supply chain bargaining problem with private information, benchmarking nine models against Perfect Bayesian Equilibrium. Three findings emerge. First, capability drives value creation: LLM agents reach agreement in 98.9% of cases and recover 95% of first-best surplus, but take 2.98 rounds versus a Bayesian benchmark of 1.25, eroding up to 39% of discounted surplus; baseline models accept individually irrational contracts in 19% of cases versus near-zero at flagship tier. Second, provider identity, not capability, determines distribution: self-play buyer share ranges from 41% (OpenAI) to 70% (Qwen), and bargaining power is relational rather than intrinsic. Third, the principal's design choices are strategic levers: most distinctively, announced patience (the agent's prompted discount factor) explains 85-99% of surplus division, making it the most consequential deployment parameter available to practitioners. This paper distills these results into an equilibrium-referenced audit framework and a three-dimensional deployment map for strategic AI agents.
