摘要：

Information design studies how a sender can influence a receiver's decision by choosing what information to disclose. The sender-receiver framework provides a general perspective on communication and decision-making when the sender and the receiver may have different objectives. Beyond its classical applications in economics, information design has found applications in multi-agent RL and LLM reasoning. In both settings, a central question is how to design information structures to achieve desired outcomes.

In this talk, we introduce the foundations of Bayesian Persuasion. We then study online Bayesian persuasion under bandit feedback, explaining an online algorithm and its regret bound. Turning to LLM reasoning, we introduce an Advisor-Receiver framework in which an advisor learns to select information for a language model. Finally, we discuss information design in multi-agent RL through Markov signaling games.

论坛简介：该线上论坛是由张志华教授机器学习实验室组织，每两周主办一次（除了公共假期）。论坛每次邀请一位博士生就某个前沿课题做较为系统深入的介绍，主题包括但不限于机器学习、高维统计学、运筹优化和理论计算机科学。