Incentivizing Effort and Honesty for High-quality Information

Abstract

In today’s data-driven world, high-quality data is the lifeblood of artificial intelligence (AI) systems. However, a pressing question arises: how can we obtain high-quality information from human agents?

A critical challenge lies in incentivizing agents to put forth diligent effort and report truthful information, even when their objectives (e.g.~earning money) diverge from the designer’s goals (e.g.~training powerful AI). Reward mechanisms that distort agents’ incentives may result in the collection of noisy and even manipulated data. This thesis addresses these incentive problems in collaborative information systems by integrating and advancing methods from information elicitation, contract design, and mechanism design.

Mechanisms in different settings must take care of agents' diverse incentives differently. This thesis first focuses on settings where agents are textbf{non-stakeholders} of the system, meaning that agents are primarily motivated by payments (or numerical scores) and are not concerned about how the collected data is used. We are primarily interested in the scenario where each agent responds to multiple tasks, and the correctness of an agent's reports cannot be directly verified with the ground truth but can only be approximated by comparing them with other agents' reports on the same tasks, known as emph{multi-task peer prediction}.

In the peer prediction setting, this thesis first addresses a long-standing limitation of prior multi-task peer prediction mechanisms. That is, truth-telling was ensured to maximize the expected reward only under the assumption that every agent uses the same reporting strategy across all tasks. Our proposed Matching Agreement mechanism provides an elegant solution to this issue.

Furthermore, we identify a practical concern in existing peer prediction mechanisms: the goals of incentivizing effort, truth-telling, and maintaining practical payments cannot always be achieved simultaneously. To address this, we propose rewarding agents based on a tournament of their performance scores computed by a peer prediction mechanism. While a naive combination does not necessarily preserve the truthful properties of the peer prediction mechanisms, we show how to mitigate this issue by modifying performance scores.

When agents are textbf{stakeholders} in the system, their reported information may affect their (non-monetary) utilities through an additional feedback loop. One example is peer review, where authors possess valuable information about their own papers, but revealing this information to the conference may influence the acceptance or rejection outcomes. The thesis first investigates how a prestigious conference should design its acceptance policy, such as adjusting the acceptance threshold, while considering while considering that authors may strategically decide where to submit or resubmit their papers based on their knowledge of the paper's quality. We further generalize the model to investigate how the tableau of different conferences and acceptance types (e.g.~oral and poster presentation), which we call conference design, drives the strategic submission and production of authors.

Lastly, the thesis designs a novel review mechanism that directly elicits a ranking of paper quality from an author with multiple papers. We prove several attractive properties of our mechanism: 1) it ensures that truth-telling is the author's best response under mild assumptions, 2) it reallocates review resources from likely rejected papers to likely high-quality papers; and 3) it incentivizes authors to write fewer, but higher-quality papers.