Please note: This seminar will take place in DC 1304 and online.
Ira Globus-Harris, Research Professor
Center for Data Science for Enterprise and Society, Cornell University
Designing effective collaboration between humans and AI systems and between different agents, is crucial for leveraging their complementary abilities in complex decision tasks. But how should agents possessing unique knowledge—like a human expert and an AI model, or multiple LLM agents with different contexts—interact to reach decisions better than either could alone?
In this talk, I will introduce a collection of tools which allow us to develop efficient “collaboration protocols”, where parties iteratively exchange only low-dimensional information—their current predictions or best-response actions—without needing to share underlying features and which guarantee that the agents’ final predictions are provably competitive with an optimal predictor with access to their joint features.
These results are an extension of ideas from game theory’s Aumannian agreement and information substitute conditions. However, rather than considering agents as rational agents, we will taking a boosting-theoretic approach that weakens the computational requirements to risk minimization over weak learners. This will leverage a collection of tools for calibration and swap regret, offering a new foundation for building systems that achieve the power of pooled knowledge through tractable interaction alone.
To attend this seminar in person, please go to DC 1304. You can also attend virtually on Zoom.