Invited Talks

The Limits of Multi-task Peer Prediction

  • Shuran Zheng, Harvard University
  • Time: 2022-04-26 11:00
  • Host: Dr. Yuqing Kong
  • Venue: Online Talk


Peer prediction refers to a collection of incentive mechanisms that have been designed for the challenging setting where truthful information elicitation about some tasks is desired but the designer has no access to the ground truth (i.e. event outcomes) for incentive alignment. This setting is fundamental to many information elicitation applications such as peer grading, surveys, product reviews, and forecasting for long-term events. Recent advances in peer prediction have progressed from single-task peer prediction, where an agent's reward on a task is solely determined by how his report on the task relates to the reports made by peer agents on the same task, to multi-task peer prediction, where reports made by peer agents on other tasks can also be used in determining the agent's reward on the task. Multi-task peer prediction mechanisms can often achieve stronger incentive guarantees or require fewer assumptions on the underlying information structure than single-task peer prediction mechanisms, thanks to the additional cross-task information. While the quest for better peer prediction mechanisms is bound to continue, we attempt to understand the limits for designing multi-task peer prediction mechanisms in this work: When is it possible to design a desirable multi-task peer prediction mechanism? We provide a characterization of the elicitable multi-task peer prediction problems, assuming that the designer rewards participants' reports for different tasks separately. The characterization uses a geometric approach based on the power diagram characterization in the single-task setting. This is a joint work with Yiling Chen and Fang-Yi Yu.



Shuran Zheng is a fifth-year PhD at Harvard University. Before joining Harvard, She got her Bachelor's degree from Tsinghua, IIIS. Shuran studies Economics and Computation with a focus on Markets for Information and Data. In this talk, she is going to introduce her work on the theoretical foundation of Peer Prediction.


  • Admission


Zoom ID: 82457812359

Passcode: 202204