Seminar, Hosted by Kempner Institute, Harvard University
Learning Efficient Multi-turn Communication
Carnegie Mellon University
Abstract
People benefit from collaborating with an AI agent when it is more efficient to communicate their goals and verify that the agent has met them than to do the task themselves. This talk presents work on training agents to communicate more efficiently with people, using simulated interactions between two agents to generate synthetic training data and showing that the trained agents communicate better with people. First, the approach is applied to train a program synthesizer that better infers user intent, exploiting the fact that a user chooses examples of program behavior cooperatively, picking those that are specifically informative about their intent. Second, the approach trains agents that communicate more efficiently about images by forming shared conventions of reference with a listener over multiple rounds of interaction. These approaches pave the way for AI agents to become helpful collaborators on increasingly complex tasks.